TL;DR
📊 The GEO study improved on baseline by up to 41% using statistics, quotations and cited sources, measured as position-adjusted visibility inside a generated answer, which is why a number beats an adjective.
🧪 C-SEO Bench found only 3 of 54 unilateral conditions produced statistically significant citation-rank gains, so rewriting a page you already have is not the strategy.
🔗 57% of AI citations point to sources brands do not control, and original data is the one asset that earns citations on both sides of that split.
🔎 87% of SearchGPT citations matched Bing's top results: a study nobody can retrieve is a study nobody cites.
📈 Telemetry counts as research when the denominator is public, as with an index built on 126 million AI search prompts.
The most durable way to be cited by an AI engine is to hold a number nobody else holds. Phrasing tactics test badly: C-SEO Bench tested ten conversational-SEO rewrite methods and found only 3 of 54 unilateral conditions produced statistically significant citation-rank gains. But an answer that uses a figure has to attribute it to whoever produced it, and if you produced it, the attribution is yours by construction. That is the whole strategy behind becoming a source AI engines cite: publish something only you can publish, then publish it so a machine can read it and credit it.
Here is the short version. Run a repeatable measurement of your own market, state its method and its denominator in plain language, date it, and put the result at a stable URL in plain HTML. The GEO study improved on baseline by up to 41% using statistics, quotations and cited sources, measured as position-adjusted visibility inside a generated answer, and the published version carries the full method. Profound found 57% of AI citations point to sources brands do not control, and original data works on both sides of that split: it earns citations on your own pages, and it gives the sources you do not control a reason to name you.
What makes a page quotable to an AI engine?
Three properties, and none of them is tone. The claim has to be quotable, meaning it survives extraction as a standalone sentence. It has to be attributable, meaning the engine can say who produced it without guessing.
And it has to be specific, meaning it carries a number and the population that number came from, with a date attached. A sentence saying your category is growing fast fails all three at once.
Retrieval gates all of it. Seer found 87% of SearchGPT citations matched Bing's top results, and an AirOps analysis of 548,534 pages mapped which page traits correlate with being pulled into an answer. Vercel's crawler research with MERJ found the ChatGPT and Claude crawlers it tested fetch pages without executing JavaScript. A study nobody can fetch is a study nobody cites.
The corrective belongs here, because this is tactics-adjacent territory. The C-SEO Bench result above rests on released code and data covering two tasks across six domains, and the NeurIPS 2025 paper carries the method-by-method breakdown. Rewriting an existing page to sound more citable mostly does not work. Holding a fact the answer cannot be written without is a different mechanism entirely.
What counts as original data if you do not have a research team?
Four kinds, all available to a company with no research budget. Scale is not the test. The test is whether you are the only party able to produce the number.
The first is aggregate product telemetry: whatever your system already counts, counted across a stated population and published with the denominator visible. Semrush's AI Visibility Index, built on 126 million AI search prompts, and Profound's 7.5 million-conversation sample are that pattern at large scale, and Cloudflare's report that 52% of crawler requests now come from AI-related bots is the same move made from infrastructure logs. A smaller company publishing a smaller count, honestly bounded, is doing the same thing.
The second is a recurring survey of your own market, run on the same instrument every year so the trend line becomes the asset. 6sense's 2025 B2B Buyer Experience Report and G2's finding that AI chatbots are now the single largest influence on B2B shortlists are quotable for one reason: someone asked a stable question of a defined population and published the denominator. Two hundred honest responses beat two thousand from a population nobody described.
The third is a benchmark you run and publish, including the code that produced it. GEO-bench and C-SEO Bench both ship the code that produced their results, so anyone can rerun them and check the number before quoting it. A benchmark is expensive once and cheap every year after.
The fourth is the teardown nobody else bothered to do: a systematic read of a corpus, a comparison run across many configurations, a count of what a public dataset actually contains. Entity-oriented retrieval research spanning 443 configurations has that shape. So does reading 200 competitor pricing pages and reporting what share publish a price at all.
How do you publish a number so it can be extracted and attributed?
Five mechanics, ordered by how often they get skipped. State the method on the same page as the result, not inside a gated PDF.
Give the denominator. Put a date on it. Keep the URL stable forever. Serve it in plain HTML.
Placement matters more than most publishing checklists admit. Liu et al. showed language models use information least reliably when it sits in the middle of a long input, so the headline figure belongs near the top of the page and again in the section that explains how you got it. Keep each finding to a self-contained paragraph: passages in one multi-hop chemistry retrieval corpus ran a median of 188 tokens with a 56.26-token standard deviation, which is a rough size reference rather than a rule.
Attribution is a provenance problem, and provenance has a vocabulary. The W3C provenance ontology models an Entity, the Activity that produced it, and the Agent responsible for that activity, which is a usable checklist even if you never write the markup: name the thing measured, the process that measured it, and the party accountable for it. Fetchability is not uniform either, since OpenAI documents four relevant user agents with different purposes, so a page reachable by one is not automatically reachable by all.
What makes a published number survive scrutiny?
A number without a stated method and a denominator is worth nothing, and publishing one costs you more than publishing nothing at all. The first reader who asks how many are in the sample and gets no answer stops treating your domain as a source.
The specific failure to avoid is dressing an anecdote as a study. Three customer calls are three customer calls, and saying exactly that is fine. The FTC's advertising guidance is the floor rather than the ceiling here: claims about performance need substantiation you can produce on request. A number you cannot reconstruct is a number you should not have published.
Report the null result too. The C-SEO Bench headline finding is a negative, and it is the finding the benchmark is known for. Publishing what did not work is the cheapest credibility available, and the part most companies cut first.
