Blog · Data and trends

AI search statistics worth quoting

Published 10 August 2026 · Updated 12 October 2026 · 8 min read

AI search statistics worth quoting
Photograph: Arturo Añez / unsplash

The short answer: Quote only the numbers whose denominator you can see. Four clear that bar: Pew Research Center's US browsing panel, where a Google page carrying an AI summary produced a result click on 8 percent of visits against 15 percent without one in March 2025, Cloudflare's July 2025 crawl-to-refer ratios from its own network, the KDD 2024 GEO paper measuring what adding statistics does to a page's share of a generated answer, and Google's documentation saying no separate optimisation exists for AI Overviews.

Everything else in the slide deck needs a trace-back before it goes in front of a board.

The four that survive a trace-back

Clicks fall when a summary is present, but the base rate was already low. Pew Research Center recruited 900 US adults from a probability panel and recorded their real browsing in March 2025. Across 68,879 Google searches, 12,593 returned an AI summary. On pages with a summary, a traditional result got clicked on 8 percent of visits. On pages without one, 15 percent. The link inside the summary itself was clicked on 1 percent of visits. Sessions ended on the page 26 percent of the time with a summary, 16 percent without. That is a sample and not a census: one US panel, one month, Google only. The honest reading is not that AI halved your traffic. Even without a summary, 85 percent of those visits produced no result click at all, so the summary widened a gap that was already there.

Crawling and referring are not the same business. Cloudflare published per-company crawl-to-refer ratios for July 2025, in a post dated 29 August 2025: roughly 1,091 crawls per referred visit for OpenAI, about 38,000 for Anthropic, and 5.4 for Google. That is one network's view of the platforms it can identify, not an industry census, and it says nothing about traffic that never passes through it. Treat the numbers as a shape, not a constant. The Anthropic figure had already fallen by 87 percent between January and July 2025, and Google's moved between 3.8 and 22.5 inside those same seven months. What the ratio tells you is directional: for the platforms in that table, a crawler can read a site very heavily while sending almost nobody, which is the whole reason How much traffic AI search actually sends is a separate question from how often you are cited.

Adding statistics changed how much of a page got used, on the engines that paper tested. The GEO paper presented at KDD 2024 built GEO-bench, a 10,000-query benchmark, and tested nine content edits. Its three strongest edits, citing sources, adding quotations and adding relevant statistics, raised position-adjusted word count in the generated answer by 30 to 40 percent, with the best method at 41 percent. Keyword stuffing went the other way: the authors report little to no improvement on their own generative engine, and 10 percent below baseline when they repeated the test on Perplexity. Those results come from one research engine and one commercial engine as they stood in 2024, so the direction is the durable part, not the decimals.

Google says there is no separate lever. Search Central's page on AI features states plainly that there are no additional requirements and no special optimisations needed to appear in AI Overviews or AI Mode, that AI feature impressions are folded into the ordinary Performance report under the Web search type, and that robots.txt, nosnippet, data-nosnippet, max-snippet and noindex remain the controls. That page carries a last-updated date, 10 December 2025 when we checked it, which is reason enough to re-open it rather than quote someone's summary of it. For Google's surfaces, a third-party claim that contradicts the first-party document is the weaker of the two. We cannot say the same for engines that publish no comparable document.

Why most of the other numbers are laundered

Watch what happens to a figure over four hops. A study measures 200 pages in one vertical and reports a rate with its sample frame attached. A newsletter quotes the rate and drops the frame. A blog quotes the newsletter and rounds it. A deck quotes the blog and calls it an industry benchmark. Nothing was fabricated at any step. The denominator simply fell off, and what arrives on the slide is a number with no population behind it.

There is a newer failure mode on top of that one. Writers producing time-sensitive content with language models need figures to fill gaps, and generated text reaches for whatever looks like a sourceable statistic. We have not found a published, reproducible demonstration of a fabricated figure travelling from a new site into assistant answers and back out into human writing, so we are not claiming that it has been shown. We raise it because the loop needs nobody to lie. It needs only that nobody opens the link, and that is a condition you can check on your own citations this afternoon.

That is also why Statistics pages: how to build one engines trust is about method and provenance rather than volume. A page of numbers with visible sample sizes and collection dates is a different object from a page of numbers.

