Every 6 months, I take a sharp look at the latest developments in AI and Search. Based on how quickly things are developing, I would almost need to increase that cadence to monthly. (And maybe I will.) But first, here’s what happened in H1 2026.

H1 2026 moved money, traffic, jobs, and market cap before anyone could prove how much value AI created. Search behavior changed, token budgets exploded, software stocks sold off, and companies blamed layoffs on AI.

Every major AI story in H1 2026 was really a story about attribution.

• We’re still figuring out how to improve accuracy in measuring AI visibility.

• Companies spent billions on inference, but there is no obvious answer to “What’s the ROI?”

• Public markets punished software companies, but were investors reacting to actual or perceived disruption?

• We’re hearing “AI caused layoffs,” but when you look deeper, the real causes aren’t connected.

• It’s clear where the traffic loss is coming from, but it’s not clear what content marketplace model will replace it.

The common thread: AI’s economic impact is expanding faster than our ability to attribute it.

How does AI search see US retail?

Refresh your ecommerce strategy for the remaining summer shopping season with data from 126M+ US prompts. Explore how AI search platforms influence brand visibility, recommendations, and discovery across the customer journey.

Build your winning strategy based on the trends that are driving AI visibility for this category’s leading brands.

Read for free

AI Search

In the H1 2025 report, I predicted Google would roll AI Mode out further, and that turned out to be true. AI Mode is now a click away from AI Overviews, which means it’s just two clicks away from regular search results.

• AI Mode reached 1 billion MAU with queries around 3x longer than classic search.

• Google called it the biggest search box upgrade in 25 years at I/O 2026.

• Gemini 3 auto browse is shipping inside Chrome.

• Nick Fox says Google’s AI Mode sends out billions of clicks to the open web.

To understand the impact of this shift, I published a series of user behavior studies and data analyses in the first half of 2026.

The through line of Growth Memo research in H1 2026:

1. Tracking the presence of a brand in AI Search is severely complex. You need to factor in the engine, personalization, reasoning levels, model updates, stochastic variability, etc. Measurement is fragmented. The citation/mention overlap between engines is minuscule: 91% of citations appear in only one of ChatGPT, Perplexity, or AI Overviews. Prompt tracking should be closer to polling and focus groups than SEO rank tracking.

2. Brand mentions in AI answers are more impactful on business outcomes than citations. Don’t get me wrong, citations shape the answer, and for some businesses, that’s most important. But the majority of vendors and merchants need to pay attention to how often they show up in a panel of prompts, in which context, what sentiment and whether they’re recommended or not (ahead of competitors).

3. Trust is the highest currency in AI search. Close to 75% of consumers pick the number one result in an AI shortlist. But if they see a trusted brand anywhere on that list, they will pick it.

4. The average US adult has high trust in AI recommendations. 88% of the time, users accepted AI Mode product recommendations as best there is. In AI Overviews, on the other hand, users click, evaluate, and compare a lot. That’s classic search behavior, but it doesn’t carry over to AI chatbots.

5. Writing style, token budget, and technical factors matter for agents. If you want to be present in AI Search and agentic recommendations, you need to lead with unique information, write in a direct, easy-to-understand way, remove fluff and keep your site technically fast and easy to access.

AEO/GEO is a brand channel disguised as a performance channel. AI recommendations shape demand (not citations). The unit of optimization is whether AI names, trusts, and recommends your brand.

The grAIpes are sour

November 2025 was a pivotal turning point in AI. Claude’s Opus 4.5 was the first model that was perceived as reliable and robust enough for agentic workflows. A month later, Peter Steinberger went viral with Clawdbot, which led to millions of installs, long lines in China, and Nvidia building a Nemoclaw clone.

Then, things got even crazier. Shopify, Uber, Meta, and many other tech giants built token leadership boards that incentivized engineers to light as many tokens on fire as possible.

“Let me give you a thought experiment. Let’s say you have a software engineer or AI researcher, and you pay them $500,000 a year. At the end of the year, I’m going to ask him how much did you spend in tokens. And [if] that person said $5,000, I will go ape something else. If that $500,000 engineer did not consume at least $250,000 worth of tokens, I am going to be deeply alarmed.” — Jensen Huang

A Meta engineer created an internal leaderboard dubbed “Claudeonomics“ that ranked over 85,000 employees by the number of AI tokens they burned. It led to massive waste as employees left AI agents running on idle or useless tasks just to climb the ranks and earn titles like “Token Legend”.

Reportedly, Meta engineers consumed 73.7 trillion tokens in just 30 days. (I wonder how much of OpenAI’s and Anthropic’s growth was ignited just by workslop.)

Good times. But in April, CFOs pumped the brakes after discovering that employees had burned through their annual token budget in 4 months. TokenMAXXXING shifted to valueMAXXXING.

Meta’s leaderboards were shut down in April (along with leaderboards at other companies), but not before the #1 user on the board averaged 281 billion tokens, which would cost over $1.4 million at standard API rates.

But what remains from the all-you-can-eat token craze is a step change in productivity that we cannot yet measure. As I wrote in AI changed my work. And yours, too:

I have used AI to raise my impact by a magnitude.

• I designed landing pages end-to-end for a major travel brand that made it into production.

• I automated topic prioritization, SEO testing, and SEO reporting for my clients with full-blown apps.

• I built an array of useful applications for myself, from automating the SSI (SEO Site Index found in the bimonthly Growth Intelligence Briefs) to Openclaw agents that help me with research and charts.

One outcome of the token tsunami is that a lot more people started using Claude, which in return fired up the growth engines for AI infrastructure companies that Claude recommended, like Supabase.

Source: Supabase Series F

SaaS bloodbath

In February 2026, as a consequence of the Opus 4.6 release and Claude Cowork, the software-as-a-service industry saw a severe market downturn that wiped out hundreds of billions of dollars in market cap over just a few trading sessions. The SaaS vertical saw a steep annual decline of +30% in the markets. Valuations for high-growth software companies collapsed, with revenue multiples shrinking dramatically compared to their pandemic highs.

If you’re working in B2B, you’ve probably felt the impact in the form of lower conversion rates, longer sales cycles and softening brand searches.

Source: It’s Time to Build, July 17, 2026

It turns out not all software was (or is) forsaken. It’s the bottom quartile of software companies that pulls the market down. Interestingly, the decline is purely associated with how “the market” views a company’s AI robustness. Company valuation drops are not related to performance, but rather to whether people think AI will disrupt their product.

Over the last 30 trading days, both the top quartile and median IGV stock have outperformed the ETF as a whole—it’s the bottom quartile (which include some of the largest companies) that’s pulling down overall performance. —Source: It’s Time to Build

So, if you’re in B2B and either forced or privileged to take a new job, pick a company the market is optimistic about.

AI washing in the labor market

You might have read or heard this recently: “We’re laying off x% of our workforce due to AI.” It’s the narrative du jour. Challenger, Gray & Christmas reported that AI was the leading cited reason for May 2026 job cuts and had been cited in 87,714 cuts year-to-date, equal to 22% of all 2026 announced layoffs through May.

Source: May 2026 Challenger Report

And don’t get me wrong. Layoffs are real:

• Tech layoffs up ~66% year over year, toward 150K

• Oracle cut 21,000 citing AI

• Challenger lists AI as the top cited reason (~87,700 YTD by May)

• Block cut 40% and its stock rose (Bloomberg’s “AI washing” framing)

The only problem is that AI is not the cause. I’ve written extensively about the Big Labs using AI replacement as a marketing narrative. But it hasn’t come true yet. A lot of AI layoffs were really to offset capital expenditures, pandemic overhiring, and economic turbulence. We can all imagine a future in which AI is so good that it makes certain roles obsolete, but that time is not now.

A year ago, in the H1 2025 report, I predicted that AI layoffs were a PR stunt. And I was right: Despite clear productivity gains from AI, the recent waves of layoffs show no indication of AI being the underlying cause. In fact, companies using AI as a reason to lay employees off are rehiring them.

The agent market fragmented

ChatGPT’s market share decreased from 78% in July 2025 to 56% in July 2026, while Gemini went from 15% to 30% and Claude from 2% to 10%.

Google is now the most likely player to win the AI consumer market, and ChatGPT has lost its significance in AI Search to Google:

• ChatGPT still has ~1.1B users, but OpenAI refocused on an enterprise pivot and shut down Sora, dropped video in ChatGPT and killed Instant Checkout. Enterprise makes up ~40% of revenue, heading to ~50% ahead of a potential IPO. I’m skeptical that we’ll be able to buy OpenAI stock. According to Ed Zitron, OpenAI spends $2.8 billion USD every month to make $1.1B.

• Anthropic’s fight with the Pentagon pushed Claude to No. 1 among free apps on Apple’s U.S. App Store. A few months later, users seem less sure if Anthropic can resist attacks from open-source models.

• The Big Labs massively subsidize tokens. SemiAnalysis found a $200 USD plan of the frontier models gives you 8-14K USD worth in tokens.

Lately, open-source models like Kimi K3, GLM 5.2, or Deepseek V4 have put pressure on the big US labs because of those subsidies. Open-weight and open-source models also show that the model itself becomes commoditized, while the harness and application layer gain more importance.

Forward-deployed-engineer job postings rose 800% in the first 9 months of 2025 and were still up 729% year over year in April 2026. Frontier Labs now pays more than $500,000 in total compensation for people who can turn a model demo into a working system inside a customer’s business.

Never have marketers faced a faster evolving environment with more platforms to measure… which is why understanding what platforms matter most for your audience is important. But from an execution perspective, model availability has become a business risk, as we can see at the hand of Fable 5. Having options is good.

Publishers and the law

I predicted up to 70% of 2024 organic traffic could be gone by 2026, but that was overeager. In the end, the numbers land at ~33% referral loss year over year, with 68% of Google searches now ending without a click. Still, as a result, publishers are moving the fight to the courts and regulators.

USA Today CEO Mike Reed explained: “We’re getting close now to the point where we’ll block Google as well and abandon the traditional search traffic that we get today.”

Here are just a few lawsuit-related highlights:

• A court in Munich ruled Google liable for false statements generated by AI Overviews (Jun 2026).

• 400 newspapers sued OpenAI and Microsoft over unauthorized content use (Jun 2026).

• The UK CMA ordered Google to give publishers greater control and transparency over how content is used in AI search (including opt-outs from AI Overviews, AI Mode, and Discover summaries), along with clearer attribution and improved reporting on how content appears in AI-generated results.⁠ Google responded by adding an impression-based AI report in Search Console.

In response to the intense publishing landscape changes, Parallel AI and Cloudflare are both trying to build publishing marketplaces for content owners and AI companies.

“Assume there’s no search. You have to have your businesses planned as if search is zero.” — Conde Nast CEO Roger Lynch

Publishers are trying to replace the content-for-traffic bargain with a content-for-training market. This leaves several open questions:

• Who sets the price and how?

• How can you prove that an article contributed to an AI agent’s answer or purchase?

• Do publishers have negotiation power?

• Are smaller publishers even needed by AI labs?

Dig in: Full list of Growth Memo research completed in H1 2026

Dig deeper into the patterns covered in this halftime report below:

• User behavior studies:

  • How consumers navigate high-stakes purchases in AI Mode

  • Users behave differently in AI Overviews vs. AI Mode

  • What to do now that AIOs turned search into reading sessions

• Data analyses:

  • The Consensus Gap

  • The ghost citation problem

  • Reasoning lift: What happens to AI visibility when AI thinks harder

  • The science of how AI pays attention

  • The science of how AI picks its sources

  • The science of what AI actually rewards

  • Shorter, Focused Content Wins in ChatGPT

  • GSC data is 75% incomplete

  • Organic rankings vs. product grids: The new e-commerce divide

  • The AI skills salary premium

  • Where AI agents get stuck on your site

  • Why most original data never gets cited

As we pass the halfway point in the year, a big thanks to my editorial and research partners, Amanda Johnson and Eric van Buskirk. And an extension of gratitude to my data partners who make these analyses possible: Semrush, AirOps, Gauge, Citation Labs, Siteline, SalaryGuide, and AudienceKey.

Premium: Predictions and planning for H2 2026

H2 2026 will separate intelligence from agency: Intelligence gets cheaper, more abundant, and easier to embed through open-weight models and token competition.

But the authority to act, spend, access data, impersonate someone, or deploy the most capable models becomes more tightly controlled by platforms, governments, payment networks, and users themselves.

Read more

Subscribe now

Some brands get away with broad topical focus. Forbes, for example, ranks well for everything from “bitcoin price” to “best pillows for neck pain.” But most brands are more visible in search when focusing their content strategy on core topics.

Topical Authority is one of the most used concepts in SEO to justify content creation. It’s gone so far that swinging too far from the base can lead to algorithmic demotions (Hubspot and Clickup, to name a few significant ones). But does it translate to AI Search? Or are LLMs such a new playing field that brands can really swing big?

Imagine a payroll software company planning its next 20 articles. It can chase broad, high-volume finance topics, or build out the payroll questions it wants to be known for: W-2 deadlines, contractor classification, payroll-tax errors, overtime rules, and state registration.

The topical-authority bet is to choose the second path, even when ChatGPT can answer some of those questions without sending a click. The goal is not traffic from every glossary page. It is to become a brand that appears consistently when people ask related questions, including the eventual “Which payroll software should I use?” question.

Topical authority matters in AI search, and the reason is durability. This study shows AI answers develop topic owners: Once a brand earns an outsized share of mentions in a category, it usually keeps it.

Build your strategy with the definitive study into AI visibility

AI visibility isn’t uniform. Performance diverges sharply across platforms and industries, meaning a strategy built for one environment can fail to perform in another.

Semrush’s AI Visibility Index tracks 126 million prompts across 22 industries and ChatGPT, Google AI Mode, Gemini, and Google AI Overviews. It’s our biggest ever piece of research into AI visibility, complete with playbooks, expert guidance, and deep-dives into the winning brands.

Build a strategy that wins across where your audiences search. Start with the AI Visibility Index.

Download the complete report

Method and limitations

To find out, the Semrush team gave me access to a big dataset based on the Semrush AI Visibility Toolkit.

• Scope: 1,094 US categories, 5 prompts per category, and monthly snapshots from January through June 2026 in ChatGPT, USA only.

• The study covers +220,000 domains and +50,000 brands, +600,000 citations and +220,000 URLs

• This study tracks which brands ChatGPT mentions most often across 5 prompts per category. It does not measure sentiment, recommendation quality, user trust, or purchase impact.

• The study has no topic-level organic rankings, organic visibility, or SEO share-of-voice data.

• Semrush’s export records whether a brand appears in an answer, not the surrounding text or sentiment. A mention can be positive, neutral, or negative.

Definitions:

• Owner: The brand with the highest share of brand mentions, named in at least 4 of the category’s 5 prompts, and at least 5 percentage points ahead of the runner-up.

• Emerging leader: The most-mentioned brand appears in at least 3 prompts but does not meet the clear-leader threshold.

• Unsettled category: No leading brand appears in at least 3 prompts.

Important note: The definition for unsettled category makes sense based on the threshold we’ve seen in this specific study, but don’t treat it as a universal truth at this time.

While almost half of cited pages in the study were hard to classify (a mix of cryptic URLs like “examplesite.gov/0982349” that don’t have common/standard page types), we found that the most cited page types are product or service landing pages, followed by classic editorial content. Homepages were only cited 4% of the time.

1. Most categories have a leader, but some have an owner

Category owners dominate the category, of course. We defined them as +80% Share of Voice (available brand mentions in AI answers) and at least a 5 percentage point lead to the 2nd place.

Most categories that we analyzed in ChatGPT have not yet settled on an owner. In June 2026, only 15.2% of categories had a clear owner, while 53.7% were open fields with several contenders for the top. Over half of the +1,000 categories we analyzed have a leader that can be dethroned by a challenger. Plenty of opportunity, which I personally don’t think is true in classic Google Search.

When we look at AI (search) volume, which describes how many people might prompt in a given topic*, we find that the biggest categories by volume do not have an owner.

*Yes, I’m aware that AI (search) volume is not perfect, but when we abstract search keywords and prompts into topics and look at aggregate search volume, we can at least guess how big a topic in AI Search is.

In Semrush’s US sample for our study, the biggest topics are the least likely to have an owner.

We ranked all 1,094 categories by AI-search volume and split them into 2 groups of 547.