Then size your claim to your variance. A 2026 variance-components study found run-to-run noise in AI answer measurement large enough to swamp real differences in small samples, and work on quantifying uncertainty in AI visibility argues the same case with confidence intervals. If your measurement rests on ten runs, write ten runs.
How do you tell whether the data actually earned citations?
By measuring the right rung, repeatedly. A mention is your name appearing in an answer. A citation is your page used as a source.
A recommendation is the engine advising the buyer to choose you. Original data moves the citation rung most directly, and its effect on the rungs above is slower and indirect.
One reading proves nothing. SparkToro's research found AI engines are highly inconsistent when recommending brands, and Search Engine Land's write-up of the same work reports that recommendation lists rarely repeat exactly. Run the same question ten or more times across several days and track the rate, never the instance.
Report the rungs separately inside your own team. The IAB splits AI visibility into Presence, Prominence, Portrayal and Persuasion instead of one composite number, which is the right instinct. And keep the purpose in view: 6sense found buyers evaluate roughly five vendors and most of that list is set before contact.
The other half of this work happens off your domain, on the review sites and community threads you do not own. That is a different playbook with different mechanics, and it is covered separately.
How does Trovance help you become the source that gets cited?
Trovance starts from the questions rather than the page. You define the market questions your buyers actually ask, and it runs them repeatedly across AI engines, preserving each answer run with the sources that carried it. That record tells you which numbers the answers in your category depend on, and whose numbers they are.
From there the work is evidence rather than volume. Your Brand Core holds the claims you are entitled to make and the proof behind each one, so a claim with no method or denominator behind it does not quietly become a draft. Recommended actions name the specific asset the record says is missing: the benchmark that would answer a competitor's quoted study, the recurring count no source in your category currently publishes. Drafts are produced from approved claims, and a person reviews everything before it publishes.
Then the cycle closes. After your asset goes live, the next analysis run puts the same questions back through the engines and compares the new answer snapshots against the preserved ones, so you can see whether the number you published started appearing, in which answers, and next to whose name.
What Trovance will not promise is that publishing original data gets you cited. No honest system can, because retrieval is probabilistic and the engines keep changing. There is no single visibility score here and no control over what a model says. What you get instead is an evidence record candid enough to tell you when a bet did not pay off, which is the part that makes the next bet better.
What should you do this week?
Pick one number you are already able to produce: what your product counts, or what a public dataset would reveal if somebody read it end to end. The constraint is rarely analytical skill. It is choosing a question your market cares about that only you can answer.
Then publish it properly: method on the page, denominator visible, date stamped, a URL you promise never to move, plain HTML a crawler can read without running JavaScript. Write the finding as a self-contained paragraph near the top. Then schedule the next run, because the second reading is what turns a data point into a trend line worth quoting.
Be honest about timing. Retrieval can happen quickly once a page is fetchable, but becoming the source a category quotes takes repeated publication across quarters, and the variance in AI answers means your early readings will be noisy. If you want the measurement side handled while you produce the data, start a free Trovance analysis and see which questions in your market are still being answered with somebody else's numbers.
What engines actually cite
What gets cited by AI - the page traits that correlate with being quoted.
How to earn third-party AI citations - the inverse problem, on sources you do not control.
AI agents need evidence, not more content - why publishing volume stops working.
A relevant page can still miss the proof - relevance and evidence are separate tests.
Publish it so it holds up
The most-quoted study in AI search, quoted correctly - what the GEO benchmark actually tested.
AI crawlers don't run your JavaScript - the fetchability floor under every study you publish.
How to track AI citations - measuring whether the number you published shows up.
There is no such thing as an AI visibility score - why one composite number hides the rungs.
FAQs
How do I make my brand a source AI engines cite?
Publish a measurement only you can produce, then make it extractable. State the method and denominator on the page, add a date, keep the URL stable, and serve plain HTML. In the GEO benchmark, adding statistics and cited sources moved visibility while phrasing rewrites mostly did not, so the number does the work rather than the wording around it.
What counts as original data for a small company?
Four sources work at any size: aggregate product telemetry counted across a stated population, a recurring survey of your market run on one instrument, a benchmark you run and publish with its code, or a systematic teardown nobody else attempted. Being the only party able to produce the number is the real test.
Does rewriting existing pages help me get cited?
Rarely on its own. C-SEO Bench tested ten conversational-SEO rewrite methods and found only 3 of 54 unilateral conditions produced statistically significant citation-rank gains. Rewriting helps when it makes a real fact easier to extract. It does not manufacture a reason for an engine to quote your domain at all.
How much data do I need before publishing a study?
Enough to state a denominator you are not embarrassed by. A few hundred responses from a defined population is usable; three customer calls is an anecdote and should be labeled one. The disqualifying failure is a sample whose size and method never appear anywhere on the page.
Where should the number appear on the page?
Near the top, in a self-contained paragraph, then again in the section explaining the method. Language models use information least reliably when it sits mid-document, so a headline figure buried on page four of a gated PDF is functionally missing. Keep each finding quotable without its surrounding context.
Do I still need third-party coverage if I publish original data?
Yes. Profound found 57% of AI citations point to sources brands do not control, so review sites and community threads still decide much of the answer. Original data helps there too, because it gives those sources a specific figure to attribute to you instead of a positioning claim.
How long until publishing data changes what AI engines cite?
Retrieval can happen quickly once a page is fetchable, but becoming a source AI engines cite for a whole category takes repeated publication across quarters. Measured answers also vary run to run, so early readings are noisy. Nobody can honestly guarantee a citation on a timeline; track appearance rates instead.
