Blog · Content that gets cited

Statistics pages, and how to build one engines trust

Published 23 September 2026 · Updated 12 October 2026 · 8 min read

Statistics pages, and how to build one engines trust
Photograph: Max Harlynking / unsplash

The short answer: Build the page so a single number can survive being cut out of it. Each statistic gets one sentence, a figure, the date it was measured, the population it was measured over, and a link to the place it came from. Put that structure in the markup as well as the prose. A page of numbers with no provenance gets read, used, and credited to somebody else.

Most statistics pages are built as link bait, then their owners wonder why assistants quote the numbers without naming the site. The two jobs are different. A citable page needs a retrieval system to lift one sentence, keep the attribution attached to it, and have enough nearby context to decide the sentence is still true.

What an engine actually does with a number

Numbers are what a model reaches for when its own weights are not enough. Dates, prices, adoption rates and market sizes are exactly the facts that go stale between training runs, so the retrieval step goes looking for them. In the sites we audit we have seen statistics pages fetched by assistant crawlers more often than their search traffic would predict. We cannot see inside an engine, only what arrives in a server log, so treat that as a pattern in the logs we have read rather than a measured rate.

One academic result points the same way. The GEO paper, accepted to KDD 2024, evaluated nine content changes against its own GEO-bench benchmark and measured how much of the generated answer each source won. The paper reports that its top methods, which include Statistics Addition, achieved a relative improvement of 30 to 40 percent on the position-adjusted word count metric over an unoptimised baseline. It also reports that domains such as Law and Government, and opinion-type questions, benefit notably from adding statistics. That is a measured effect on one synthetic benchmark evaluated with GPT-3.5 as judge, not a law that holds across live engines, and the same paper states the efficacy of these methods varies across domains.

What the paper does not tell you is the second half of the trade. A number travels well. Attribution does not. Once a figure is in an answer, the next writer copies the figure and links to whoever the engine named, which may not be you.

The anatomy of a number worth reusing

Write each statistic as a standalone unit. The test is brutal and simple: cut one sentence out of the page, show it to someone who has never seen your site, and ask whether they could defend it in a meeting.

A unit that passes has five parts in it.

The figure, written once, in a form that does not need the heading to make sense. The population, which is the thing counted and the thing not counted. The window, the dates between which the measurement was taken. The method, one clause about how, with the sample size if there is one. The origin, a link to the dataset, the filing or the paper, not to an article about the dataset.

Most pages have the first part and skip the other four. That is the whole gap. "AI referral traffic grew 10x" is unusable. The same claim written as a unit would name the domains counted, the share of sessions at the start and end, the months between them and how the referrer was identified. Those are invented placeholders, not our figures, and the point is the shape: a sentence in that form can be quoted by anyone, including a model, without the quoter taking on risk they cannot assess.

This is the same property that makes any page quotable, and it is worth reading what makes a page citable by an AI engine alongside this, because the mechanics overlap. It is also why comparison pages that get cited look so similar structurally: both formats win by being easy to cut up.

Markup, including the fields almost nobody fills

Prose alone leaves the parser to guess. Schema.org has types built for this and they are underused.

If your page publishes a dataset, Dataset is the type, and Google's dataset documentation requires only name and description, with the description between 50 and 5,000 characters. The recommended properties are where the trust lives: identifier, license, sameAs, variableMeasured, temporalCoverage, citation and measurementTechnique. Google states it will also read the equivalent structure expressed in W3C's Data Catalog Vocabulary, so a team already publishing DCAT does not need a second representation.

For individual figures rather than whole datasets, Observation is the type to know. As of schema.org v30.1, released 16 September 2026, it carries observationAbout, observationDate, observationPeriod, measuredProperty, measurementMethod, marginOfError and value. A marginOfError field in public markup is close to a costly signal. Nobody publishes an error bar on a number they invented.

None of this forces a citation. What it does is remove the ambiguity that makes a cautious system hedge, and hedging reads as absence.

