SIGNALS
How AI search works

Do AI engines check whether the statistics on your page are real?

The short version

At retrieval time, mostly not, and that is the problem rather than the opportunity. Qi Liu and colleagues published a fabricated concept and 8 of 10 AI search platforms cited it within seven days, and one platform cited content from a $14 optimisation package within an hour (arXiv 2610.11932, 2026). So the engines are not currently a filter on whether your numbers are true. Two other findings point the other way. When Runze Kong and colleagues pointed a verifier at pages rewritten to add statistics, it flagged all 80 descriptions produced by that strategy for numerical faithfulness (arXiv 2611.01126, 2026). And Zeyu Zhang and Bradly Stadie found that one fabricated article dropped into a model's context reversed 74% to 82% of forecasts that disagreed with it (arXiv 2611.15102, 2026). Verification is being built, it already works on unsupported numbers, and the common advice to add statistics is half an instruction. The half that does the work is the source beside the number.

Do AI engines check whether the statistics on your page are real?

AI engines do not reliably check whether the statistics on your page are real before quoting them, on the evidence published so far. Nothing in the retrieval path of a live answer engine is documented as verifying a figure against its claimed source, and the experiments that have tested the question by planting false numbers found them propagating rather than being filtered out.

That is an uncomfortable answer for everybody in this category, including us, and it has to be said before the practical advice, because the advice only makes sense once you accept the mechanism. The reason to source your numbers is not that an engine is auditing you today. It is that unverifiable material is a fragile asset: it is being checked more as verification layers get built, it costs you the reader who looks, and it is the one thing in this whole discipline that can turn a visibility programme into a liability.

There is also a narrower finding worth holding onto, because it cuts against the fatalism. Where researchers have put a verifier into the loop, it caught unsupported numbers at a high rate. So the figure that cannot be traced is not invisible to machines in general, only to the live retrieval step as it currently works.

The three sections below are the published experiments, each with its scope, and then what the evidence actually supports doing on a page. None of them is a reason to stop putting numbers in your writing. All of them are reasons to put the source next to the number.

What happened when researchers published a fabricated concept?

When researchers published a fabricated concept on ordinary public platforms, 8 of 10 AI search platforms cited it within seven days. The experiment sits inside a larger citation-mapping study covering 17,211 citation instances across 6,356 unique source domains on 10 AI search platforms, reported by Qi Liu and colleagues in "From Public Posts to AI-Search Citations: Measuring the Fragility of AI Search" (arXiv 2610.11932, submitted 8 October 2026).

The study's other numbers describe why that was easy. Of 22 publication platforms tied to cited source domains, 15 had low or medium barriers to both account creation and posting. The authors also bought a $14 generative engine optimisation package, which produced 13 public posts, and one AI search platform cited that content with their designed markers still in it within one hour. One article on a platform the engines prefer carried more citation impact than more than 20 matched posts elsewhere.

Two conclusions follow and they point in opposite directions. The first is that being cited is not evidence of being correct. Keep that in mind before reading anybody's citation count as a quality signal, ours included. The second is that the cheapness of the attack is itself the argument for the discipline: in a retrieval environment where a fabricated concept travels in a week, a firm whose numbers survive checking is accumulating something that a seeded network cannot buy.

What nobody should take from that paper is a tactic. Its methods are now explicitly prohibited by at least one platform: Anthropic's usage policy, effective 12 November 2026, bars users from seeking to "Manipulate the sources from which search engines or AI systems draw answers", including "seeding them with content that misrepresents its origin, authorship, or independence" and "networks of sites posing as unaffiliated sources corroborating the same claims", in its published usage policy (read 17 November 2026).

What did a verifier find in pages rewritten to add statistics?

A verifier pointed at pages rewritten to add statistics flagged every single one. In a benchmark that published optimised content as live web pages and let real engines retrieve and choose between them, the judgement verifier flagged 86 of 400 rewritten product descriptions for numerical faithfulness, and those 86 included all 80 of the descriptions produced by the statistics-adding strategy, according to Runze Kong and colleagues in "Do GEO Gains Survive the Live Web? Evaluating In-the-Wild Visibility Gains of Generative Engine Optimization via LIVE-GEO" (arXiv 2611.01126, submitted 31 October 2026).

