Why most original data never gets cited
Primary research is rare, but when it appears, it is much more citation-dense than ordinary pages.
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.”
In this memo:
First-party research is rare in AI citations, but it earns 3.3x more
75 of 90 primary-research citations came from a single content format
Owning data is not the asset. A benchmark is.
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.
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.