The failure modes

The orphan figure. A number appears in a chart image with the source in a caption three paragraphs away. The passage that gets retrieved contains the number and nothing else. Keep the source in the same sentence or the one immediately after it.

The compounding copy. Your number gets quoted by a site that does not link, then quoted again from that site. Within a few months the figure is attributed to a chain you are not in. There is no clean fix, but a distinctive phrasing and a named methodology make the trail traceable, and an identifier or DOI makes it much harder to lose.

The frozen page. A statistics page dated 2024 that still says "updated for 2026" in the title is worse than one that admits its age. Whether an engine weights recency is not something we can observe from outside, and we have not found a published weighting in Google's, OpenAI's or Perplexity's documentation, but a date in the markup that disagrees with a date in the prose is a contradiction a parser can see without any judgement at all.

The number you cannot defend. Publishing a plausible figure with a vague attribution can look good for a while, because retrieval does not verify arithmetic. It verifies form. Build on that and the asset collapses the moment anyone checks.

Access. A page nobody can fetch is not a source. OpenAI documents four agents with different jobs, three of them relevant here, and notes that because ChatGPT-User fetches are started by a person, robots.txt rules may not apply to them, which changes how you reason about a blanket block. Published IP ranges exist for each agent, so you can verify a hit in your logs rather than trust a user-agent string.

Five checks you can run this week

  1. Grep your statistics page for digit strings and count how many have a four-digit year

    2. within thirty words. We use roughly four in five as our own working threshold in audits; it is a house rule of thumb, not a measured cutoff. Falling well short of it is a provenance problem, not a writing one.

  2. Open the page with JavaScript disabled. If the numbers are rendered client side, they do

    3. not exist as far as most fetchers are concerned.

  3. Pick your three best figures, search each as an exact phrase, and see who is quoting you

    4. without a link. That list is your outreach list.

  4. Validate the JSON-LD and confirm dateModified agrees with the visible date.
  5. Filter server logs by the published crawler IP ranges, not the user-agent header, and

    6. check whether the statistics page is being fetched at all.

The self-check covers the rest of the page-level work in the same spirit. For the figures themselves, we keep a running list of AI search statistics worth quoting with the source and date attached to each, which is also a working example of the format argued for here.

Honest limits

This post does not establish that adding statistics raises citations on a live engine. The one measured result here is from a 2024 benchmark paper scored by a language model, and a benchmark is not ChatGPT, Perplexity or AI Overviews. We cannot separate how much of any uplift is the number itself from how much is the sentence structure that tends to surround a number, and we have not run that test. Our log observations come from the sites we audit, which are a self-selected sample and not a census of the web, and a server log shows a fetch, never a citation. The markup fields described here are documented by Google and schema.org, but neither party says that filling them causes a citation, and we are not claiming they do: as assistants begin to transact directly, the question of who gets credited for a fact is tangled further by agent payments and the trust signals in the way.

FAQ

How many statistics should a page carry? Fewer than you think, each fully sourced. Twenty defensible figures outperform two hundred loose ones, because the loose ones drag the credibility of the whole page down when one is checked and fails.

Do I need original research? It helps, and it is not the only route. A well-maintained curation where every entry has a date, a method and a working link is a real contribution, and it is a lot cheaper than running your own survey.

Should the page be one long list or many small pages? One page per question beats one page for everything, unless the figures genuinely belong to a single dataset. Retrieval works at passage level, so the risk of one page is dilution, not length.

What if my source is behind a paywall? Say so in the sentence. An honest "paid report, Q2 2026" is better than a link that returns a login screen, which reads as a broken source to a fetcher.

Does updating the page help? Changing the date alone does not. Changing the figures, and saying which ones changed and when, gives a system a reason to prefer your version over the copy of it published elsewhere.

Sources

All links below were opened and checked on 12 October 2026.

Related reading

What makes a page citable by an AI engineComparison pages that get citedAI search statistics worth quotingAgent payments and the trust signals in the wayDoes AI-written content get cited less by AI enginesAnswer-first writing, and why the opening paragraph carries the answer