Read that carefully, because it is the sharpest thing in this whole literature about the advice everybody repeats. The study found that a rewriting strategy whose instruction was to add statistics produced unsupported numbers in 100% of its output. Not as an accident of one bad run: as what the instruction does when the writer has no sources to hand. That is the mechanism by which well-meant optimisation advice turns into fabrication, and it is why the useful version of the instruction has to be "add a statistic you can attribute" rather than "add a statistic".

The study's scope limits what else it can be used for. It ran on 80 fictional brands across 18 product families in a shopping task, with prices frozen so price could not do the work, judged by three live engines. It is one preprint, not replicated, and it measures product recommendation rather than citation of an article.

Two other findings in the same study are worth knowing for a different reason. The live engines picked the same top product for only 48.8% to 66.3% of brands, and the leading strategy's top-one rate on ChatGPT moved from 73.8% under one fixed candidate order to 93.8% under the reverse. In other words a large part of what looks like a content effect is an artefact of the test setup. That is a good reason to be sceptical of anybody's single-number claim about what a rewrite achieves, including the ones in this field's most-quoted studies.

Does one fabricated source change what a model concludes?

One fabricated source changes what a model concludes far more than its single-document weight suggests. Zeyu Zhang and Bradly Stadie planted one fake news article into the retrieval pool of forecasting systems and found that, once it was in context, it reversed 74% to 82% of forecasts that had opposed it, across every frontier model tested, in "One Fabrication Is Enough: A Fake News Article Changes the Forecasts of LLM Forecasting Systems" (arXiv 2611.15102, submitted 11 November 2026).

The retrieval half of that result is the part that matters for a page. The authors found that the fabricated article was retrieved in about 90% of runs by the architectures of two leading deployed forecasting systems, so the document did not need to be authoritative to be read, only to be retrievable and on topic. That is the same property a well-structured page on a buying question has, and the symmetry sits at the centre of this subject.

The study's defence tests are where the finding becomes an argument for restraint rather than a reason for despair. Passage filtering and partitioned retrieval each stopped the single-article attack, each was bypassed by a one-prompt change, and the two together changed 29% of clean decisions and discarded 39% of genuine news. A filter aggressive enough to catch fabrication throws away a third of the real material with it.

Scope matters here too. The task is forecasting rather than citation, the deployed systems were re-implementations against a frozen archive, and this is one preprint a few days old at the time of writing. What it supports is narrow and useful: in a retrieval system, the cost of a single unverifiable document is not proportional to its size, and the cost of filtering it out is paid by everybody who wrote in good faith.

So should you still put statistics on your pages?

Yes, put statistics on your pages, with the source in the same paragraph as the number, and stop treating the statistic itself as the lever. The evidence that formatting choices move citations is much weaker than the advice implies, and the evidence for what does move it points somewhere else entirely.

The largest published analysis of this gives alignment between the page and the question asked a standardised coefficient of +0.37 with a 95% confidence interval of +0.33 to +0.41, significant after false discovery rate correction at q below 1e-73, across 2 million AI citations over six months and 10,000 crawled cited URLs, according to Discovered Labs in "What actually drives AI citations" (read 17 November 2026). Its format coefficients are far smaller: page length at +0.13, title-prompt similarity at +0.09, an FAQ section at +0.07 on third-party pages and a summary block at +0.05. Two caveats belong with those figures. The study requires a predictor to survive at least four of nine robustness checks rather than all of them, and Discovered Labs sells AEO services, so the choice of what to measure is not disinterested.

Put beside the verifier result above, that gives a clean practical rule. The number is not what earns the citation; the question-shaped page earns it, and the sourced number is what keeps the page defensible once a reader or a verification layer arrives. Our own methodology weights it that way, with alignment as the dominant dimension and grounding scored on whether each claim carries a named source.