The top group holds 98% of all volume-based AI-search demand in the sample. It also has the lowest owner rate: 11.3%, versus 19.0% in the bottom group.

Stack those up and 89.3% of estimated AI-search demand sits in categories with no clear owner. Demand concentrates, but in this dataset, ownership does not.

When we split the data out by quartiles or deciles, we see the same trend of top-heavy categories being more open-field opportunities than smaller ones.

A category with no owner is an open budget conversation waiting to happen. The AI SEO Budget Reallocation Planner generates 5 growth-priority scenarios and a screenshot-ready visual in under 7 minutes, so the categories you choose to attack come with a number attached. Find it in the Premium Subscribers Resource Library.

2. Owners win on brand mentions, not citations

This study counts whether a brand appears in an answer, not how prominently it appears or whether it is recommended. A category owner is therefore the brand named across the largest share of its tested prompts.

Previous Growth Memo user-behavior studies have repeatedly shown that users pay little attention to citations and most to brand mentions. In shortlists, users demonstrate strong preference for the top results for purchase decisions (75% of users pick the top brand). In other words, brand mentions are AEO gold. I applied this lens to this data as well.

From How Consumers Navigate High-Stakes Purchases in AI Mode:

In classic search, 56% of participants built their own shortlist from multiple sources. In AI Mode, only 8 out of 147 codeable tasks produced a genuinely self-built shortlist. The user’s comparison process didn’t just shrink when using AI Mode. For most participants, it didn’t happen at all.

64% of AI Mode participants clicked nothing at all during their task. They read the AI’s text, sometimes scrolled through inline product snippets, and declared their finalists. The no-click rate varied by category.

Just like in classic search, the top answer carries outsized weight. 74% of participants chose the item ranked first in the AI’s response as their top pick. The mean rank of the final choice was 1.35. Only 10% chose something ranked third or lower.

While citations influence the AI answer to a prompt, the relationship between citations and brand mentions is slightly negative (-0.229). In the study, the most cited domain was rarely (20.8%) the most mentioned. But the most mentioned brand was often cited (69.9%) at least once.

All industries exhibit clear owners, except for legal. In retail, we see the highest owner rate with 61.6% (meaning 61 out of 99 categories we analyzed have a clear owner).

3. Characteristics of category owners

The available data lets us identify traits associated with ownership, but not prove how ownership was created. When compared to the the 2nd place in a category, we do see that owners have:

• Higher brand search volume (55.7%)

• Higher organic traffic (48.4%), and

• Higher authority (52.5%, as measured by Semrush’s Authority Score)

The branded-demand result means that, in 577 of 1,035 owner-runner-up pairs, the owner had higher branded search volume.

It does not mean branded demand explains 55.7% of ownership. These are useful clues, but not a recipe. The better working hypothesis is:

Brand strength may help a domain enter the answer set. Category ownership is more likely determined by unmeasured topic-specific factors, including source relevance, cited content quality, reputation, and how well a domain fits the prompt.

4. Can a challenger win a category?

Category owners are defined by their position. They’re hard to dethrone.

It’s not impossible to dethrone them, but the odds depend heavily on how settled the category is. A clear owner retained first place in 90.4% of next-month (or month over month, MoM) comparisons. Emerging and unsettled leaders change much more often.

Leaders switched in 1,950 of 5,470 MoM comparisons. But those changes were concentrated in categories where the initial lead was narrow: Switching leaders had a median 1.3-point lead, versus 2.9 points for leaders that retained first place.

3 examples where a clear leader was replaced by the previous month’s runner-up:

As you can see in the examples, you can trend up but reverse course (example 1), be flat and overtaken by a competitor (example 2) or grow and still be overtaken by a competitor that grows faster (example 3). Your trend line says little about your safety. You need a big enough margin (at least 5%).

We cannot say why these switches occurred from this dataset alone. They could reflect changes in sources, answer generation, prompt interpretation, or real changes in each brand’s underlying relevance.

What the data does show is the condition for disruption: Narrow-lead categories are contestable; wide, sustained ownership is much harder to displace.

5. How to compete for category ownership

Use topical ownership to decide where to focus, not to declare that more content will make ChatGPT mention you:

1. Watchlist. List 10 to 20 commercially important categories for your brand. For each, track at least 5 representative prompts: a definition, comparison, alternatives, use case, and buying question. Read more about how to make your prompt tracking more accurate here.

2. Classify. A clear owner is the most-mentioned brand with a strong majority of prompts and a clear lead over the runner-up. Everything else is either an emerging-leader category or an unsettled category.

3. Invest. Prioritize commercially valuable categories with no clear owner, or a small gap between the leader and runner-up. Defend categories you already lead. Be more selective about attacking a category where an established brand leads across nearly every prompt.

4. Gaps. Close gaps across the prompts where competitors appear and you do not. That can mean clearer product positioning, comparison pages, documentation, expert guidance, proof points, and credible third-party coverage. The study does not show which of these creates ownership. It gives you a disciplined place to test them.

5. Measure. Track brand mentions as the ownership metric, citations as a source-authority metric, and referrals, self-reported pipeline, or revenue as the business metric. A monthly change in the most-mentioned brand is a reason to investigate, not proof that a competitive win created demand.

Closing prompt gaps starts with knowing which pages AI systems already rely on. The AI Citation Audit saves a junior analyst about 2 hours of manual checking per cycle and tells you which cited pages are carrying your category. Premium is only $150/year for now. See the full library.

Topic focus is worth the “content constraint” because ownership compounds over time. A category you win in is a category you tend to keep, and the brands that spread their content creation across every topic that’s semi-adjacent to their core offering rarely reach the mention share that makes a lead stick.

The Semrush data shows that, currently, the window is open… and it will not stay open. 89.3% of AI-search demand sits in categories with no owner (as defined by this study), which means the cost of picking the wrong 5 categories today is lower than it will be in 12 months.

Pick the categories you want to be the answer to. Defend the ones you already lead. Skip the ones where an established brand appears in every prompt. Broad coverage still has a job in marketing, but it does not build the position that makes ChatGPT name you.

Premium: Topic category defense playbook

Clear topic category owners held first place month over month in 90.4% of the observed comparisons. That makes ownership durable, but not permanent.

This playbook explains how to protect your category lead before a close challenger turns into the new default answer.

Important: In this study, ownership means brand-mention leadership, not citation volume. A category owner appears in at least 4 of 5 representative prompts and leads the runner-up by at least 5 percentage points.

Citations matter because they help shape AI answers. Brand mentions matter because they are what users see and act on.

What this playbook covers

• Protecting a lead in topic categories your brand already owns

• Finding weak prompts before competitors take them

• Strengthening the pages and third-party sources that support AI answers

• Building external proof that makes your brand harder to displace

• Setting an operating cadence for competitive response

This is not a claim that any one tactic creates topical ownership. The study identifies the conditions worth defending. Teams still need to test which actions change mentions in their category.

Read more

Subscribe now

The next mass frontier of AI is agentic:

• Google introduced agentic tasks in Search

• The web gets more visits from bots than humans now

• Salesforce found 20% of sales coming from agents marks a signpost

• 60% of companies use agents live in production, and 3 out of 4 companies invest in AI agents

To figure out how ready B2B sites are for agentic visitors, I teamed up with David Kaufman, founder of Siteline, and I analyzed how agents scan websites and where they get stuck. The answer: most sites are agent-ready, but there is one critical breaking point.

Agents don’t read websites like humans. They receive a task, search the web, fetch pages, extract facts, and cite the sources they used. A page can persuade a human and still fail an agent if the facts are hard to find (opacity), hard to fetch (machine-readability), or hard to cite (access friction).

AI agents turn websites from showrooms into barcodes.

Methodology

How we looked at agent behavior:

• The agent had to find the official site itself. We did not provide starting links, eg to homepages.

• We gave agents 3 buyer-related tasks for 100 B2B products: find pricing/features, integrations, and security/compliance. We ran each task 5 times to measure the impact of the probabilistic nature of LLMs.

• We weren’t comparing whether or not the information existed somewhere on the web; instead, we measured whether the agent could reliably answer from the vendor’s own site.

Everyone’s bracing for Google Zero. Cyrus Shepard studied 400+ sites and found the ones still growing share five specific traits.

Zero-click search is real. The “SEO is dead” takes are not the full story.

Cyrus mapped what’s actually driving organic growth right now, plus the 17 content types most likely to survive Google Zero. We’re hosting him for a practical session on what every content team is trying to decide: what to keep investing in, what to rethink, and where AI actually helps.

You’ll walk away with:

• The five traits behind sites still growing in Google

• Which content formats still earn their keep, and which are getting harder to justify

• How to turn proprietary assets you already own into content

• Where AI fits now: what to stop producing faster, what to build instead

Built for SEO and content leaders, content directors, and B2B growth teams.

July 23, 3pm ET. Live on Zoom, recording sent to everyone who registers.

Save your spot

Featuring

1/ Pricing breaks first-party sites

The moment a prospect looks at pricing, they stop browsing and start comparing. #High buyer intent, bottom of the funnel. That makes pricing the hardest and most important test of whether a vendor site can serve agents directly.

Pricing also sits in a triangle of 3 “wants” that good pricing pages need to satisfy:

• Companies want to control pricing disclosure.

• Buyers want fast comparison.

• Agents need clear, fetchable, citable facts.

When AI agents try to retrieve pricing, they get stuck much more than for security or integrations.

• Pricing/features: 79% first-party answer rate, 84% first-party citation share

• Integrations: 93% and 99%

• Security: 92% and 99%

• Pricing/features produced 77% of all third-party citations

If you wonder whether that’s because some B2B companies don’t publicly show pricing, you’re only half right.

2/ Hidden pricing is only part of it

Hiding prices forces agents to look elsewhere, but published prices do not fully solve the problem. Among pricing prompt runs where the vendor did not disclose a real price, 45% cited at least one third-party source. The other 55% stayed on first-party citations, usually by saying the vendor required contact sales or did not publish a concrete price.

Even when the vendor showed a numeric public price, agents still cited at least one third-party source in 18% of runs, suggesting price can be on the page but still be hard for the agent to extract, trust, or cite cleanly.

You can try to hide your pricing, but you better make sure no one else knows and writes about it. Once it’s “out there”, it’s too late. If you have complex pricing methodology, the best way is to explain it clearly and make it accessible to agents.

Some pricing pages are visible to humans but not reliable enough for agents to parse and cite. You can’t always trust your eyes.

3/ Agents fail for 3 reasons

Agents fail to retrieve pricing from a brand for 3 reasons: opacity, machine-readability, and access friction.

1. Pricing opacity simply means the brand doesn’t publicly disclose the price, or it’s vaguely packaged. Opacity explains elevated fallback, meaning agents have to rely on 3rd parties for information.

2. Machine-readability describes the situation when prices exist, but agents still do not confidently extract them. Machine-readability explains fallback despite disclosed pricing. Machine-readability fails when the price is hard to extract because of page structure, JavaScript, calculators, toggles, screenshots, PDFs, or ambiguous tables.

3. Access friction is what most people expect to be the problem with agents. The agent hits fetch failures, rate limits, blocking, or unreachable pages, making agent runs more costly.

Access errors were not the main reason agents left first-party sources, but when they happened, they were severe. They appeared in only 7% of all runs. In pricing runs, access errors pushed third-party fallback to 77%, compared to 17% without access errors.

The impact of errors on agent run cost (tokens, web searches, fetches, retrieves, time) is significant when comparing the 90th with the 10th percentile in our study:

• Cost: 4.4x

• Token: 4.7x

• Time: 2.0x

Brands don’t pay that bill directly, but it is a useful proxy for friction. The harder your site is to retrieve, the more work an agent has to do before it can answer from your page. If your pricing page is blocked, slow, hard to fetch, or hard to parse, the agent has two choices: spend more work on your site… or get the answer somewhere else.

4/ The fallback web is messy

Fallback is when agents have to rely on 3rd party sources as opposed to 1st party, as a result of the 3 failure modes. This is the biggest risk because 3rd party information is spotty, and it’s not in your control.

Agents do not fall back to one clean source category. They reconstruct pricing from a mixed web of explainers, directories, app stores, partner pages, and low-trust aggregators.

Key stats from the 580 pricing third-party citations:

• 52% were editorial (blogs, media articles, comparison guides, explainers, and other article-style pages)

• 46% fell into the directory category (review, procurement, and software-listing sites such as G2, Capterra, Vendr, Tekpon, and similar domains)

• 2% from broader ecosystem pages (app stores, marketplaces, partner pages, and integration directories tied to another platform)

The examples show the risk of missing pricing transparency and agent stumbling blocks on your site.

Example journey:

5/ How to make your site agent-proof

An agent-proof pricing page is how you keep the agent quoting you instead of a directory like Vendr. The fixes map to the three failure modes.

Disclose the fact (opacity)

• Publish real prices in text for every self-serve tier. If a tier is genuinely custom, say what drives the number instead of “contact sales.”

• Keep plan names, prices, limits, and features on one canonical pricing URL, and point every other mention back to it.

• Mark legacy plans clearly so third-party content can’t keep stale tiers alive.

Make the fact extractable (machine-readability)

• Put prices in crawlable HTML. Many agent fetches never run JavaScript, so a price rendered client-side is invisible. In testing, prices in server HTML got read in under a second; a JavaScript-only price got missed.

• Add schema.org Product and Offer markup with price and priceCurrency. This single lever moved a page from 73 to 93 in the readiness test.

• Explain usage-based pricing in text, not a calculator-only widget.

Let the agent in (access friction)

• Allow AI crawlers in robots.txt (GPTBot, ClaudeBot, PerplexityBot, Google-Extended). Check you aren’t allowing Googlebot while blocking them.

• Don’t block server-side AI fetches on pricing pages. Access errors hit only 7% of runs, but they push fallback from 17% to 77% when they do.

• Keep the price early in the DOM and the page light. A 1 MB pricing page taxes every agent and pushes cheap runs to route around you.

Fix opacity and machine-readability first; they drive most of the fallback. Then run the query yourself, “Find all pricing and features for [product].”,” and measure it with the skill below.

Premium: Check the agent-readiness of a page

A well-known B2B SaaS pricing page scored 73 out of 100 the first time the skill ran, and 92.5 the second. Same page, same hour. The 20-point jump came entirely from how the page got fetched.

The skill simulates an agent answering a buyer question from one page: Fetch it, pull the fact, measure the cost across five weighted dimensions.

Subscribe now

Read more

Welcome to another Growth Intelligence Brief, where organic growth leaders discover what matters and guidance on how to stay ahead of the competition in under 5 minutes. Data window: SEO visibility and AI prompt counts June 1 to June 29, 2026 (28 days, matched windows).

Exec Summary

The 5 biggest companies by AI mentions all gained SEO visibility and lost AI Overview mentions in the same 4 weeks to June 29.

For 2 quarters, SEO trended down and AI up. No mas.

• The 5 biggest companies by AI mentions all gained SEO visibility and lost AI Overview mentions in the same 4 weeks to June 29. Their AI Mode mentions went the other way and grew.

• LinkedIn gained 43.3% in visibility over the 4 weeks to June 29 (289.4 to 414.7), while AI mentions held flat at -2.0%. The gain is all organic, and it still leaves LinkedIn 9.1% below its April 2024 peak of 456.3.

• Barnes & Noble lost 55.4% of its visibility in 4 weeks (24.76 to 11.04), the steepest drop of any tracked company with a real base.

• AI mentions across the index have been flat for 8 weeks near 6.0 million per week. That flat number hides a rotation between AI surfaces that reverses how the divergence reads.

• Home improvement bucked the flatline with 25% to 50% AI-mention growth in 4 weeks, while the same cluster’s organic visibility stayed flat. Summer project demand is showing up inside AI shopping answers now, not in 2027.

Let’s dig in.

SEO landscape check

The biggest platforms got their visibility back. Reddit, Facebook, Instagram, LinkedIn, and YouTube each gained between 11% and 43% in SEO visibility over the 4 weeks to June 29, and and each lost AI Overview mentions over the same period. GIB 20 flagged Reddit declining on both dimensions for the first time. Half of that fall reversed inside 3 weeks.

Google’s May core update finished rolling out on June 2, and the volatility ran through mid-June (Search Engine Land), reshuffling organic shelf space back toward the largest platforms while AI Overviews moved the other way.

