AI engines quote numbers. When ChatGPT, Perplexity, or Google AI Mode assembles an answer, sentences that carry a specific figure get lifted far more often than sentences that carry an opinion. In our June citation audit, 61 percent of the 214 AI citations we tracked across client sites pointed at a sentence containing a number.
Most businesses never pull that lever. They publish claims. The sites that win citations publish measurements — original, dated, clearly denominated statistics that a model can lift verbatim and attribute to a named source.
Here is how we write statistics at Avakata so engines quote them, and how to mint your own without a research department.
Why engines reach for numbers first
Generative engines prefer statistics because a number is a pre-compressed fact. It is specific, checkable, and easy to attribute, which lowers the model's risk of being wrong. A vague sentence forces the engine to paraphrase. A statistic lets it quote. The Princeton research that named GEO found that adding statistics to a passage raised its visibility in generative answers by up to 40 percent, and our own audit data keeps agreeing with that finding.
There is a second reason. Answer engines are competing on trustworthiness, and numbers read as evidence. A cited figure makes the engine's answer look researched. Your statistic improves their product, so they use it.
Numbers also survive paraphrase. An engine can rewrite your argument into its own words and drop you entirely, but it cannot rewrite 61 percent into something else. The figure forces attribution.
The demand side is settled. The question is whether your pages give the engines anything worth taking.
What makes a statistic quotable
A quotable statistic has four properties. It is original, meaning you measured it and nobody else has it. It is specific, with a real number rather than a rounded gesture. It is denominated, so the reader knows what was counted and out of how many. And it is dated, so an engine can judge freshness. Miss any one of the four and your citation odds drop sharply.
Original is the property most sites skip. Republishing a Gartner number adds nothing, because the engine already has Gartner. Your leverage is the data only you can produce, at whatever scale you actually operate.
Compare two sentences. "Most product pages are too thin to get cited" is a claim. "Of the 68 Shopify stores we audited in 2026, 43 had product pages under 120 words, and none of those pages earned a single AI citation" is a source.
The second sentence took one spreadsheet and ten minutes longer to write. It is the one that gets quoted.
Write the number into one liftable sentence
Engines lift sentences, not paragraphs, so every flagship statistic needs one self-contained sentence that carries the full fact: subject, number, denominator, timeframe, and source. Keep it under 30 words. If a reader saw that sentence alone on a blank page, they should understand exactly what was measured and who measured it. Never split the number from its context across two sentences.
We draft the sentence before we draft the section around it. The pattern is boring on purpose: across this many things we measured, over this period, this happened.
Then we resist the urge to decorate it. Adjectives lower extractability. The sentence is a specimen, not prose.
Test it the way an engine would. Paste the sentence alone into a chat and ask the model what it means. If it has to guess at the subject or the timeframe, rewrite until it does not.
Show the denominator and the method
A statistic without a stated sample size reads as marketing, and engines increasingly skip it. State what you counted, how many, and over what period, in one plain line near the figure. "Based on 412 posts published across nine client sites between January and June 2026" costs you fifteen words and buys the credibility that separates a quotable measurement from a slogan.
Small denominators are fine. Honest beats big. A finding from 14 client projects is still more citable than an unsourced industry cliche, because nobody else has it.
What kills trust is hiding the method. If the engine cannot tell where the number came from, it will find one where it can — usually on a competitor's page.
One line is the right length. A full methodology section is for research papers. State the count, the period, and the source, then move on.
Date every figure you publish
Every statistic should carry a visible date, either in the sentence itself or in an as-of stamp beside it. Engines weight recency heavily when choosing between competing sources, and an undated number gets treated as an old number. In our audit set, pages that dated their figures were cited roughly twice as often as pages carrying the same figures undated.
Dating also protects you. When a figure ages out, you re-measure, update the number, and refresh the stamp. That update is itself a freshness signal.
We re-run our core measurements quarterly. Ninety minutes of re-counting keeps a page citable for another season.
An honest stamp beats a gamed one. Engines cross-check content changes against the date, and a bumped date on unchanged text erodes exactly the trust you are trying to build.
Place statistics where models actually look
Position matters. Put your flagship statistic in the first 100 words of the page, restate it once under the most relevant question-shaped heading, and include it in your key takeaways block if you run one. Passages near the top and passages directly under headings get retrieved most often, so a number buried in paragraph fourteen is a number the engine never sees.
One restatement is enough. Repeating a figure five times does not multiply citations, and it reads as padding to human visitors, who still matter.
Key takeaways earn their keep here. A four-sentence takeaway block at the top of a post is effectively a menu of liftable claims, and we watch engines order from it constantly.
If a page has no natural home for the number, that is usually a sign the page answers the wrong question. Move the statistic to the page where buyers actually ask about it.
Mint a flagship statistic every quarter
The sustainable version of this is a quarterly measurement habit. Pick one question your buyers actually ask, count something you uniquely can count — your own projects, audits, tickets, or benchmarks — and publish the result as a one-sentence finding with the method and date attached. One original statistic per quarter is enough to make your site a source rather than a commentary.
Ours come straight from operations. Citation audits, publishing logs, edit rates. The data already exists. Publishing it is the only new work.
The bar is lower than you think. Fourteen projects, sixty audits, a hundred support tickets — any of those counts is publishable if you state the method honestly.
Being the source compounds. Once an engine quotes your number, that answer surfaces your brand to people who never searched for you, and other writers start repeating the figure with your name attached.
Count something this quarter.