Blog · Content that gets cited

Original data gives an engine something only you have

Published 27 August 2026 · Updated 8 October 2026 · 7 min read

Original data gives an engine something only you have
Photograph: Tsimur Asayonak / unsplash

The short answer: Original data gives an engine something only you have because it is the only kind of content an engine cannot substitute. Twenty pages explain the same concept equally well, so the engine picks whichever one it already trusts, and that is rarely you. One number that exists nowhere else leaves it a single candidate. A week of counting your own records buys what a quarter of link outreach often does not.

That is a claim about cost per citation, not about effort, and it only holds if you count what the alternatives cost.

Why substitutability decides who gets cited

An answer engine assembles a reply by fetching candidate documents, cutting them into passages, keeping a handful and writing from what it kept. At the keeping step it is making a choice between passages that often say the same thing. When ten passages carry the same general claim, the tiebreak falls back on signals you cannot buy quickly: domain history, how often that site has been a safe source before, whether the brand resolves cleanly as an entity. A new site loses that tiebreak nearly every time.

Now change one passage so it carries a fact the other nine do not have. There is no tiebreak left to run. If the answer is going to contain that fact, the citation has to point at the page that holds it. Scarcity, not quality, is what moves you into the answer. That is the whole mechanism, and it is why what makes a page citable by an AI engine keeps coming back to specificity rather than polish.

Google's own guidance now draws the same line in plainer words. Its May 2026 guide on optimising for generative AI features in Search splits content into commodity and non-commodity, and describes commodity content as material based on common knowledge that could have come from anyone and adds little unique insight. The same page says there is no special schema.org markup that makes a page eligible for these features. Read those two statements together and the conclusion is uncomfortable for a lot of content plans: the lever is what the page knows, not how it is annotated.

What the research actually measured

The clearest public experiment on this is the GEO work presented at KDD 2024, which built a benchmark of 10,000 queries and then rewrote source passages nine different ways to see which rewrites made a generative engine more likely to use them. Adding statistics, and adding citations to sources, each moved visibility by roughly 30 percent in relative terms against the unmodified passage, with the paper's headline ceiling at 40 percent across all nine rewrites. Keyword stuffing moved almost nothing. Writing in a more authoritative tone moved almost nothing either.

Two cautions before anyone quotes that at a budget meeting. The study measured rewrites of existing passages, not the business of generating a number in the first place, and the gain varied a lot by subject area. Treat it as evidence that engines are drawn to concrete quantified claims, not as a promised multiplier for your next report.

What cheap means here, in practice

Most teams hear original data and picture a commissioned survey with a panel vendor and a four figure invoice. That is the expensive version, and usually the worst one, because a bought panel produces a number any competitor can also buy.

The cheap version is a count of something you already hold. A payments company knows its own decline rates by issuer. A recruiter knows median days to offer by role. A clinic knows its own no-show rate by appointment slot. None of that needs new collection. It needs a query, a denominator, a date range and somebody willing to publish a number that is not flattering. Three hundred rows from your own system is a dataset. It is also something no rival can replicate, because it is about you.

A reasonable first pass is one afternoon of SQL, one page, one method paragraph. If the page earns nothing in a quarter, you have lost an afternoon. The comparison worth making is against what a single placement in a trade publication costs, or what three months of outreach costs, and against the fact that neither of those produces an asset you own.

The four ways original data fails to earn the citation

Having the number is not sufficient. Most data pages that go nowhere fail at one of these.

The number is trapped in an image. A chart rendered as a PNG with the figure only in the bars is invisible to a passage extractor. Every number you want quoted has to exist as a sentence in the HTML, in prose, with the units attached.

The number has no denominator or date. "Decline rates rose 12 percent" is unusable. An engine that quotes it inherits a claim it cannot qualify, which is exactly the kind of claim retrieval systems are tuned to avoid. Write the population, the window and the method in the sentence next to the figure, not in a footnote.

The number is not bound to your name. This one costs more citations than the other three together. If the quotable sentence reads "the median time to offer was 34 days", any site can restate it, and the restatement will often be the version that gets cited, because the restating site has more domain history than you do. Bind the entity to the figure in the same sentence: "In our 2026 analysis of 4,100 hires, median time to offer was 34 days." Now the sentence cannot be lifted without carrying your name.

The page is not reachable. A PDF behind a form, a page blocked in robots rules, a figure rendered client side after the crawler has stopped waiting. Access failures are boring and they are the most common single cause of a good asset earning nothing. The self-check covers the ten you can test yourself in an afternoon.

A check you can run this week

Pick the single best number you already have. Search its exact phrasing in quotes on a normal web search, and see which sites come back holding it. If other sites have your figure and your name is not in the sentence they used, you have a binding problem, and rewriting one sentence will fix more than a month of promotion.

Then ask three engines the question your data answers, with retrieval on, several times each. Record whether your number appears, whether it is attributed to you, and whether it is attributed correctly. A single run tells you nothing, which is why the number of runs matters more than the number of prompts.

The same discipline applies to the other formats that get quoted. A comparison page that gets cited works for the same reason a data page does, because it carries a judgement nobody else has written down. And if an agent reaches your page and then cannot act on it, the citation leads nowhere, which is the subject of whether an agent can book, buy or fill your form.

Honest limits

A citation is not a visit. Pew Research Center tracked the browsing of 900 US adults across roughly 69,000 Google searches in March 2025 and found that users clicked a link inside an AI summary in about 1 percent of visits to pages carrying one, and clicked a standard result in 8 percent of those visits against 15 percent when no summary appeared. If your business case rests on sessions, original data will disappoint you. If it rests on being the named source in the answer a buyer reads, it is the fastest route there.

We also cannot tell you how long a data asset holds its position. We re-run our own prompt set monthly and the sources move. And a number that is interesting once becomes commodity when three competitors publish the same cut, so the asset has a half life nobody has measured properly yet.

If you would rather see how this gets diagnosed before you commit, the sample audit report shows the shape of the finding. Anybody selling you a subscription instead of a finding is selling something else, which is the point of buying an AEO audit without being sold a dashboard.

FAQ

How much data is enough to count as original? Enough that you are not ashamed of the denominator, and enough that the method holds up when somebody pokes at it. Three hundred records out of your own system, cut one honest way, beats a thousand panel responses two competitors also paid for.

Does the data have to be about our product? No, and it is often better if it is not. Data about your customers' behaviour, your category's pricing or your operational reality is quotable in answers where a product claim would be ignored.

Will structured data make the number more likely to be cited? Google states there is no special markup requirement for its AI features. Mark the page up properly because it helps a machine resolve who you are, not because it promotes the figure.

What if a bigger site republishes our number without crediting us? That is the binding failure described above and it is partly preventable. Put your name and the sample size inside the sentence that holds the figure, and keep a dated canonical page the restatement can point back to.

Sources

Related reading

What makes a page citable by an AI engineComparison pages that get citedCan an agent book, buy or fill your formHow to buy an AEO audit without being sold a dashboardPricing pages and what AI says your product costsStatistics pages, and how to build one engines trust