The platforms clawed back their visibility

Social Networks gained 21.1% in visibility as a vertical over the 4 weeks to June 29, the third-strongest move in the index. And quite coincidental with Google’s announcement of organic social visibility tracking in Google Search Console. The recovery concentrates on the names that got hit in GIB 17 and GIB 20.

Reddit is the clearest case.

• Its visibility bottomed at 2,297.6 on June 8, climbed to 3,041.2 by June 22, then settled at 2,765.7, up 18.1% on the 4-week window.

• GIB 20 measured Reddit down 11.7% through June 8. The decline reversed inside 3 weeks.

The rest of the vertical moved in lockstep.

• Facebook gained 20.1% in visibility

• Instagram gained 21.4%

• LinkedIn gained 43.3%

• X gained 38.6%

All of this occurred over the 4 weeks, ending June 29. The jump clustered in the June 15 to June 22 window for every one of them, which is the fingerprint of an algorithm change rather than organic demand.

But their AI Overview mentions went the opposite way:

• Reddit lost 12.8%

• Facebook dropped 12.1%

• TikTok lost 12.1%

• Instagram decreased 9.7%

The platforms Google just rewarded in organic are the same ones AI Overviews cited less often in June. That split is the story of the month. SEO visibility and AI citation used to drift together for the mega-platforms. In June, they came apart.

Movers & Shakers

Top 5 Movers (4-week SEO)

1. Yellow Pages (Local Search) +66.3%. Vis 14.52 to 24.13, an all-time index high. 2nd consecutive issue on this list after +137.2% in GIB 20‘s window. AI mentions stayed flat at +0.8%, so Google routing is doing all the work.

2. Trustpilot (Ecommerce) +54.3%. Vis 63.51 to 97.96, also an all-time index high, adding 34.5 raw visibility points. GIB 20 asked whether the local routing favors Yelp specifically or review aggregators broadly. Trustpilot answers that: broadly.

3. Ace Hardware (Ecommerce) +43.9%. Vis 9.66 to 13.90, with AI mentions up 45.8% over the same window. Both dimensions point the same way, and it sits inside the home-improvement cluster below.

4. LinkedIn (Social Networks) +43.3%. Vis 289.43 to 414.74, adding 125.3 raw visibility points. Full breakdown in the Winner callout below.

5. Bed Bath & Beyond (Ecommerce) +41.6%. Vis 13.44 to 19.03, with AI mentions up 23.3% in the same window. Ecommerce as a vertical gained 4.3% in 4 weeks, so this is domain-specific.

Top 5 Shakers (4-week SEO)

1. Barnes & Noble (Ecommerce) -55.4%. Vis 24.76 to 11.04. 2nd consecutive issue on this list after -32.0% in GIB 20‘s window. Full breakdown in the Loser callout below.

2. Alibaba (Ecommerce) -40.6%. Vis 7.05 to 4.18. AI mentions stayed flat at +0.9%, so nothing is absorbing the loss.

3. Shutterfly (Images) -34.4%. Vis 6.90 to 4.53. A full round trip: GIB 20 ran Shutterfly as its #2 mover at +88.9%, and it now sits back below the 5.98 where the April slide began.

4. Angi (Local Search) -29.5%. Vis 4.84 to 3.41 while AI mentions grew 78.9%, the sharpest split in the index. Angi has moved against its own vertical for 2 straight issues, and Local Search gained 22.8% in 4 weeks.

5. VRBO (Travel) -25.8%. Vis 6.92 to 5.13, with Airbnb down 23.9% over the same 4 weeks. The vacation rental squeeze GIB 20 flagged has not let up.

Winner: LinkedIn

LinkedIn gained 43.3% in visibility over the 4 weeks to June 29, from 289.4 to 414.7. It climbed every week after June 8: 280.7, then 335.7, then 398.4, then 414.7.

AI mentions held roughly flat at -2.0% over the same window, so the gain is all organic. LinkedIn still sits 9.1% below its April 2024 peak of 456.3, which makes this a recovery rather than a breakout. Its user-generated pages are exactly the large-platform real estate the June reshuffle favored.

Loser: Barnes & Noble

Barnes & Noble lost 55.4% of its visibility over the 4 weeks to June 29, from 24.76 to 11.04, the steepest drop of any tracked company with a real base.

Mid-tier retail took the other side of the platform trade: As Google handed shelf space to the largest sites, names like Barnes & Noble, Kohl’s (-20.1% in 4 weeks), and JCPenney (-15.2% in 4 weeks) gave it up. Macy’s, the surprise winner from GIB 16, slid another 11.9% in June after the bulk of its round trip landed in GIB 20.

The flat AI line hides a surface rotation

AI mentions across the index have been flat for 8 weeks. The summed total for AI Overviews and AI Mode has held in a band around 6.0 million per week since late April, with June 29 at 5.96 million. The explosive growth that defined GIB 15 through GIB 17 stopped at the aggregate level.

That flat line is 3 trends stacked on top of each other. Over the 4 weeks to June 29:

• AI Overviews fell 3.2%, from 5.47 million to 5.29 million

• AI Mode grew 22.5%, from 547,193 to 670,163

• ChatGPT mentions fell 28.4%, from 1.20 million to 860,743

AI Overviews carries roughly 8x the volume of AI Mode, so the sum barely moves while the mix underneath rotates.

The platform decline is mostly AI Overviews. Total AI Overview mentions fell only 3.2% while the 5 biggest platforms fell 10% to 17%, so the platforms lost share within the surface.

Commercial sources took it: Yelp’s AI Overview mentions grew 14.9%, Home Depot’s 21.5%, and Ace Hardware’s 35.1% over the same 4 weeks. AI Mode coverage held steady near 342 tracked companies all month, so the AI Mode growth reflects Google citing more sources rather than the index tracking more companies.

The signature divergence trade went quiet with it. Screening for companies with at least 2,000 AI mentions at baseline, AI mentions up more than 10%, and visibility down more than 5% over the 4 weeks to June 29 returns 5 names:

• Zapier

• Office Depot

• GetYourGuide

• Nordstrom

• Hilton

All sit on small bases. Angi runs a sharper split than any of them and falls just under that mention floor.

Keep in mind what flat means here: AI engines still mention these 2,600 companies about 6 million times a week. The growth curve bent, but the level held.

Home improvement is where AI demand is still rising

One cluster bucked the plateau. Home-improvement and big-box retail posted 25% to 50% AI-mention growth over the 4 weeks to June 29:

• Wayfair +48.1%

• Ace Hardware +45.8%

• Lowe’s +32.8%

• Home Depot +30.9%

• Best Buy +28.8%

• Michaels +29.2%

Ecommerce was the second-fastest vertical for AI mentions at +10.9% over the same window.

Organic search never moved. The same cluster’s combined visibility fell 0.7% over the identical 4 weeks, and June has been flat for these companies every year the index has tracked them: 0.0% in 2023, -2.8% in 2024, and +0.9% in 2025. Seasonal project demand is real, and it only shows up in AI surfaces.

The caveat sits in the data window: AI mention tracking began in February 2026, so June 2027 is the first chance to confirm the pattern repeats. If it does, AI surfaces reprice seasonal commercial intent faster than the organic index does, and AI shopping answers become a seasonal channel a growth team can plan against.

Double Click: Google runs two rankers

Read more

Part 1 tackled those all-important third-party citation signals, while Part 2 made the case for publishing original data: It is the strongest single predictor of page originality, and the bar to earning visibility/authority via this play is low.

This memo has more ammo to back up your use of proprietary data in content creation.

Publishing the number is necessary. But it’s not always what gets cited. We pulled Gauge’s citation data to find out what AI actually rewards when it comes to publishing first-party data, and the answer is narrower and more useful than “original data wins.” (Although original data does, in fact, win.)

AI rewards one format almost to the exclusion of everything else: The benchmark that answers “which is best.”

Build your strategy with the definitive study into AI visibility

AI visibility isn’t uniform. Performance diverges sharply across platforms and industries, meaning a strategy built for one environment can fail to perform in another.

Semrush’s AI Visibility Index tracks 126 million prompts across 22 industries and ChatGPT, Google AI Mode, Gemini, and Google AI Overviews. It’s our biggest ever piece of research into AI visibility, complete with playbooks, expert guidance, and deep dives into the winning brands.

Build a strategy that wins across where your audiences search. Start with the AI Visibility Index.

Download the complete report

First-party research is scarce and punches above its weight

We worked from Gauge’s cited-URL set: 301 live pages that AI systems cited (316 unique prompts across 7 verticals), carrying 1,075 citations between them.

After a full URL audit, only 8 of those 301 pages qualified as primary research, meaning the original source of the data and methodology are on the page… rather than a writeup of someone else’s numbers.

8 pages out of 301 is 2.7% of the set. Those same 8 pages earned 90 of 1,075 citations, or 8.4% of citation volume. First-party research shows up rarely, then over-indexes 3x on citation share when it does.

The cleaner way to see it is density.

Primary research averaged 11.3 citations per page. Everything else averaged 3.4. A primary-research page was 3.3x as citation-dense as a non-primary one.

Primary research compounds citations.

This is the same shape as the information gain finding discussed in Part 2, viewed from the AI side instead of the classic 10 blue links side.

There, original data correlated with page originality more than any other trait. Here, original data correlates with citation density. Both point the same direction: The number only you can produce is the lever.

Original research wins when the question has a benchmark

Here’s where the “original data wins” filter gets sharper.

The 90 primary-research citations are not spread across the 8 pages evenly, and they are not spread across topics evenly.

75 of the 90 came from one cluster: cloud data warehouse benchmarks. Fivetran’s warehouse benchmark alone took 44 citations, just under half of every primary-research citation in the set. (More on that below.)

Reality: Strip the benchmark cluster out and first-party research barely registers in the citation set. The win is not “we published original data.”

The win is “we published a benchmark that answers a buying comparison,” and almost nobody builds one. (“Benchmark” meaning you measure a set of named things against each other on a specific yardstick, and publish the results as numbers.)

Original research is most effective when it is packaged in a way that directly answers commercial comparison queries.

This is what Google is after with non-commodity content: new and helpful information that is hard to get.

Primary-research citations clustered where the prompt asked AI to compare options on measurable specs: speed, cost, latency, yield, or performance.

That explains the warehouse benchmark spike. The “HR Tech / Compensation” label is noisy, but the citations inside that bucket mostly came from cloud data warehouse benchmark prompts. Fivetran, Estuary, and ClickHouse had numbers AI could use.

Crypto / Solana shows the same pattern at a smaller scale. Marinade and Helius earned citations because staking and MEV questions need first-hand ecosystem data, not generic explainers.

The pattern disappears in topics without a clear benchmark. B2B SaaS / CRM, Education / TEFL, and Product Analytics returned listicles, product pages, explainers, and case studies. After cleaning, none of those topics produced a cited primary-research page.

A closer look at the content that held 44 of the citations

Fivetran’s warehouse benchmark took 44 of this data set’s citations on its own, and Fivetran’s 2 benchmark pages together took 58 of the 90 primary-research citations in the set. Why?

It’s a piece of content from 2022, but when you examine it, it’s easy to see why LLMs prefer it.

1/ It answers a measurable comparison head-on. Named warehouses, BigQuery, Redshift, Snowflake, and Databricks, ranked on speed and cost. It’s entity rich and not afraid to name all the major players.

2/ It runs on real first-party data. Fivetran tested against actual customer usage rather than synthetic assumptions, and called out that choice directly.

3/ It shows the method, step by step. Trust signals. Separate sections walk through what data they queried, what queries they ran, and how they configured and tuned each warehouse. A reader (or a model) can see exactly how the numbers were produced.

4/ The structure is built to be lifted. Descriptive headings (”Results,” “How much did performance improve?,” “Why are our results different from previous benchmarks?”) let AI map a question to the one passage that answers it.

5/ It links to its raw data and sources. The page footnotes its references, including the C-Store paper, and points to the underlying data, so every claim is verifiable. Not many brands put this much work into a data-backed content piece, let alone offer the full data set for transparency.

6/ It shows its seams. Dated correction notes from December 2022, named qualitative limits, and an honest “performance floor” caveat make the quantitative claims more credible… not less. They also note corrections.

7/ The URL never moved. A 2022 page is still collecting citations in 2026 because it stayed put at one canonical address.

The data behind a page like this is easier to pull and analyze than it has ever been. What is not easy is everything around it: the clean method, the linked sources, the corrections, the navigable structure, the willingness to name what the numbers do not prove. That’s craft, and that’s the moat here.

This first-party-data focused piece isn’t a sloppy press release with half-assed pulled data. It took a lot of work, and it’s holding authority 4 years out.The takeaway: AI does not reward “original data” by default. It rewards first-party research when the page gives it a clean answer to a measurable comparison that’s built to signal depth of expertise and trust.

The open opportunity here is to publish a retrievable dataset for a buyer question where AI currently has no clean benchmark source. This maps onto the unanswered-questions finding from Part 2: the open door exists, and in these verticals nobody has walked through it with a real dataset.

Original data needs a citation-ready package

Original data gives a page something AI cannot get from another explainer. But AI still has to retrieve it, parse it, and map it to the question.

That is where many brands lose the citation. They publish proprietary numbers, but bury them in narrative, gate them behind forms, move the URL, or skip the methodology. The data exists. The citation does not.

The pages that won in this dataset had both: original numbers and a clean citation shape. Stable URL. Clear method. Named comparison. Results that answered a buyer question directly.

• Who wins: Brands sitting on proprietary product, usage, or pricing data who package it into a comparison a buyer can act on, one that informs LLM outputs for recommendations.

• Who loses: Brands publishing original numbers buried in narrative, on slow or unstable pages, with no comparison frame for AI to lift.

A citation-ready research page has four parts:

1/ Lead with the comparison result. The headline finding (“X is fastest, Y is cheapest at scale”) goes in the first 30% of the page. Result, then method, then nuance.

2/ Box the methodology. Sample, time window, what was measured, how. Attribution confidence is part of what makes a number citable. Make your methodology clear on the page.

3/ Explicitly frame it as a comparison if it is one. AI reaches for benchmarks on “which is best” prompts. A table that compares named options on named specs is the shape it lifts.

4/ Keep the URL stable. One canonical page, kept live, not migrated or renamed every redesign. The citation you earn this quarter only compounds if the page is still there next quarter. Of 365 cited URLs in this data set, 64 were dead, redirected, or otherwise broken, taking 203 citations down with them.

This is the work behind a citable benchmark, and it is more involved than it looks.

HockeyStack documented its own version in a playbook on launching research reports: they published 18 original reports built entirely on anonymized first-party customer data, the kind no competitor could replicate. Their process names every step the Fivetran page demonstrates: list the data points you need, get a teammate to pull them with SQL, define and document the method so the numbers hold up to scrutiny, then structure the report around a real ICP question. They call methodology non-negotiable for a reason, noting that without it someone will always dispute your data.

With AI analysis, the data is the easy part now. Building the content into something that is citable, demonstrates EEAT, and is still earning visibility 4 years out for commercial queries is where the hard work lies.

What sites are already trusted for your topic? When a benchmark you did not publish is taking the citations in your category, the Citation Source Mapper maps that trusted set into a ranked, pitchable target list. It is in the premium library.

Premium: Benchmark draft builder - Claude skill

Turn the data your product already produces into a benchmark AI cites as the canonical source.

The Benchmark Builder skill fits into your existing Claude workflows and interviews you one question at a time: what data you own, which buying question it answers, which comparison is worth publishing. It pushes back on incomplete comparisons, helping you refine with data points you have already on hand.

Read more

Subscribe now

Publish original numbers. It’s the single most reliable lever for making a page more original, and the most defensible numbers are a byproduct of the business itself… not data you assembled to feed a content calendar.

The old play was paying a PR or research firm for a survey loosely tied to your product, like a car insurance FinTech buying road-trip research to land in Yahoo. That’s outdated. Almost every product now generates data worth publishing, and pulling it has never been easier.

You don’t need a research team. The bar to clear the field is lower than you think.

Meet the AI visibility leaders in retail, plus new tactics

AI search is reshaping customer journeys and how shoppers discover and choose products. But many retail marketers don’t yet have a clear benchmark of AI visibility, who’s winning and why.

Semrush’s AI Visibility Index changes that. Our new study is built on a US prompt database of 126 million prompts, mapping exactly which retailers are dominating AI recommendations across ChatGPT, Google AI Mode, Gemini, and AI Overviews. As well as where the opportunities lie.