Three questions that kill a bad statistic

What is the denominator, and who is in it? "40 percent of buyers start in ChatGPT" is meaningless until you know whether the sample was 300 self-selected newsletter subscribers or a probability panel. Pew's figure is quotable partly because the panel design and the 68,879-search base are published alongside it.

When was it measured, and against which engine? These surfaces move under you: Google's AI features page carries a revision date, and Cloudflare's Google ratio went from 3.8 to 22.5 and back to 5.4 inside seven months. A citation-rate study from eighteen months ago may describe a model and a retrieval stack that have since been replaced, and from outside you usually cannot tell. Dates belong in the sentence, not in the footnote.

Who paid for the measurement, and what do they sell? A number produced by a company that sells the fix is not automatically wrong. It is automatically in need of a second source. When the only available figure comes from a vendor, say so in the sentence and let the reader discount it themselves.

The ten-minute check before you quote anything

Open the citation. Then open its citation. Keep going until you reach something with a methodology section, or until the chain dead-ends in a blog post that links to another blog post. Plenty of numbers die at hop two or three. We have not counted how many across the field, so take that as a warning rather than a rate.

For the ones that survive, write a one-line provenance note next to the figure in your own notes: source, sample, date, measured or modelled. Six months later that note is the difference between defending the number in a meeting and quietly dropping the slide. We keep one for every figure that appears in a client report, and the sample audit report shows what that looks like when the evidence sits next to the finding rather than in an appendix.

One more discipline worth adopting: separate traffic numbers from quality numbers. They move independently, and conflating them produces the most common bad slide in this category, the one where a small referral count is dismissed before anyone has checked what those sessions did. Do AI referrals convert better exists because the two questions have different answers.

What we cannot tell you

We have not found a credible public figure for how often a brand mention inside an answer leads to a purchase. Several of these surfaces pass no referrer we can see in client analytics, and a mention often produces no click at all: in the Pew study, the link inside a Google AI summary was clicked on 1 percent of visits. Anyone quoting a conversion rate for AI answer mentions is modelling, not measuring, and should say which assumptions they used.

Nor have we found a stable, published citation-share figure per engine that holds across industries. Each audit we run is a sample rather than a census: roughly 45 prompts across five engines, with retrieval on and off, run on the dates of that audit. The spread we see between verticals is wide enough that we will not publish a cross-industry average from it. If that prompt set returns almost nothing for your category, that is a finding, and When answer engine optimisation is not worth doing yet is the honest next read.

Honest limits

This post checks four sources. It does not survey the field, and three of the four rest on a single measurement each: Pew's figures come from one US panel of 900 people, one month, Google only; Cloudflare's ratios are one network's view of the platforms it can name; the KDD 2024 results come from one research engine plus one commercial engine as they stood in 2024. None of them tells you what an assistant will do with your pages this quarter, and we have no measurement of our own that fills the gap. The fabricated-statistic loop above and the hop-two-or-three attrition rate are both cautions we could not source, and we have flagged them as such in the text. Our own audits are samples of about 45 prompts across five engines on a given date, which is enough to find a problem on one site and not enough to publish a benchmark.

FAQ

Which single AI search statistic is safest to quote? The Pew click comparison, 8 percent against 15 percent, because the panel, the date and the search base are all published and the effect size is modest enough to be plausible. Quote it with its March 2025 date and its scope, a US panel on Google, attached.

Are vendor-published AI visibility studies usable at all? Yes, with the label on. Describe the finding, say it is a vendor estimate, and avoid building a forecast on it alone. The problem is not that vendors measure things. The problem is a figure travelling without the sentence that qualified it.

Does putting statistics on my own pages help me get cited? The KDD 2024 GEO-bench results show a lift of 30 to 40 percent in how much of an optimised page appeared in the generated answer for the three strongest edits, one of which was adding relevant statistics. That was measured on one research engine and one commercial engine in 2024, so treat the direction as more durable than the percentage. We have no equivalent measurement for today's engines.

How often should I refresh a statistic I use publicly? Re-open the source every quarter and check whether it has been updated or superseded. Engine documentation carries a last-updated date and network reports get re-issued, so both will tell you when they moved.

Sources

All four were re-opened and checked against the sentences that cite them on 12 October 2026.

Related reading

How much traffic AI search actually sendsDo AI referrals convert betterStatistics pages: how to build one engines trustWhen answer engine optimisation is not worth doing yetCitation churn, and how fast the cited set changes