The advice that fails is the version that treats statistic density as a target. A page with fourteen numbers and no sources scores zero on grounding in our model and reads as invented to a human, and as the LIVE-GEO result shows, it is exactly what an instruction to add statistics produces when nobody has the sources. Our guide on whether rewriting a page for AI search can make it worse covers the other ways a well-intentioned rewrite backfires.

What does a properly sourced statistic look like on a page?

A properly sourced statistic names who measured it, when, and on what, in the same paragraph as the figure, and links to the place a reader can check it. The table below is the standard we hold our own pages to, with the version that fails beside it.

Element What passes What fails
Attribution Named author or organisation, in the sentence "Studies show", "industry reports suggest"
Date Publication date of the source, and the date you read it No date, or only the year the page was written
Sample What was measured and how much of it A percentage with no denominator
Location In the same paragraph as the number In a bibliography at the foot of the page
Link To the primary source, opened and read To a secondary post repeating the figure
Conflict Stated when the publisher sells in the category Omitted, leaving the reader to find it
Scope The limits the authors state, carried over The headline number without its conditions

Two of those rows are the ones most often skipped and they do the most work. A figure placed in a bibliography is unreadable to a reader who arrived mid-page from a citation, which is how a growing share of readers arrive. And a figure quoted from a secondary post is how rounding errors and misattributions spread: the number circulating in blogs is frequently a variant of the published one.

The last rule is the one we treat as non-negotiable rather than as a style preference. Where a page is too thin to support a grounded claim, the correct output is to say so and leave the field empty, not to supply a plausible figure. A tool that scores other people's pages on grounding cannot be loose with its own, and if you want to see what that reading produces on a real page, our teardown of one of our own pages shows the score of 28 out of 100 it gave us and the 71 we reached after the fixes. A free visibility assessment runs the same reading on yours.

What else do people ask about statistics and AI citations?

Marketers writing for AI search ask these five questions most often about statistics and sourcing.

Will an AI engine penalise a page for an unsourced statistic?

No published evidence shows a penalty applied by the engines themselves. What is documented is that verification layers do exist in research settings and that they flag unsupported numbers reliably, and that an unsourced figure gives a reader and a model nothing to check. Treat the risk as selection against you rather than a sanction, and treat the trend as moving towards more checking rather than less.

Does putting a source next to a number make a page more likely to be cited?

The strongest page-level finding in the current evidence is not about statistics at all, it is that the match between a page's language and the question asked carries a far larger measured effect than any formatting choice. A named source makes a claim quotable and checkable. That is reason enough on its own, but it should not be sold as a citation lever with a number attached to it.

Is it safe to cite a statistic you found in another article?

Only after you have opened the original. Figures drift as they are repeated, and the version circulating in secondary posts is often a rounded or misattributed variant of the published number. The cost of opening the source is a few minutes; the cost of publishing a number that does not survive a reader checking it is the credibility of everything else on the page.

What should you do with a claim you cannot source?

Cut it, or write the sentence without the number. A page that says a practice is common is weaker than one that says 73.5% of citation URLs appeared exactly once and never returned, according to Trakkr Research's 2026 study of 108,650 of them, and both are better than a confident figure with nothing behind it. Where a page is too thin to support any grounded claim, the correct output is to say so rather than to fill the space.

Is seeding content to influence AI answers against the rules?

Against at least one platform's rules explicitly. Anthropic's usage policy, effective 12 November 2026, prohibits manipulating the sources from which search engines or AI systems draw answers, seeding them with content that misrepresents its origin, authorship or independence, and networks of sites posing as unaffiliated sources corroborating the same claims.

Related guides

The Assessment

See how your pages read to something that checks.

A free visibility assessment samples your buying questions across ChatGPT, Perplexity, Google AI Overviews and Claude, and scores your pages on whether every claim carries a source a reader can follow.

Request a free visibility assessment →
SIGNALS · A BlackSig Systems company