Download free study

First-party data: The strongest correlation of originality

On-Page.ai’s recent information gain study scored 150 top-3 Google pages across 50 keywords and 10 verticals on how much each adds beyond the rest of its ranking cohort, grading the contribution from 0 to 100 by meaning rather than wording.

The median page scored 52, and original data correlated with that score more than any other page-level trait, including length.

Pages with at most 1 unique figure averaged an information gain score of 40.2, while pages with 15 or more averaged 62.1, and the score climbed steadily at every step in between

Good news, the bar to beat is low. The study found that it might not take much original evidence to outclass top-visible pages in classic Google search: top organic results typically have only 4 unique data points on average. Publish a page that includes more than 4 real original claims, figures, or answers, and that’s one more lever to pull for increasingly competitive organic visibility.

This analysis also found that in almost every search, there are plenty of adjacent unanswered questions that are being left on the table. On-Page ran their analysis with synthetic reader questions, a sample of plausible questions generated for the study, that were a closely related set to the search topic of each query, and results showed an open door for new pages to answer them and stand out. (Remind you of anything? Query fan-out, perhaps?)

We had a similar finding in an analysis of ChatGPT citations:

“A single evergreen page covering 10+ query intents is worth more in AI citation reach than 10 single-intent pages. The ROI of comprehensive content is front-loaded: one well-built page compounds citation reach over time. The long tail exists, but the top 5% of pages capture a disproportionate share of ongoing citation activity.” - The science of how AI picks its sources

In fact, your brand’s high-intent prompts should be monitored across a journey for this purpose. Turn them into journeys that follow the buyer across the 5 stages from Reasoning Lift: Problem, Exploration, Comparison, Validation, Selection. (Read more about more accurate AI prompt tracking here.) Answer those questions on the page with the knowledge and expertise that only your brand can to stay competitive against this finding in the analysis.

The main takeaway from the findings: Most pages are middling on originality, genuinely original pages are a minority, and scoring high enough to stand out is achievable… without being an extraordinary feat or lift.

The gap in the findings? This study focuses on classic search visibility and rankings (the SEO concept of information gain was birthed out of a Google patent language, after all). It doesn’t take AI citations or mentions into consideration, and there’s no mention of including AI Mode or AI Overviews in the analysis.

Caveat: Being the primary source alone may not win the citation

This is the part most proprietary data-publishing advice skips. Everyone says publish original research. Few test whether AI rewards the brand that originated the number or the page that presents it most readably.

More data analysis is coming next week, but what we do know from analyses we’ve completed at Growth Memo this last year:

• The entity types that predict ChatGPT citations the most are DATE and NUMBER (from The science of what AI actually rewards). High-cited pages are dense with specific entities: a particular methodology, a precise statistic, a named comparison. Even if your proprietary findings are picked up by another source that is cited instead, i those external third-party authority signals are likely to build.

• Entity-richness and balanced sentiment matter (from The science of how AI pays attention). Generic advice is risky and vague, but specific entities are grounded and verifiable. Proprietary data produces, verifies, or validates and creates entity-rich content. (Think: why a feature works to save % of dollars, how many hours saved by clients vs. competitors they worked with before, etc.). Add balanced sentiment into the analysis and explanation of your data, and you’ve got yourself a 2-for-1 tactic.

If the hypothesis that first-party data is crucial in the era of AI search holds, publishing content based on proprietary data is necessary… but it is not sufficient. LLM extraction structure (along with the sites that AI search engines trust for the topic) decides who gets the citation, even when your brand owns the data.

Unfortunately, an aggregator who repackages your benchmark into a cleaner answer-ready page can collect the citation your research earned. (Truth sucks.)

• Who wins: Brands sitting on proprietary product, usage, or pricing data who also structure it for extraction and don’t ignore other organic brand authority building plays. (Learn more in How to build an AI SEO strategy that outlasts tactics.)

• Who loses: Brands producing opinion content any tool can replicate, ignore other crucial ways of building off-site authority, along with primary sources who bury their own numbers in narrative instead of surfacing them.

Whether some verticals reward data content more than others, we do not know yet. The science series found citation signals vary sharply by vertical, so a uniform payoff would be the surprise, but we will not assert a pattern without data.

Head’s up: Reading this data is one thing. Running the training session that changes how your writers write is another. Premium subscribers get the How AI Pays Attention training deck plus the AEO writing checklist. The slides and the exercises are already built.

How to structure data for extraction

Owning the data gets you into the fight for visibility. But how you structure your content could be what wins it.

We analyzed 18,012 verified ChatGPT citations and found a ski-ramp distribution: 44.2% of all citations come from the first 30% of a page. The middle 30-70% earns 31.1%, and content buried deep in a long post is roughly 2.5x less likely to be cited.

The follow-up analysis across 7 verticals sharpened the target: The 10-20% band of a page is where AI reads hardest in every vertical, while the first 10% is typically navigation and intro filler that AI skips. The bottom 10% of any page earns 2.4-4.4% of citations regardless of vertical.

Applied to a data study, the structure of your content writes itself:

1/ Lead with the headline statistic. Your strongest number goes in the first 30% of the page, ideally right after the title block where the 10-20% band begins. Number → comparison → implication, in the first screen.

2/ Define the metric immediately. One sentence on what the number measures and the population it covers. An undefined statistic is harder to lift with confidence.

3/ Box the methodology. Sample size, time window, collection method, in a short labeled block. Attribution confidence is part of what makes a number citable.

4/ Front-load every secondary finding. Findings ranked by strength, strongest first. The 20-paragraph narrative buildup is a human-retention pattern that costs you machine citations.

5/ Skip the suspense close. AI reads like a busy editor, not a patient student. The payoff-at-the-end structure that worked for ultimate guides actively works against extraction.

Premium: 6 first-party data plays by category

You don’t need a research budget to publish original data. Every business produces data it never publishes.

Below are 6 plays, each tied to a type of data your business already produces, so you can match the play to the byproduct you already own.

Subscribe now

Read more

This page collects my 2026 Growth Memo research through June 30, 2026. The focus is AI search, Google, consumer behavior,…

Read more

The biggest positioning mistake in AI marketing is selling your product as a replacement for people. It wins attention now, but costs you trust later.

In this memo:

• Why “substitution positioning” feels good short term and decays your brand long term

• The data showing AI isn’t actually replacing people

• How to better position AI

Getting retrieved by AI search is not the same as getting cited.

Strong rankings and clean pages still lose when AI systems decide which sources to reference. The gap is everything that happens between the prompt, the fan-out, and the final citation.

On June 25, AirOps Head of Research Oshen Davidson breaks down where citation wins actually come from, alongside Melanie Dell’Olio (AI Solutions) and Josh Spilker (Head of Search Marketing).

You’ll walk away knowing:

• Why many retrieved pages never make the final answer

• Which page elements get content parsed and reused: headings, lists, tables, stats, schema

• How to find pages being surfaced but not selected, then prioritize the right fixes

• What to measure beyond rankings: citations, mentions, repeat visibility, refresh timing

Tactical, no fluff, built for teams who want AI visibility that moves pipeline.

Save your spot → From Research to Revenue: 5 AEO Actions to Take Now. June 25, 3:00pm EST.

The cardinal sin of positioning in the AI era is replacement. I call it “substitution positioning.” Short term, it’s yummy. Long term, it plants cavities in your perception.

In January 2026, Anthropic CEO Dario Amodei predicted software engineering jobs to go away in the following 6-12 months. Demand for software engineers is reaching new heights today.

“I think we might be 6 to 12 months away from when the model is doing most, maybe all of what SWEs do end to end.”

— Dario Amodei (source)

In September 2025, OpenAI CEO Sam Altman predicted customer support jobs will go away. Almost immediately after he said that, customer service hiring started outpacing the broader job market.

“I’m confident that a lot of current customer support that happens over a phone or computer, those people will lose their jobs, and that’ll be better done by an AI.”

— Sam Altman (source)

AI companies use fear as a marketing tool. Specifically, the fear of being replaced. They got me, too. I was deeply worried about my future when Opus 4.5 got traction.

However, as long as I see that even Anthropic is hiring copywriters and SEOs, I sleep much better.

Times were more uncertain before AI came on the scene (I’ll spare you the TED talk about unprecedented times). Fear sells. Tapping into our primal fight-or-flight response is a sure way to get attention. And layoffs act as a double agent in this environment: They make companies that underperform or overhired look good (“Look, we don’t need as many people anymore because we’re so innovative!”) and feed the substitution positioning narrative.

But the facts show that recent layoffs are closer to AI washing than AI spring cleaning:

• New York state added an option for companies to indicate when they lay staff off due to AI (technological innovation or automation). In March of this year, over 160 companies filed mass layoffs of ~28,300 workers. Not even one chose AI as the reason. That list includes Amazon and Goldman Sachs.

• Researchers at Yale looked at the CPS, a government survey that tracks employment across the US, over the last 33 months and found no evidence of job displacement from AI. The way AI impact works looks very similar to how computers and the internet changed work.

Stop trying to make “replacement” happen. It’s not going to happen.

— Regina George from Mean Girls in an alternate multiverse

AI is just not there… yet.

Synthetic intelligence in its current form is heavily paradoxical: AI can do some things better than humans and others not even at the level of a chimpanzee.

This phenomenon is called the Jagged Frontier. It’s based on a paper where BCG and Harvard researchers conducted a randomized controlled trial of 758 knowledge workers and found that you get the most (12.2% more tasks, 25.1% faster) out of AI if you know what it’s good and bad at.

That’s why some people are lukewarm on AI and others see it as new salvation.

“Currently, measures of exposure, automation, and augmentation show no sign of being related to changes in employment or unemployment.”

— From “Evaluating the Impact of AI on the Labor Market: Current State of Affairs,” Gimbel et al.

Microsoft came to the same conclusion. In its 2026 Work Trend Index Annual Report, the Windows mother found that ~16% of the 20,000 AI users surveyed are “Frontier Professionals” or the most advanced AI users in the research set who use agents for multi-step workflow systems.

80% of those Frontier Professionals said they produce work they couldn’t have a year ago, compared to 58% across the board. They are not just using AI more, but also know which mode a task calls for. They have the judgment to navigate the Jagged Frontier.

Despite some of the gains you can see from AI, we’re far away from replacing people.

Do you truly trust any AI workflow enough to leave it on full automation without any maintenance or quality assurance?

Look at a company that tried: Klarna publicly bragged its AI did the work of 700 agents and cut headcount ~40%, Then they reversed and rehired humans after the CEO admitted cost-cutting produced “lower quality” and that customers want a human option.

As a result, choosing substitution positioning means you gain short-term attention but sacrifice long-term credibility. Anthropic calls for a public pause on AI development on June 4 and launches Fable 5 on June 9. OpenAI did the same: Altman told the U.S. Senate in May 2023 that AI needs licensing and a global watchdog, then opposed binding rules and lobbied against California’s SB-1047 and mandatory federal pre-launch approval. The words don’t match the actions.

Substitute positioning could work if customers (1) actually wanted the replacement and (2) if the technology was there. Neither are true.

FYI: The Adapting for AI-Based Search deck gives directors a validated, data-backed story to present to executives who keep asking what changed. Find it in the Premium Subscribers Resource Library.

Antagonism

Cost reduction is a much stronger AI-use argument than productivity growth because it hits the P&L immediately. Productivity gains show up later. They build slowly inside a company and even slower across the economy, the same lag we saw with computers and the internet.

However, because replacement positioning goes way beyond just cost-cutting (it’s people-cutting), you also directly antagonize the people you sell to:

• Duolingo’s AI-first memo drew a “torrent of pushback” on social media/TikTok. CEO von Ahn said he “did not expect the amount of blowback” and walked it back, reframing AI as a “tool to accelerate,” not replace contractors

• ~31% of workers have refused to use AI tools at work, often citing job-loss fear

• Pew: 50% of US adults are more concerned than excited about AI in daily life; only 10% more excited than concerned, up from 37% concerned in 2021

• 71% of Americans are afraid of being replaced by AI

• A survey of 1,294 knowledge workers shows those who sense the goal is replacement produce more low-quality “workslop”

• Surveys show that leadership is commonly more excited about AI than the employees who use it.

In April, UBER execs bragged about burning their annual token budget in just 4 months. Today, that’s a massive liability. Token cost is growing despite being heavily subsidized. But soon, it will reach a price point that makes junior employees start to look interesting again.

Entry-level jobs might soon be cheaper than tokens accomplishing the same output, and young people don’t have the same baggage and reflexes that experienced experts have.

What is the replacement for “replacement”?

Instead of trying to replace people, the solution is to position your AI use as enhancement. The opposite of fear-based marketing is aspiration and empowerment.

AI goes both ways: You can reduce the number of people or grow output with the same number of people. And the data shows that productivity gains have a higher ROI than substitution.

In a paper by the National Bureau of Economic Research, Researchers asked 750 executives about the impact of AI on productivity and labor markets. While larger firms have a higher interest in replacing labor cost (people), the highest ROI comes from growing productivity.

The moral of the story? Doing more with what you have works better than firing copywriter Alex because only Alex can judge good copy.

Premium: A Claude skill to assess positioning in your content

Building products has gotten easier, but distribution has not.

When supply explodes, the scarce thing is being the product that actually gets chosen, and positioning (right alongside product quality) is exactly how you get chosen.

Premium subscribers get the Claude skill that performs a quality check on your content’s positioning for any AI-based tooling within your product, offering quality reframes for your team that avoid substitution positioning.

Read more

The common advice is to build authority off your own site through digital PR, mentions, and high-authority links. Grow your third-party mentions, grow your AI visibility, right?

Right instinct. But the source set that an AI relies upon differs for every topic, so your off-property authority-building efforts must also be topic-driven.

This week:

• Why AI trusts a different set of sources for every topic, and what that changes about where you build it

• Why scattered off-property authority wastes budget, and how to aim it

• How to find the exact sources AI cites for your topic and earn your way in

Premium subscribers also get a citation mapper skill for Claude or Codex that helps you build a pitchable list of citations.

Discover how people are really searching, discovering, and engaging online

AI continues to reshape how users discover, engage, and click across search environments. But most teams are still making decisions without a clear picture of what’s actually shifting.

The Datos State of Search Q1 2026 report, created in collaboration with Rand Fishkin, maps real search behavior across millions of users in the US, EU, and UK. AI adoption, regional contrasts, e-commerce discovery, and intent signals are all explored in detail using real clickstream data.

Get your free report

AI builds a different trusted-source set for every topic

AI search engines rebuild which sources to trust around the subject of the question. Ask about invoicing and it leans on one set. Ask about starting a business, and it leans on a mostly different one.

This data is a sample set from an anonymous pool of clients, but it illustrates the gap is large. AI citations have a source-type pattern that follows the specific topic.

In invoicing questions, competitor domains hold 33.5% of what AI cites. In starting-a-business questions, that same source type holds 7%. Same model, two topics, and the kind of source it reaches for nearly flips.

This small sample analysis confirms something I wrote about in Operationalizing your topic-first SEO strategy:

You should be employing a topic-based backlink strategy, too. You don’t just want backlinks - you want links that have authority in your target topics and/or with your audience.

Meanwhile, Video and Social surfaces run on separate mechanics and deserve their own separate play in your visibility mix, cited at ~6.5% across this sample set. YouTube stays an exception across LLMs, and UGC platforms like Reddit behave differently again. But this memo sets the Video / Social / UGC slice aside on purpose and covers the publisher, research, and expert sources that behave like earned media.

So where you build authority depends on the topic you want to win. Lift a PR plan from a neighboring topic or vertical and you’re likely to aim at the wrong sources.

AI trusts entities it already recognizes

AI does not form a fresh opinion about your brand on every query. It reuses the trust already attached to the sources it pulls from, and it favors documents and entities it already associates with authority on that topic. (Which is why, yes, building topical authority still matters.)

Your owned blog/site is one input; it’s a crucial input, but it’s likely one of the weakest. The publications, analysts, experts, competitors, and communities that mention you carry significant weight.

That gap is why 2 brands with identical on-page optimization can get cited at different rates: off-property reputation the model already trusts.

Add to that, a named author with a byline appears to beat the same content published under a brand. We don’t have a clean dataset on this across platforms yet and it’s based on one-off qualitative findings via clients and industry chatter, so treat this as a working belief, not a measured finding.

LinkedIn’s own analysis of AI visibility factors reported that authorship and timestamps tracked with better performance: fresh, expert-authored, clearly time-stamped content earned the fastest visibility and citation gains in their testing.

Our early testing showed meaningful lift in visibility and citations across the topics we focused on, with owned content delivering the fastest and most scalable gains so far…. Publishing authoritative, fresh content improves visibility. LLMs favor content that signals credibility and relevance, authored by real experts, clearly time-stamped, and written in a conversational, insight-driven style on platforms like LinkedIn.

Source: How LinkedIn Is Adapting to AI-Led Discovery

The mechanism is reasonable: A human author with a track record (someone who has written on the topic across other sites, has an active social presence, holds a license, or sits on your executive team) gives the model an entity to attach authority to. A faceless brand post gives it less to anchor on.

Authority pays out in steps, so depth in your topic beats spread

More quality mentions should mean more citations. Authority does not pay evenly though. It pays in steps.

Our analysis of 1,000 domains with Semrush showed this for backlinks: Authority Score is the strongest backlink predictor of AI mentions in the study at 0.65 Pearson, ahead of raw link count, and the curve bends rather than climbing straight.

Why it matters: A “little more” third-party authority in a crowded middle tier probably won’t change how often AI cites you. The thing that changes citation is getting into the top tier of authoritative sources for your topic.

Read alongside the topic finding, the move is clear. Depth in your topic’s top-tier sources beats spread. 3 placements in a top-decile source move you more than a dozen scattered across low-authority sites. That’s where you invest.

One question stays open: whether citation rises smoothly with mentions or trips past a tipping point, a volume of in-topic mentions after which a brand gets cited consistently instead of occasionally. We don’t know yet. More data and time will tell.

Whatever the tipping-point answer turns out to be, authority-building through (1) third parties in your topic via (2) authoritative sites in that topic is not an open-ended ask to your executives. It’s a target you can budget for and measure against.

How to build authority in the sources AI cites for your topic

Not every third-party signal carries equal weight. A research publication picked up across high-authority industry blogs likely moves citation more than a few executive podcasts a year, while YouTube stays an exception across LLMs. (UGC will be tackled in a later post as part of this building authority series.)

Test these moves, in order.

1/ Pick 2 to 3 willing SMEs.

They don’t need an existing audience, and they don’t need to be a founder or on the executive team. They need:

1. Credibility in your topic

2. Deep understanding of your brand and product, and

3. Willingness to publish.

Give them a process (and permission) to develop sharp content with a strong branded point of view geared toward specific personas.

How-to guides and roundups together account for 62.3% of cited source rows (see chart below), so build your SME’s work in those shapes.

2/ Map the set AI already cites, then target the entity, not the logo.

Run your highest-intent prompts through an audience research tool like SparkToro and record which domains, social accounts, YouTube channels, and named authors show up. Do the same across AI search engines, and note overlap or inconsistencies in who gets cited and when.

Then chase the people, not only the publications. Get your SME quoted by the same journalist, booked on the same webinar series, or co-authoring with the author the model and your audience already trust. Co-occurrence with a trusted entity pulls you into the candidate set faster than a standalone post.

3/ Concentrate on the authority tier.

Depth beats spread: 3 placements in a top-decile source move you more than a dozen scattered across low-authority sites. Rank your target set by authority and spend there first. Use tools like Qwoted and HARO to discover opportunities.

4/ Mine nofollow on purpose.

Don’t skip these. The Semrush link study found nofollow pulls almost the same weight as follow for AI mentions (0.509 Spearman vs 0.504). They are easier to earn and the models still count them. Build a list of the nofollow-heavy sources your category links to and pitch them deliberately.

5/ Ship embeddable data under your expert’s name.

Publish original charts and infographics bylined to your SME and let other sites embed them with attribution. One chart can earn citations across dozens of pages you never pitched. The format matters: AI leans hard on answer-ready pages.

6/ Use LinkedIn as one fast lane.

According to their own testing, named-author posts get indexed and surfaced quickly, and first-hand reports in our space describe brands entering AI answers within weeks of consistent publishing under a person rather than a page (sometimes days, depending on a built-in LinkedIn audience). Possible play: Partner with LinkedIn experts and influencers in your area of authority.

Premium: Citation Source Mapper Tool - Build your targeted off-site authority

Below is what that mapping looks like in practice: The top cited domains change by topic, which means the outreach target list should change by topic too.

Read more

Welcome to another Growth Intelligence Brief, where organic growth leaders discover what matters and guidance on how to stay ahead of the competition in under 5 minutes. Data window: SEO visibility and AI prompt counts May 11 to June 8, 2026 (28 days, matched windows). Year-over-year comparisons run against June 7, 2025.

Exec Summary

Yelp’s all-time record, Macy’s round trip, and Reddit’s first decline show the same mechanism: Google reallocates shelf space within categories whose demand hasn’t changed.

• Yelp posted the highest visibility reading in index history (646.5 on June 1), carrying Local Search to +30.5% in 4 weeks, the strongest vertical move in the index.

• Macy’s lost 53.8% of its visibility in 4 weeks, a full round trip of the March gains we flagged in GIB 16. Its AI mentions grew 36.6% over the same window.

• Reddit is declining on both dimensions for the first time since AI tracking began in February: visibility -11.7% and AI mentions -10.9% in 4 weeks. X gained 40.9% over the same window.

• SaaS AI mentions grew 34.5% (like-for-like) while visibility slid another 3.9% in 4 weeks. The divergence GIB 19 flagged kept widening: the visibility decline slowed, but AI mention growth doubled.

• Airbnb (-44.0%) and VRBO (-40.3%) slid into booking season while delivery went the other way: GrubHub +45.1%, DoorDash +43.4%, all in 4 weeks.

Movers & Shakers

Top 5 Movers (4-week SEO)

1. Yellow Pages (Local Search) +137.2%. Vis 6.54 to 15.51. The smallest baseline on this list, and a directory written off a decade ago riding the same local routing shift as Yelp.

2. Shutterfly (Images) +88.9%. Vis 3.83 to 7.23. A full reversal of the slide we flagged in GIB 19 (-32.5% in that issue’s window). It now sits above the 5.98 where that slide began in April.

3. Serious Eats (Publisher) +55.9%. Vis 44.81 to 69.86. The biggest 4-week gainer with a baseline above 40. Publisher as a vertical was flat at -1.3% in 4 weeks, so this is domain-specific.

4. X (Social Networks) +40.9%. Vis 210.93 to 297.27, with 5 consecutive weekly gains since May 4. The other side of the Reddit story. Full breakdown in the Double Click.

5. Yelp (Local Search) +29.7%. Vis 495.71 to 642.78. An all-time index high and 147 raw visibility points added, more than any other domain this month. Full breakdown in the Winner callout below.

Top 5 Shakers (4-week SEO)

1. Macy’s (Ecommerce) -53.8%. Vis 110.20 to 50.93. Full breakdown in the Loser callout below.

2. Twitch (Entertainment) -48.2%. Vis 6.82 to 3.54. Entertainment as a vertical gained 7.1% in 4 weeks, so this is Twitch-specific.

3. Airbnb (Travel) -44.0%. Vis 8.69 to 4.87. 2nd consecutive issue on this list after -24.9% in GIB 19‘s window. The vacation rental story continues in the Bigger Picture below.

4. VRBO (Travel) -40.3%. Vis 11.88 to 7.09. Moving in lockstep with Airbnb, which points to a category squeeze.

5. Barnes & Noble (Ecommerce) -32.0%. Vis 27.87 to 18.96. Part of a broad Ecommerce slide: the vertical fell 10.0% in 4 weeks and 24.0% YoY.

Winner: Yelp (+29.7%)

Yelp hit 646.5 visibility on June 1, the highest reading since SSI tracking began in November 2022, and held 642.8 through June 8. The 4-week gain accounts for nearly the entire Local Search move: 147 of the vertical’s 155-point lift. Yelp’s AI mentions grew 39.1% over the same window (13,479 to 18,743), so both dimensions point the same way.

Local Search leads the index in 4-week growth at +30.5%, and its 168.5% YoY gain trails only the far smaller Events vertical. The catch: Angi moved against the vertical, down 18.5% in visibility, while its AI mentions grew 31.1% in 4 weeks. Google is concentrating local real estate on the biggest review aggregator and starving the rest.

Worth watching whether the routing favors Yelp specifically or review aggregators broadly. Yellow Pages at +137.2% in 4 weeks suggests broadly.

Loser: Macy’s (-53.8%)

Macy’s gave back the gift. Visibility jumped from 48.8 to 102.0 in the single week ending March 16, the cliff that made Macy’s the surprise winner of GIB 16 and GIB 17. Now the round trip: 110.2 on May 11, then 4 straight down weeks to 50.9 on June 8, putting Macy’s back near its February baseline.

The decline outpaces the market around it. As a vertical, Ecommerce fell 10.0% in the same 4 weeks, so Macy’s dropped at 5x the vertical rate. But its AI mentions grew 36.6% over the same window, to 20,613, among the largest retail AI footprints in the index.

Algorithm-gifted visibility is a rental.

Double Click: The Reddit trade unwinds

Read more

By now, you understand that LLMs are probabilistic systems and that AI answers are highly variable. That fact has convinced a lot of people that prompt tracking is extra noise. But discounting prompt tracking as nonsense is the wrong conclusion.

This memo assumes that you’re already:

1/ Operating under the philosophy of persona-based prompt design, argued for in Synthetic Personas for Better Prompt Tracking.

2/ Bought into doing AI SEO / AEO and need a measurement system that actually tracks your progress vs. noise. Check out How Much Can We Influence AI Responses to learn more.

AI Mode is expanding. The Search box just got its biggest AI upgrade in decades. Search agents now crawl the web and answer for users in the background, before anyone reaches your page. If you’re still grading yourself on rankings and traffic, you’re scoring a game that no longer exists.

Chris Long, co-founder of Nectiv, breaks down what changed, what it means for B2B visibility, and how to build the workflows that get you cited in AI search.

You’ll walk away knowing:

• Why the SGE to AI Overviews to AI Mode shift is bigger than it looks

• The I/O 2026 updates that belong on your roadmap now

• How to measure visibility past rankings: citations, prompt coverage, mentions

• What makes content surface in fan-out and decision-stage queries

• The workflows to audit, refresh, and publish content built for AI search

Date: June 17 · 3pm ET · Zoom · Recording sent after

Spots are limited. Claim yours before Search moves again.

The prompt-tracking backlash is only half-right

Prompt tracking critics are not wrong. 5 people running the same prompt get 5 different answers. Within-LLM variance from sampling alone hits 10-34% on identical prompts.

Reporting a point estimate from one run is astrology. Together with AirOps, I looked at 815,000 prompt-page pairs and found that after running the same prompt 3x in ChatGPT, only 2.2% of citations remain.

Every prompt is n = 1. Given that the average prompt is 5x longer than classic search keywords, the chance that 2 people around the world use the same exact prompt is close to 0. We currently don’t have any insight into what users prompt, and we might never get that data (although both Bing and Google are keeping us satiated, for now, by offering some AI-visibility data).

But “probabilistic = unmeasurable” is lazy thinking. Weather is probabilistic. Credit Scores are probabilistic. We still forecast and track them.

Keyword tracking was never as clean as we’d like to remember

Classic keyword tracking was more deterministic, but not as much as you think:

• For local searches, results were personalized by location and device.

• Google re-scores results daily, so every rank tracker reports a position range, not a fixed number.

The industry standardized the sampling, fixed location, clean profile, daily crawl, etc., until the noise disappeared. Prompt tracking needs the same move, applied to a harder problem. An added challenge: Keyword tracking was focused on Google, but now we have tons of engines. As the market consolidates, tracking simplifies.

I’d argue there’s no escaping this either as Google transitions from classic search to AI search. More searches than ever show AI Overviews (AIOs), all while AIOs and AI Mode increasingly merge.

At I/O 2026, Search head Liz Reid said users increasingly ask “longer, more natural-language questions,” and Sundar Pichai described Search as “less about individual queries” and “more like an ongoing conversation” (blog.google[KI1] [KI2] /innovation-and-ai/sundar-pichai-io-2026, May 2026).

Where common prompt tracking breaks

Over the last 2 years, prompt tracking tools multiplied while the methodology behind them stalled. Where’s the innovation?

The common prompt-tracking approach looks something like this:

• Define 25–50 prompts (brand/category/problem split)

• Run each prompt once per platform

• Track daily

• Score for citation, mention, sentiment, position

Here are the problems I see with that approach:

1. Variance: Only 2.3% of citations remain after 3 prompt runs [The Consensus Gap]. One run is a coin flip with the answer hidden.

2. Reasoning: High vs. low reasoning opens an 18 percentage point citation-rate gap and changes how the model searches, with high reasoning firing 4.6x more fan-out queries [Reasoning Lift]. An aggregate score blends 2 different engines into one misleading number.

3. Personalization: Most prompt-tracking is not persona-specific, so it reports generic answers that no one sees.

4. Monthly cadence: SISTRIX tracked 82,619 prompts over 17 weeks and found Google AI Mode replaces 56% of its cited sources every week, while ChatGPT replaces 74%. At that drift, monthly tracking is like checking your bank account once a quarter.

5. Cross-platform aggregation: Blending your ChatGPT + Perplexity + Gemini visibility into one “AI visibility score” is like averaging your Google rank with your Bing rank.

6. Conversations: A single Turn 1 query tells you whether you get mentioned. It says nothing about whether you survive Turn 2 onward, when the user asks about alternatives, pricing, integrations, or risk. AI is a conversational interface, so the journey is the unit of measurement, and a one-shot prompt misses most of it.

7. Context: Pure mention counting with no context treats every appearance as a win. Get named first for “what are the worst CRMs to avoid?” and a mention tracker still records a victory.

So, while we can’t remove AI answer variance, we can run prompts multiple times and measure what parts, brand mentions, and citations of the AI answer remain.

Mirroring follow-up prompts is hard because we don’t know exactly what people will ask, but we can use AI to estimate likely follow-ups, enrich them with real conversation transcripts, and track the follow-ups LLMs suggest inside their own answers. We can also record the attributes a brand gets mentioned with, not only whether it shows up.

What good prompt tracking looks like in practice

Worked example: B2B SaaS, CRM category.

• Prompt set: 40 seed prompts, weighted toward problem prompts where purchase intent lives (12 brand, 12 category, 16 problem).

• Platforms: ChatGPT, Perplexity, Gemini, Google AI Overviews. Tracked separately.

• Run config: 5 reps per prompt per platform, every week.

• Personas: the 28 category and problem prompts are customized for 3 key personas (CFO, IT, marketing).

• Metrics: mention rate (± CI), citation rate (± CI), average position when mentioned (1-5), sentiment, and the attributes attached to each mention.

Level it up by adding the journey layer. A flat list of 40 prompts only measures Turn 1. To measure conversations, build the high-intent prompts into journeys that follow the buyer across the 5 stages from Reasoning Lift: Problem, Exploration, Comparison, Validation, Selection.

Each seed prompt for Turn 1 becomes the “seed prompt,” and each stage adds a natural follow-up prompt on subsequent turns.

For a buyer evaluating CRMs, one journey runs:

1. Problem: “How do I know if my sales team needs a CRM?”

2. Exploration: “What types of CRM software exist for B2B SaaS?”

3. Comparison: “HubSpot vs. Salesforce vs. Pipedrive for a 50-person sales team”

4. Validation: “Is HubSpot worth the price for mid-market B2B?”

5. Selection: “How do I get started with HubSpot Sales Hub?”

Run the full sequence as 1 conversation rather than 5 isolated prompts, and score every turn. The payoff is persistence: in Reasoning Lift, a brand cited at the Problem stage carried all the way to Selection in 4 journeys under high reasoning and in zero under minimal. Persistence is the metric a one-shot tracker can never see.

Scope it so the run volume stays sane. Track all 40 seed prompts at Turn 1 for breadth, and build the 16 problem prompts into full 5-stage journeys for depth.

Insight example: HubSpot is mentioned in 78% ± 6pp of fiproblem prompts on ChatGPT vs. 34% ± 9pp on Perplexity. Perplexity pulls from comparison posts (G2, Capterra); ChatGPT pulls from HubSpot’s own blog plus integration and compliance docs.

Action: invest in integration guides and API docs to win ChatGPT. Invest in G2 review velocity and comparison content to win Perplexity.

The next generation of tracking looks like polling

Prompt tracking won’t become keyword tracking. AI answers are too variable, too personalized, and too dependent on source selection. But that doesn’t make them unmeasurable.

The next iteration of prompt tracking will look less like rank tracking and more like polling: repeated runs, clear sampling rules, confidence intervals, segmented panels, and raw-answer audits.

Until existing tools get there, you can build a simple version yourself.

Premium: How to upgrade your prompt tracking

You now know why single-run, monthly, blended prompt tracking misreports reality. Here is the build to fix it.

Read more

Scale up your brand visibility, everywhere search happens

Semrush for Enterprise empowers businesses to own every layer of brand visibility.

By unifying SEO, AI search, and technical site health into one platform, teams gain comprehensive performance tracking and complete optimization capabilities. Swap busywork for impactful strategy with advanced automations that are backed by the market’s leading search database.

It’s how winning brands ensure they show up wherever their customers search.

Explore Semrush for Enterprise

The new mental model of search intent

Last week, I shared how Eric Van Buskirk of Clickstream Solutions and I analyzed anonymized clickstream data from approximately 846K U.S.-based Google search sessions.

The most significant finding? The time-on-page for a user on the SERP is no longer dependent on search intent when an AIO is present. The AIO compresses search intent behavior to look similar across intent types.

• Old mental model of search intent: Navigational searches are “fast;” informational ones are “slow.” Time-on-SERP follows intent, and SERPs without an AIO clearly show this pattern (similar to classic search pre-AI outputs), demonstrated by 12% of navigational searches vs. 32% of local searchers still being on the SERP after 21 seconds.

• New mental model of search intent: There is barely a difference in how long users spend time on SERPs between user intents when there is an AIO present. 42-48.5% of users are still on the SERP after 21 seconds across all the 5 major intents.

From The same user behaves differently in AIOs vs. AI Mode (bold added):

At 21 seconds into a session without an AIO, only 12% of navigational searchers are still on the page. 32% of local searchers are. In classic search, time-on-page has always followed intent: navigational users leave fast because they know where they’re going, local users stay because the SERP is dense with maps and listings, informational users fall somewhere in between.

With an AIO present, the spread compresses to barely 6 points. All five intent types (informational, local, navigational, transactional, video) cluster between 41.9% and 48.5% time-on-page at 21 seconds.

Notice how much longer average SERP sessions are - almost 4x! So, we can conclude that AI Overviews don’t just compress user intent but also prolong the time users spend with search results.

The reason? Additional context. Direct answers from the AIOs provide more information and take longer to read. The intent behind the initial query matters less.

This is the gap between links and answers in the new AIO-filled SERP. In the past, giving users a list of (ten blue) links meant the user was responsible for verifying accuracy and finding the information after the click through. Therefore, Google gets user feedback from their behavior.

But when Google (or other LLMs) gives the answer directly, that onus is on the answer engine.

Bing’s blog Evolving the role of the index brings this to a point:

Grounding an AI–generated answer introduces a fundamentally different constraint: The system is no longer just pointing to information, it is using it. The goal shifts from “fetch the best documents” to “fetch the best information to synthesize into a reliable, verifiable answer.”

Lastly, it also means there is utility in tracking branded prompts more diligently and making sure LLMs return the desired information about a brand. Just like companies bid on their brand as a defense mechanism, they should monitor branded prompts, not just product- or painpoint-related ones.

Why this matters

For 20 years, what you searched told Google and SEOs how you’d behave. Type a brand name (navigational search), and you’re in and out in seconds. Search “best CRM for startups” (comparison search), and you settle into a set of comparison pages. Intent sorted everyone.

The AIO erased that tell. By dropping a block of answer text at the top of the page, it pulls every searcher into a reading session, no matter why they came. The brand-name searcher reads the AIO. The product researcher reads another. Both slow down, both stay, both behave alike on the SERP page. That flattening is the intent compression.

Most Google users never chose this, because most Google users are not AI early adopters. They meet AI through Google’s search results as Google forces guides them into AIOs and AI Mode at the top of the results. AI is changing its search behavior passively and without searchers’ explicit consent. Many Google users might not even realize they’re using AI, and that’s one reason we’re seeing growing installs of DuckDuckGo.

Google reports more than 1.5 billion people use AI Overviews, so this is not an edge case. It’s how the web is searching now.

More than 1.5 billion users around the world use AI Overviews for help with their questions. Source: Google

What winning the second impression looks like

The AI Overviews vs. AI Mode behavior analysis revealed the significance and optimization opportunity of the second look. The second impression is what searchers see on the back-scroll, after they’ve already passed your listing once. It’s like a double-take made by a grocery shopper in the cereal aisle who has dozens of options: The shopper scans every box in view, then circles back to reread the one that caught their eye.

Metadata is the trigger for selecting search results. Rich snippets catch attention early on, but they might not be enough to convert users to a click, especially if new search behavior has shifted to include a thorough reading of the SERP and a second scroll up. What can earn the click is what shows up next to your listing on that second pass, and it has to be relevant and trustworthy.

Different page types require different relevance and trust indicators.

Product detail pages (PDP)

Users compare star ratings, review count, price, and stock status.

3 things to control outside the meta:

1. Product schema with aggregateRating, review, offers, and availability. Miss any one of these, and a competitor’s listing renders fuller than yours.

2. Review count is a comparison field. 47 reviews next to a competitor’s 2,300 loses on the second pass even if your description is sharper. Review velocity is a competitive moat.

3. Multiple images in the schema array so Google has options for different SERP layouts.

Category detail pages (CDP)

Category pages compete with the AIO’s own list. If the AIO already enumerated 5 options, your category page has to look like the place where the user goes to actually choose between them.

3 things to control outside the meta:

1. ItemList schema on the category page so Google can render product carousels in the SERP. A carousel takes more vertical space than a single listing and dominates the back-scroll.

2. Filter and sort UI visible in the SERP preview. Google sometimes surfaces sitelinks for category facets (”by price” or “by brand”). Internal linking to those facets makes them eligible.

3. Page count and depth. A category page with 12 products competes badly against one with 240. The second impression carries an implicit “Is this comprehensive?” check.

Blog content

The AIO has already given the user the answer. To earn a possible “validation click”, or at least a thorough second impression for your brand, what the user is looking for is credibility on who said it and when.

2 things to control outside the meta:

1. Visible datePublished or dateModified in the SERP. A 2024 article next to a 2026 article likely loses, whether the user considers the description or not.

2. Article schema with a named author field that links to a sameAs URL (LinkedIn, author bio page). This makes the author an entity Google can resolve, which matters for E-E-A-T scoring even if no visible card renders.

What intent compression means for operators

The last 3 years of intent-based content work produced the right pages for the right queries. What’s shifting is one prediction layer on top of that strategy: how long users stay, where they look, when they click.

That layer is now AIO-driven, not intent-driven. User search intent still drives what your brand needs to write, but it is no longer a good estimate of how users will behave on a SERP page.

Therefore, more optimization work moves to the SERP, focusing on how your listing reads against the AIO and the results around it, separate from the intent logic that decides what the page content should be to answer the query.

Your core optimization efforts don’t change. A content team built around intent clusters keeps its cluster map intact. What gets updated is the optimization pattern per page (meta descriptions, title tags, the second-impression framing from finding 7 shared last week with Premium subscribers), not the underlying taxonomy or content strategy.

(Premium-only) Claude skill for identifying 3+ competitor words at scale

Last week, I instructed Premium Growth Memo readers on how to optimize for more competitive metadata within the SERP.

Today, I’ve included a Claude skill that’s ready to do that for you.

Also included: A short helpful explainer to educate your stakeholders on the user-behavior changes we found across the AIO-based SERP.

Read more

Exec Summary

Three things moved this week. The SaaS slide we celebrated as “finally turning” in GIB 18 reversed inside 30 days. The Jobs vertical is having its strongest month since I started tracking. Local Search posted +44% on the back of a single domain. Each of these tells a different story about where Google is routing traffic right now.

• SaaS visibility dropped -7.8% in 30 days, erasing last issue’s rebound. AWS (-22.7%), Oracle (-17.0%), Stripe (-16.8%), Atlassian (-15.6%), and GitHub (-14.2%) all fell in lockstep while their AI mentions grew 25 to 90%.

• The Jobs vertical posted +21.2%, with Glassdoor (+64.2%) and ZipRecruiter (+53.5%) doing nearly all the work.

• Local Search rose +44.0%, almost entirely on Yelp, which gained more raw visibility points (151) than any other site in the Index this month.

• SaaS AI prompt counts grew +16.9% over the matching window. The divergence economy has resumed at a wider gap than any prior reading.

• Trustpilot (+39.5%) and Glassdoor (+64.2%) extend the GIB 18 thesis: review aggregators continue to absorb traffic that used to go to original sources.

Movers & Shakers

Top 5 Movers (30-day SEO)

1. Glassdoor (Jobs) +64.2%. Vis 14.31 to 23.49. The single biggest 30-day gainer with a meaningful baseline. Full breakdown in the Winner callout below.

2. ZipRecruiter (Jobs) +53.5%. Vis 40.09 to 61.56. Running the same play as Glassdoor in parallel, which is what makes the Jobs story a routing story, not a single-domain win.

3. Nextdoor (Social Networks) +43.2%. Vis 20.14 to 28.83. Third consecutive month of double-digit gains. Local community signal keeps compounding while Threads and TikTok lose ground.

4. Yelp (Local Search) +42.7%. Vis 354.60 to 505.91. Single-handedly responsible for the entire vertical’s +44.0% lift. Yelp added 151 raw visibility points, more than any other site in the Index.

5. Trustpilot (Ecommerce) +39.5%. Vis 39.79 to 55.53. Third review aggregator in three issues to make the top risers, following G2 and Capterra in GIB 18.

Top 5 Shakers (30-day SEO)

1. Shutterfly (Images) -32.5%. Vis 5.98 to 4.04. The slide flagged in GIB 18 continues. AI mentions are up, but the SEO loss is steeper than the AI gain.

2. Costco (Ecommerce) -24.9%. Vis 57.32 to 43.05. Surprise loser this month. Worth watching whether this is a one-month dip or the start of something structural.

3. Airbnb (Travel) -24.9%. Vis 7.23 to 5.43. The Travel vertical itself is only -1.5%, so this is Airbnb-specific, not category-wide.

4. Substack (Content Platform) -22.0%. Vis 42.56 to 33.22. Notable one for the readers of this brief.

5. AWS (SaaS) -22.7%. Vis 10.40 to 8.04. Biggest SaaS faller of the month. Full breakdown in the Loser callout below.

Winner: Glassdoor (+64.2%)

The single biggest 30-day SEO gainer for any company with a meaningful baseline. Glassdoor went from 14.31 visibility to 23.49, with ZipRecruiter (+53.5%) running the same play in parallel. The Jobs vertical as a whole posted +21.2%, and those two aggregators account for most of the lift. Smaller job boards and individual employer pages barely moved.

The likely driver is Google routing more “[company] reviews” and “[role] salary” queries back to the big aggregators after a brief period of generating those answers directly inside AIO. This looks like the same pattern we tracked with G2 and Capterra absorbing SaaS comparison queries in GIB 18, only now applied to employer research.

Worth watching whether Indeed (+13.4% this month, much smaller move) catches up by next issue, or whether Glassdoor specifically is getting favored by the routing change.

Loser: AWS (-22.7%)

The biggest SaaS bellwether had its worst month in the Index history. AWS visibility dropped from 10.40 to 8.04 in 30 days, while its AI mentions for the same window grew only modestly. The product isn’t the story. The story is what AWS shares with Oracle (-17.0%), Stripe (-16.8%), Atlassian (-15.6%), Square (-17.9%), and GitHub (-14.2%): every major SaaS brand is bleeding visibility at similar rates this month.

What makes the pattern interesting is the diversity of products. Cloud infrastructure, payments, dev tools, project management, all moving in the same direction within the same window.

Double Click: The SaaS slide resumed at a wider gap

Read more

The average Netflix user spends 18 minutes browsing the home screen before picking what to watch. They scroll past tiles, hover for trailers, scroll back to a show they almost picked, then circle back to the row they started in. The browse IS the experience.

Eric Van Buskirk of Clickstream Solutions analyzed anonymized clickstream data provided by Surfer SEO, drawing findings from approximately 846,000 U.S.-based Google search sessions collected in February and March 2026.

It’s the fifth user-behavior study on Google’s AI features in the last 12 months. The 70-user UX study from May 2025 used think-aloud and screen recording. The 250-session AI Mode study from October 2025 captured how users behave inside AI Mode itself. This one trades qualitative depth for the scale to find behavioral patterns the smaller studies couldn’t see.

For context: Prior public Google SERP mouse-tracking studies measured dozens of people. The largest had a few thousand tasks. This study analyzed queries from a panel of tens of thousands of Google Search users.

The most significant pattern shown: Users behave in opposite ways in AI Overviews and AI Mode. AI Mode is autoplay. AI Overviews is the Netflix browse.

This article covers 4 findings from the new study (full methodology here) and what they mean for how you write title tags and meta descriptions in 2026.

Complete brand visibility: Whether you’re a global brand or nationwide nonprofit

Good Guys SEO secured buy-in and built an end-to-end search strategy for one of the US’s most recognized nonprofits using Semrush for Enterprise.

The result: +37% AI share of voice for homelessness services, with visibility connected directly to campaign traffic and donations.

Complete SEO and AI search capabilities in one platform. The outcomes speak for themselves, whatever your industry or business model.

Read the full story

1/ Same user, opposite behaviors

In AI Mode, users often accept the answer while continuing to evaluate the SERP in AI Overviews.

The April 2026 study of 185 high-stakes purchases found that in 88% of AI Mode tasks, users took the AI’s shortlist as-is, 74% picked the item ranked first, and 64% clicked nothing at all. AI Mode often behaves like a closed loop: The user reads the answer, picks from inside the answer, and moves on.

New cursor tracking data on AI Overviews shows a different behavior pattern. Cursor positions spread across the equivalent of 83% of a viewport, compared with 66% when no AI Overview appeared. Users kept their cursors still 44% of the time, compared with 29% without an AIO. And in the median session, reverse-direction scrolling accounted for nearly half of scroll movement.

Taken together, this suggests that AI Overviews turn the SERP into more of a comparison environment. The user reads, pauses, weighs, returns to earlier content, and reconsiders before clicking.

The implication for strategy: AI Mode optimization and AI Overview optimization are not the same job. Showing up in an AI Mode shortlist is a visibility problem at the model layer. Showing up in an AIO is a comparison problem on the SERP itself.

2/ Half of all scrolling now goes backward

Among users who reverse direction in an AI Overview SERP, the median user spends 47.5% of total scrolling going back up the page. Without an AIO present, that figure is 27%.

The cursor data shows scrolling isn’t a one-way trip. On AIO SERPs, scrolling is roughly a 50/50 split between going down and going back up. That’s the behavior of someone re-reading, not someone scanning.

The 70-user study from May 2025 found this pattern qualitatively. 38% of AIO sessions in that study showed reassurance-seeking clicks where users opened a second link “just to be sure.” The new clickstream data shows that the validation work has now moved onto the SERP itself. Users used to leave Google to validate a result. Now they validate by reversing back over results they already saw.

This is exactly the Netflix browse pattern. You hover on a tile. You scroll past it. Something pulls you back, so you scroll up to read the description again. The decision happens during the reversal, not the first pass.

For e-commerce and high-consideration decision categories, this is the most consequential finding in the study. Your listing in an AIO SERP isn’t getting one impression anymore. It’s getting 2 or 3, and the second impression is when comparison happens.

3/ Search type no longer predicts behavior

For 2 decades, search intent has been the foundational segmentation framework in SEO. You could predict how long a user would stay on a Google SERP by knowing what kind of search they did. The cursor data shows that’s no longer true when an AI Overview is on the page.

At 21 seconds into a session without an AIO, only 12% of navigational searchers are still on the page. 32% of local searchers are. In classic search, time-on-page has always followed intent: navigational users leave fast because they know where they’re going, local users stay because the SERP is dense with maps and listings, informational users fall somewhere in between. That 20-point spread is what every SEO mental model is built around.

With an AIO present, the spread compresses to barely 6 points. All five intent types (informational, local, navigational, transactional, video) cluster between 41.9% and 48.5% time-on-page at 21 seconds. Intent stops predicting how long users will stay in the search results.

This is the most novel finding in the study. Nothing in prior research has shown AIO collapses time-on-page differences between intent types into a single band. The 70-user UX study segmented by query risk, the 250-session AI Mode study by task type. Both showed intent matters for engagement. The new clickstream data says intent stops predicting how long users stay on the SERP once an AIO appears.

The scoping matters. This is a time-on-page finding, not a “behavior” finding. Scroll depth still varies by intent under AIO, and it actually reshuffles: local jumps from third deepest to first, video falls from first to third. That part of the story is in the next section.

The practical implication is uncomfortable. Most SEO playbooks recommend different optimization patterns by intent. Local pages get one treatment, transactional pages another, informational pages a third. However, our time-on-page data suggests that when an AIO is on the SERP, users stay for similar amounts of time regardless of why they searched.

The main takeaway here is that intent-based segmentation still matters for what content you write, but it matters less for predicting how long users will stick around on the SERP itself.

4/ Brand searches lost their shortcut

Cursor activity for navigational queries (someone typing a brand name into Google) increased 40% when an AIO is on the page. Even users who came knowing where they wanted to go are sweeping the page first.

Without an AIO, navigational searchers were the most focused group in the study. They scored 19.7 on the cursor scatter measure (essentially, how much people bounce around the page), the lowest of any intent type. Only 12% were still active at 21 seconds. They came to Google with a destination, found it, and left.

With an AIO present, that profile changes completely. Cursor scatter for navigational searchers jumps to 27.5. 45.8% are still active at 21 seconds. The brand-name shortcut, the fastest path through Google search for the past 20 years, no longer works the way it used to.

The 70-user study from May 2025 found that brand and authority is the first gate users apply when reading a SERP. They check who’s cited before they check what’s said. The new clickstream data shows that the gate now fires even when there’s nothing to gate against. A user who typed “Lenovo” is still doing the authority check on the AIO content first.

This is the same Netflix pattern again. You open a show you’ve already decided to watch, but you pause first to read the synopsis and check the rating. The brand recall got you to the page. The browse decides whether you click through.

The implication for branded search: Brand recall is no longer enough on its own. Even users who searched for you specifically are now evaluating what’s around you on the SERP before they click through.

Premium: Where AIOs reshape the SERP the most, plus the 3 rewrites that can earn the click

The above data shows how AIOs converted the SERP to a browse environment. The next 3 sections cover the following:

• Cross-study reconciliation (why the 70-user study said one thing about AIO scroll depth and the new clickstream data says another, and how both are right)

• Sectors where the reading-mode shift hits hardest and the one where it doesn’t hit at all, and

• Specific rewrites to make in your title tags and meta descriptions to win in this new environment.

Read more

This Memo was sent to 26,659 subscribers. Welcome to +100 new readers! Upgrade to Premium for the full archive, research, frameworks, and templates

AI offers a conversational experience. We use LLMs through chatbots. But no one has yet looked at how citations and mentions evolve in a conversation.

I analyzed data from the Semrush AI Visibility Toolkit to review 20 buyer journeys across 4 different verticals to compare high vs. low reasoning for ChatGPT5.2.

In this analysis:

• Why high reasoning cites a nearly different web (only 25.6% domain overlap with minimal) and which source types gain or lose ground

• Why TOFU content has a payoff again: Grands cited at the Problem stage are more likely to persist all the way to Selection under high reasoning, and never under minimal

• How to split your prompt tracking by reasoning mode so your AI visibility reporting reflects 2 different systems, not an averaged one

Premium subscribers also get bonus data from the full analysis and the Buyer Journey Prompt Map: a fillable worksheet that turns these findings into a stage-by-stage audit of where your brand persists across an AI buyer journey, and where it goes missing.

Discover how people are really searching, discovering, and engaging online

AI continues to reshape how users discover, engage, and click across search environments. But most teams are still making decisions without a clear picture of what’s actually shifting.

The Datos State of Search Q1 2026 report, created in collaboration with Rand Fishkin, maps real search behavior across millions of users in the US, EU, and UK. AI adoption, regional contrasts, e-commerce discovery, and intent signals are all explored in detail using real clickstream data.

Get your free report

Methodology

Data comes from the Semrush AI Visibility Toolkit, which captures the prompts, citations, and fan-out queries ChatGPT generates per response.

• We ran 100 prompts twice through GPT-5.2, once with minimal reasoning and once with high reasoning, for 200 total responses.

• Prompts span 20 buyer journeys across 4 categories (B2B SaaS, Finance, Consumer Tech, Health/Lifestyle), with 5 stages per journey: Problem, Exploration, Comparison, Validation, Selection.

• Citation rate is the share of prompts where the response cited at least one external source.

• Avg citation counts sources per cited response.

• Fan-out queries are the sub-queries the model fires internally to research the prompt before answering, surfaced via the Semrush API.

GPT 5.2’s high reasoning cites and searches more

Turn high reasoning on, and the citation rate jumps from 50% to 68% (+18 percentage points), the average sources per response nearly doubles (2.6 to 4.5), and fan-out queries go up 4.6x. High reasoning also pulls from 173 unique domains across the test set vs. 127 for minimal; 99 of those domains never appear under minimal reasoning.

This is grounding at its finest. When the model thinks harder, it relies more on web search. Reasoning plays a major role in brand visibility, though we don’t know how many users activate reasoning vs not.

Query intent is a cleaner proxy than user demographics. Free-tier users have reasoning access too, just rate-limited, and ChatGPT auto-routes hard prompts to Thinking mode without the user clicking anything. So the question isn’t who can afford reasoning. It’s which prompts trigger reasoning automatically.

Multi-criteria comparisons, evaluation frameworks, regulatory and compliance questions, and complex shopping builds are the prompts most likely to fire reasoning regardless of plan. Map your audience by query type, not by paywall status.

High reasoning fires more fan-out queries deeper in the funnel

Users move through problem-solving and purchase decisions in stages, often within the same conversation. The gap between minimal and high reasoning isn’t constant. It scales with where the user sits in the journey.

What the 5 stages look like in practice. Take a buyer evaluating CRM software:

1. Problem: “How do I know if my sales team needs a CRM?”

2. Exploration: “What types of CRM software exist for B2B SaaS?”

3. Comparison: “HubSpot vs. Salesforce vs. Pipedrive for a 50-person sales team”

4. Validation: “Is HubSpot worth the price for mid-market B2B?”

5. Selection: “How do I get started with HubSpot Sales Hub?”

The 3 patterns hold across all 20 journeys:

1. Citation rate climbs through the funnel under both modes, but high reasoning closes the early-stage gap most aggressively: +35pp at Problem, only +5pp at Validation. The model treats early-funnel questions as research tasks when high reasoning is on, whereas it answers-from-memory when it’s off.

2. Fan-out queries peak at Comparison. High reasoning fires 24 sub-queries per response there vs. 5.5 for minimal. Selection runs 15.4 vs. 2.6.

3. Average citations per response peaks at Comparison (9.8 high, 5.8 minimal) and narrows at Selection (4.7 high, 2.6 minimal). The model resembles an hourglass across funnel stages.

At the aggregate level, minimal reasoning fires 245 search queries across 100 prompts. High reasoning fires 1,130. When the model operates with high reasoning, it runs a mini investigation per prompt, and most of the investigation happens at the Comparison and Selection stages.

What does a fan-out actually look like? A B2B SaaS prompt under high reasoning comparing Salesforce, HubSpot, and Pipedrive for a 50-person sales team breaks into separate queries about API rate limits per vendor, SOC 2 / ISO 27001 compliance, SAML/SSO/SCIM support, webhook architecture, OAuth flow, developer documentation, enterprise pricing tiers, and change-data-capture support. Each becomes its own retrieval. The brand that wins the answer is the one whose documentation surfaces clean for each sub-query, not the one that ranks for the parent prompt.

The Selection stage has the widest per-response query variance: 0 to 40 fan-out queries on the same five-stage cohort. The driver is prompt specificity. Bounded prompts (like “should I finance through the dealer at 0% APR or use a bank?” or “draft an RFP to 3 SEO agencies”) run zero queries because the answer’s structure is given. Open-ended product builds (”shopping list for a $3,000 home gym” or “which travel card ecosystem fits our grocery spending?”) run 28 to 40 queries. The Selection stage isn’t bounded by one type of question, and the model’s research effort tracks how many degrees of freedom the prompt leaves on the table.

For marketers: Early-funnel visibility is a reasoning-mode story. If your buyers use ChatGPT with reasoning on, problem-stage and exploration-stage content is in play. If they don’t, you’re effectively invisible until Comparison.

Reasoning affects how brands appear in a conversation

An LLM session is a conversation, not a single query. The question that it opens up: Does a brand cited at the start of the journey carry through to the end? If yes, early-funnel visibility compounds. If not, every stage is a fresh fight.

When a brand gets cited in the Problem stage (step 1), does it survive to the Selection stage (step 5)? When using minimal reasoning: No. Zero journeys show this kind of persistence. In high reasoning: Yes. Brand continuity is maintained in 4 journeys across all 5 stages.

Within a single response, high reasoning also anchors harder on individual sources. 51 of 100 high-reasoning responses cite the same domain more than once in the same answer, vs. 26 of 100 for minimal. High reasoning quotes a source repeatedly when it commits to it.

Brand mentions tell a softer version of the same story. If you loosen the test from the cited domain to the brand name in the answer text, persistence shows up in 3 high-reasoning journeys (HubSpot across CRM Selection, American Express across Business Credit Cards, Sony and Canon across Mirrorless Camera) and 2 minimal-reasoning journeys (HubSpot, Mercury). Consumer Tech shows up here even though it doesn’t show up in the citation persistence table. Brands like Sony and Canon are mentioned through the conversation without the model linking out to them, which is its own form of category dominance and worth tracking separately.

High reasoning builds a consistent mental model of the solution space throughout a session. The headline finding: TOFU prompts have value. If a brand shows up at the Problem stage, it tends to carry through to Selection. Top-of-funnel content isn’t just brand awareness for AI visibility. It’s a leading indicator of where the model lands at decision time.

Two more implications:

1. All 4 persistent journeys are in Finance, which suggests persistence rides on the same authoritative-source content (regulatory pages, official brand sites) that drives the +28pp Finance lift overall.

2. For marketers running an account-based or category-creation play, reasoning-mode visibility is the prize. It’s the only mode where early-funnel content compounds into selection-stage citations.

Reasoning mode is a separate search engine

The brand that wins under minimal reasoning is not the brand that wins under high reasoning: 3 in 4 cited domains are different. The mix of source types is different. The stages where citations appear are different.

I’m excited about 2 findings in particular from this analysis:

1/ The first is measurement. We need to track low vs. high reasoning in our prompt trackers. It’s best to avoid an aggregate view because the mechanisms are truly different.

Bad news: This adds more effort and cost to prompt tracking. Good news: We can make prompt tracking a lot more accurate.

2/ The second is funnel stages. In the latest AI Mode user behavior study, I found that users react strongly to shortlists, demonstrating a similar behavior seen with Google’s classic search results where the top result matters most. That result made it seem to me that focusing on BOFU prompts that return shortlists is the game.

However, now we know there is value in TOFU prompts because of persistence: Brands that appear early in the buyer journey can persist all the way through. The best way to find that out for yourself is to map buyer journeys and track your persistence.

For Premium subscribers: winner & loser categories in high reasoning, buyer prompt journey worksheet and how Reddit performs

Most AI visibility strategies treat ChatGPT as a single system. The data says otherwise.

Reasoning mode and minimal mode pull from different corners of the web. Different source types. Different domains. Different winners. If your content strategy is built on the AEO / GEO / AI SEO you saw last year, it’s likely underperforming.

Below, premium subscribers get extra findings from this analysis, including which source types are losing ground, which ones are quietly taking over, and the categories where the gap is widest.

You’ll also get the Buyer Journey Prompt Map, a fillable worksheet for running an audit on your own brand across all 5 stages.

Read more

Welcome to another Growth Intelligence Brief, where organic growth leaders discover what matters, get insights into the bigger picture and guidance on how to stay ahead of the competition.

As a free subscriber, you’re getting the first big story. Premium subscribers get the whole brief.

Subscribe now

Today’s Growth Intelligence Brief went out to 660 marketing leaders.

This week, we’re looking at:

• Microsoft openly redefining what the search index is for in the age of AI answers

• Ramp’s 32-day A/B test that killed the schema-first playbook for AI agents

• Amazon’s LLM-built knowledge graph that quietly added hundreds of millions in revenue

I’ll also connect the dots on what this all means for you.

Microsoft just redefined what the search index is for

Here’s what happened:

Microsoft’s Bing team published a post on May 6 making the most explicit case I’ve seen from a major platform on how AI search inverts the role of the index. The core argument: When an AI system grounds an answer, the index is no longer pointing to information, it’s using it. The unit of value shifts from documents to “groundable information,” discrete facts with clear provenance that a model can responsibly cite.

The post lays out 3 specific shifts.

1. The goal moves from “fetch the best documents” to “fetch the best information to synthesize into a reliable, verifiable answer.”

2. Factual fidelity becomes the key quality metric: Does the indexed representation of a page accurately preserve the meaning of the original content?

3. Freshness failure carries a categorically different cost when the index is grounding an answer instead of surfacing options.

Why this news matters:

This is the first time a major search platform has publicly written down what the new optimization target is. Most of the AI search discussion still treats the problem as “how do I rank in AI Overviews,” which is the wrong frame. Microsoft is saying the right frame is “how do my facts get retrieved into a synthesized answer.”

That changes what counts as a win. A grounding function is accountable for the quality of evidence it provides, not the order of links it returns. The metric is whether the model picks up your specific facts, attributes them correctly, and surfaces your provenance to the user when the user wants to verify.

My take on this:

This is the doc I’ve been waiting for. Volume sits with Google’s surfaces, but the conceptual leadership is coming from Microsoft. Bing is the one writing down the playbook, partly because they have less to lose and partly because they got there first with Copilot integration.

The piece I’d add to Microsoft’s perspective: Grounding has a discoverability problem. A model can only ground against content it has seen, parsed correctly, and stored as retrievable evidence. That’s a 3-step funnel, and most SEO teams are still optimizing for step 0, which is “rank in a SERP.”

The new funnel? Crawl, parse, retrieve. If there’s a break in any of those 3 steps, you don’t get cited.

Here’s what to do:

Audit your content for 3 things this month:

1/ Are your key facts written as distinct, attributable claims rather than buried inside narrative paragraphs?

2/ Do your pages carry author bylines, publication dates, and visible source links that a model can use to assign provenance?

3/ Can a non-Google crawler (try a tool like Firecrawl or run a Cloudflare bot log audit) actually retrieve the parts of your page that matter, or are they hidden behind JavaScript that AI crawlers won’t execute?

The fastest tactical win: Rewrite your top 20 product or category pages so the load-bearing facts (price, feature, comparison, capability) appear in clean text inside the first 600 words. That’s the part the grounding layer will read.

Markdown beats schema for AI agents

Read more

This Memo was sent to 26,564 subscribers. Welcome to +115 new readers! Upgrade to Premium for the full archive, research, frameworks, and templates

Most teams talk about “AI visibility” like it’s one thing. New data on 3.7 million citations across ChatGPT, Perplexity, and Google AI Overviews suggests it isn’t. And the gap between the 3 engines is wider (and more strategically important) than your dashboard likely admits.

Today’s memo breaks down:

• Why a blended AEO score hides the only finding that matters

• Which page types and domains actually travel across engines

• The shift from measuring AI presence to measuring portability

Premium subscribers also get the YouTube exception data, the universal-winners domain list, and 3-step operator guidance for separating engine-level visibility from market-level claims.

I get pitched a lot of AI search tools. Most tell you where you rank. Almost none tell you what to do about it.

AirOps just launched the first serious answer to that gap I’ve come across, and I think it resets the category.

Not a dashboard. Not a writing assistant. A growth platform built for marketers operating in AI search: see where you’re cited, ship the work that closes the gaps, and measure what actually moves.

On May 21, the team is going live with a walkthrough of what they shipped and how to use it.

Live demo with the team building it:

• Amr Shafik (VP Product)

• Eoin Clancy (VP Growth)

• Melanie Dell’Olio (Manager, AI Solutions).

May 21, 3pm ET. Live on Zoom.

Recording goes out after. Live room gets the unfiltered version.

Register now

One of the biggest differences between AEO and SEO is that AEO plays on more platforms.

Omnia data shows across multiple samples that only 2.35% to 2.45% of cited URLs appeared in ChatGPT, Perplexity, and Google AI Overviews for the same prompt. 91% of citations appeared in only one engine.

Conclusion: AI visibility is not a single leaderboard. Instead, it’s 3 different distribution systems that sometimes overlap and usually do not.

Only 2% of URLs get cited by all 3 engines

Most people would guess that if a URL gets cited by one major AI engine, it has a reasonable shot at appearing in the others.

But the 20,000 prompt sample shows only 2.37% of cited URLs show up across all 3 engines for the same prompt.

Meanwhile, 91.07% show up in only one. Those 2 numbers belong next to each other because they explain each other. The remaining ~7% overlap in pairs, which means engines are drawing from largely disjoint pools rather than ranking the same pool differently.

For AEO / SEO teams, that means a single composite visibility score is the wrong unit of measurement. Averaged AEO scores hide this. A brand can look strong in aggregate and be invisible in 2 of 3 engines. Teams chasing one blended AI visibility number are compressing 3 ranking systems into one metric and calling it strategy.

The 2% holds across every cut

The ~2% overlap rate and ~91% exclusive rate stays almost perfectly flat across 4 samples.

That consistency matters more than the exact decimal point. The consensus gap is not an artifact of one query set or one time window. It looks structural.

In Q3 2025, universal overlap was 2.2%. In Q4 2025 and Q1 2026, it rose to 2.7%. Engine-exclusive citations fell from 90.1% to about 88%. So yes, a small amount of convergence. But even after that shift, fragmentation still dominates.

Commercial prompts don’t converge either

The intent split is one of the quietest but most useful parts of the dataset. You could argue that commercial queries should produce more consensus. When someone searches for the best CRM, best running shoes, or best project management software, the pool of acceptable sources feels narrower than it does for broad informational prompts.

Surprisingly, the data does not support a big difference.

Commercial prompts show 2.4% universal overlap. Informational prompts show 2.0%. Even when the query should narrow the answer set, the engines still choose different sources most of the time.

That pushes against a common instinct in SEO and content strategy. Teams often assume high-intent queries are where shared authority will show up. The opposite looks closer to true. Even in commercial territory, each engine’s own retrieval logic, what sources it trusts, what formats it prefers, is doing most of the work.

Guides beat homepages by 2x

The page type breakdown below shows guides and tutorials have the highest cross-engine overlap at 2.3%, followed by blogs at 1.8%, category pages at 1.6%, product pages at 1.2%, and homepages at 1.1%.

Two lessons:

1. First, explanatory content travels better than brand or transactional assets. If you want the best shot at showing up across engines, the strongest candidate is not the homepage and not the product page. It is the page that helps, explains, compares, or teaches, but keep in mind that these are also content formats that AIs can answer directly well.

2. Second, even the best page types perform badly in absolute terms. Guides are not winning across engines in any meaningful sense. The right read on this is not “publish more guides and you will win everywhere.” It’s simpler than that: Helpful content travels better than brand content.

Visibility is not the same as portability

One of the easiest mistakes in this space is to confuse citation frequency with citation portability. Wikipedia is the cleanest example. It appears 16,073 times in the dataset, but only 1.3% of those appearances are universal across engines. Reddit appears 14,267 times, but only 0.1% are universal. Reuters shows up 1,202 times and still lands at 0.0% universal overlap.

That is why an important metric is portability. A domain can show up all over one engine and barely travel, which means a brand looking dominant in an aggregate dashboard may be one platform’s habit away from invisibility. Presence tells you whether you are visible. Portability tells you whether that visibility is resilient.

What this means for operators

The practical implication is simple: Stop treating AI visibility as one thing. Examine the comprehensive visibility of your domain by measuring:

1/ Presence, the % of your tracked prompts where your domain appears in any engine. Presence tells you whether you’re visible.

2/ Portability, the % of your cited URLs that appear in all 3 engines. Portability tells you whether that visibility is resilient.

3/ Concentration, the % of your citations that come from a single engine. Concentration tells you which engine your current dashboard is secretly built on.

If overlap between engines is this low, a single AEO strategy is too abstract to be useful.

When we approach AI visibility from a holistic perspective, it forces sharper questions:

• Which engine matters most for us?

• Which of our assets travel across engines, and which only work in one?

• Are we measuring presence when we should be measuring portability?

This also changes how brand teams should think about diagnostics. A weak homepage across engines may not be a homepage problem. It is a symptom of something broader: Engines favor utility over brand centrality. In that world, visibility comes less from being the official source and more from being the useful source.

The strategic question is no longer How do we rank in AI? We should instead be asking ourselves How do we build assets that survive different engine preferences? That is a narrower question. It is also a better one.

Methodology

There are a few caveats to this analysis:

• The dataset is skewed toward Omnia’s customer base.

• The intent and page-type cuts rely on regex classification, which is useful for directional analysis but not perfect taxonomy work.

Those caveats do not weaken the main finding much. The biggest signal is not precision at the edges. It is consistency at the center. No matter how the cuts change, the same pattern resurfaces: very little overlap, very high engine-specificity, and only modest differences by time, intent, or page type.

Dataset size and time window

The analysis draws on 4 prompt samples. Three cohorts of 5,000 prompts each, tracked from 1 January 2025, 1 July 2025, and 1 January 2026. A separate 20,000-prompt random sample underpins the headline 2.37% and 91.07% figures. The time-view cut spans Q3 2025 through Q1 2026 (to date) and covers 3.7 million URL citations in total. Commercial / Informational / Other intent splits are drawn from roughly 2.6 million URLs across the combined sample. Page-type splits span 4.1 million URL appearances.

How prompts were selected

The 20,000 prompts are drawn as a random sample from Omnia’s live prompt monitoring pool. The pool reflects what real marketing teams chose to track, weighted toward Omnia’s customer geography (Spain-heavy, plus UK, Nordics, and other EU markets). Each prompt runs in its country’s primary language, so Spanish is overrepresented versus a US-only dataset. Industry mix is fintech/insurtech, travel, SaaS, B2B services. Treat findings as directional for European AI search.

Engine coverage

The study covers 3 engines: ChatGPT, Perplexity, and Google AI Overviews. Each fires the same prompt concurrently within the same minute, twice a day, with country localization and each engine queried in its default web-enabled, unauthenticated state. Perplexity tracking runs on sonar, while ChatGPT and Google AI Overviews use each vendor’s default production model for logged-out web browsing (which neither OpenAI nor Google pins publicly to a specific version).

Classification methodology

Intent and page type are assigned by regex. Intent buckets are Commercial, Informational, and Other. Page-type buckets are Guide / tutorial, Article / blog, Category page, Product page, Homepage, Wikipedia, and Other. The rules are keyword- and URL-pattern-based, which makes them fast enough for a multi-million-URL dataset but coarse at the edges. Edge cases fall into Other, which is why Other carries a high share in both the intent and page-type tables. Treat the regex cuts as directional, not authoritative.

Premium: The YouTube exception and the few universal domain winners

Domain-level data sharpens the page-type finding. Most social and UGC platforms are invisible across engines. As mentioned above, Reddit is at 0.1% universal overlap. But the UGC platform pattern holds: LinkedIn is at 0.1%. TikTok is at 0.1%. Facebook is at 0.0%. Quora is at 0.0%.

Then there is YouTube. And a small set of outliers that illustrate even some “portable” winners are only partially so.

Read more

This Memo was sent to 26,431 subscribers. Welcome to +97 new readers! Subscribe to get the free memo weekly or upgrade to Premium for the full archive, research, frameworks, and templates.

Come join me at AirOps Next in NYC on May 13th. We’re joined by the greats from Airbnb, Ramp, Anthropic, etc.

Subscribe now

How LTM Interactive achieved +400% project ROI with Semrush for Enterprise

Global businesses operate under relentless pressure. But LTM Interactive successfully juggled more clients, growing expectations, and tighter margins by consolidating their search operations into Semrush for Enterprise.

After cutting fragmented tools and turning wasted manual work into high-value client delivery, their teams improved search project ROI +400%.

Read LTM’s full story

Am I still an advisor? Or a builder? I’m having an existential moment.

My work has forever changed in a way I’m still trying to understand. 6 months ago, agentic vibe coding crossed a threshold. Since then, I have used AI to raise my impact by a magnitude.

• I designed landing pages end-to-end for a major travel brand that made it into production.

• I automated topic prioritization, SEO testing, and SEO reporting for my clients with full-blown apps.

• I built an array of useful applications for myself, from automating the SSI (SEO Site Index found in the bimonthly Growth Intelligence Briefs) to Openclaw agents that help me with research and charts.

The work I shipped improved, while it also became harder to define. But when the cost of building collapses due to AI, judgment is the only thing that doesn’t compress. Meanwhile, most operators are still hiring, budgeting, and measuring as if execution is the constraint.

I’m not alone: AI companies are reaching $100M ARR faster than ever, in large part because they’re AI native. Their whole product development philosophy is fundamentally different. Heck, Anthropic went from $9 to $30 billion USD in 6 months and is now worth about as much as Starbucks, Mastercard, or McDonald’s.

And I have my feelings about Matt Schumer’s essay “Something Big Is Happening,” but with reportedly 80 million views, it clearly hit a nerve.

So, I want to take a beat from publishing research this week and take measure of how agentic coding changes software, distribution, and people.

The effect on software

In 2024, I made a bold prediction that AI agents would hit 100 million users in 2025. I was off by about a year. Agents didn’t hit 100 million users in 2025, but they did hit production in 2026, and the gains are measurable:

• METR found 1.5 to 13x (!) time savings when technical staff used Claude Code.

• A 40% reduction in cost and 60% reduction in time from agentic AI is not unrealistic.

• Bain & Co estimates a 30-50% gain in productivity from deploying AI agents and automation.

What happens to software when non-engineers can ship code?

After the iShares software ETF (IGV) cratered 24% in Q1 2026 (steepest quarterly drop since Q4 2008), you could sense a panic in the air that AI would make software companies redundant. But software is more than code.

Enterprise software has strong guards against AI redundancy. Anyone who has ever purchased a CRM or migrated to another vendor knows how hard this is and how much is involved.

Enterprise software is more than code. It’s code plus integration, security, uptime, sales and support… all wrapped up in procurement cycles, IT review, and legal sign-off.

AI can chip away at any one of those pieces. For example, an agent can handle an integration, run a security audit, even book a demo. But no agent shows up to get sued when a mission-critical system goes down at 3 am. Accountability is the part that doesn’t unbundle. Enterprise companies don’t replace this stack; they build their own agents and AI workflows on top of it.

Self-serve software is a different beast. Anyone can now spin up a simple task tracker in a Kanban format. I personally rather pay a few dollars a month and spare myself the hassle of bug fixing, but it’s possible and quick. Self-serve products need to move upmarket. The playbook shows up in Notion’s, Figma’s, and Canva’s moves into enterprise.

In this shift, 2 archetypes stand out:

1. Data providers

2. System of records

1/ Data providers provide value by making data that the market could not otherwise access. These companies lose leverage from their user interfaces but gain it from their data. For example, let’s say a data provider gives you app store rankings. The user interface for that company is slowly turning into friction as more people can code their own dashboards. But their data becomes much more interesting. The durable levers for APIs / MCPs in this world are data completeness, uniqueness, stability, and cost. The logical move is to shift to a headless experience for early adopters and keep the user interface for legacy users.

2/ Systems of record (SOR) are the canonical place where a company’s own data lives. Salesforce, Workday, or Coupa are the bane of existence for many people, but they’re billion-dollar companies because they’re extremely hard to replace. The moat is the tangle of permissions, audit trails, integrations, compliance posture, and decades of workflow conventions built around that data. An agent can generate a CRM in an afternoon; replacing Salesforce at a Fortune 500 is a multi-year change-management project. These companies have already started and will continue to use AI more to provide better user experiences. But their levers are depth of integrations, compliance and audit posture, switching cost, and the quality of their agents. The winners in the SOR space are the ones whose agents make the existing system of record more useful, not the ones trying to replace it.

The effect on distribution

Distribution is more important than product, or so the saying goes, but getting it in 2026 is hard. Platforms are closing (by reducing clickouts and keeping users inside), and they’re taking opportunities away to convert or build direct relationships with visitors outside of the platform.

• AI Overviews and AI Mode make more than 50% of clicks redundant and keep users on the Google platform.

• AI chatbots send a tiny fraction of traffic out.

• Social is flooded, word of mouth is uncontrollable, and paid gets more expensive.

From The Brand Tax:

Cost per visit climbed 9.4% in 2025 alone, adding to a 30% cumulative increase over 3 years. Conversion rates fell 5.1%.

How do you get distribution in this AI-first world? 2 levers compound:

1. Velocity

2. Product

1/ Velocity means you execute faster (and better) than your competitors. When all distribution channels decline and no alternatives open up, the only way to grow is to leverage them better. Play the game better than the competition. Fast shipping speed becomes table stakes, and ideas + compute become the differentiators.

PwC found AI speeds content production up by 3-10x. In plain words, we need to automate more. But not at the cost of trust. When you lose trust, you lose the game.

2/ Product is the marketing now, with 2 distinct effects:

• AI sees through marketing gloss. Agents can read ingredient lists, parse reviews, compare specs. “We’re the best X in the world” doesn’t survive an agent that actually checks. But strong products get chosen consistently.

• Free product is the new top of funnel. Standalone tools that solve a real problem are easier to build than ever, and they acquire better than ads. Ramp Sheets routes users toward Ramp’s core product without a marketing budget.

When product is the marketing, the emphasis shifts to product growth: onboarding, engagement, retention. The fastest growing products these days all have product-led growth motions. So, marketing and product development melt together.

The effect on people

AI capability is racing ahead, but human cognition… isn’t. Until we reach AGI (God knows when; I hope not any time soon), human cognition is what limits AI productivity. We can only ship as much as we can review.

AI tools can take in more input than ever before, while our own human attention span is declining: AI’s context windows grew 3,906x (!) over the last 10 years, from 512 to 2 million tokens, while human attention has shrunk. We’re outsourcing thinking faster than we’re learning to check it.

Two cost curves are racing each other: the Cost to Automate (exponential decay) versus the Cost to Verify (biologically bottlenecked). In “Some Simple Economics of AI,” Catalini et al. argue that tasks with a verifiable output will be automated the fastest. Work that requires a human to check it compounds slower, so we’ll automate work that’s easy to measure faster. I feel it whenever I’m running four terminal windows at once: the focus drain is as high as the throughput. At scale, what holds us back is how much we can proofread and direct.

When anyone can build anything, the ways we’re limited change: Skill and tools matter less. But judgment, ideas, and time decide whether you run in the right direction or in circles. It’s very easy to get distracted with AI because the cost to build is now so low.

Judgment is the part that doesn’t compress. I can ask Claude Cowork for a contract review, but I have to know what it missed. Claude will happily write me a Q4 plan, but it’s only as good as my read on which market to attack and what my competitors are about to do.

Over the last 6 months, I implemented more agentic and automated systems than I’ve done hands-on work. My clients now have access to unique software they can’t get anywhere else that solves their unique problems.

3 things are now compressing toward zero: the cost to build software, the cost to produce content, the cost to spin up a tool. But another cost is trending far away from zero: The cost to know whether any of it is right.

I’m not directly “doing” the work anymore in a traditional sense. I’m now building the thing that does the work, then checking it. The work that matters now is the part I can’t hand to an agent… knowing what to build, what to kill, and what the agent missed. And I’m here to figure out what that means - with you.

Premium: My 4 top rules for AI automation

At Shopify, we built the initial concept of a “keyword universe” when I was there between 2020 and 2022. I wrote about it publicly for the first time in February 2024. In June 2025, I republished the post with a Claude Artifact that helps you build your own keyword universe. In March of this year, I built a full web app in Next.js for my client. I could probably charge a decent subscription fee if I were to publish it.

That keyword universe arc shaped how I work with agents today.

The same patterns power an SEO testing system I built for another client: 4 rules carry across both.

Read more

This Memo was sent to 26,313 subscribers. Welcome to +85 new rea…

Read more

This Memo was sent to 26,187 subscribers. Welcome to +95 new readers! Subscribe to get the free memo weekly or upgrade to Premium for the full archive, research, frameworks, and templates.

When an AI …

Read more