AI SEO statistics

What the grade distribution shows

Ten of sixteen at partial is not an accident of which pages happened to get written. It is what happens when the organisations best placed to measure something sell the measurement. Visibility vendors have a direct commercial reason not to publish the method behind their corpus. A figure can therefore carry a real company name, a real sample size, and still not be reproducible by anyone outside it. That is a structural feature of the field, not a failing of any individual vendor. Grading around it honestly, rather than flattening every row to the same confidence, is what this page is for.

The master table: 16 categories, graded

Traceable means the study behind it publishes its sample, method and date. Partial means a named source stands behind it but the full method is not public. Broken chain means the claim does not line up across the sources we checked, or the research needed to support it does not exist. This hub holds no original numbers: every figure lives on the linked page, with its own sourcing and method notes, which is why the hub cannot go stale relative to its own sources.

How a grade gets assigned

Grades are not impressions. Each comes from running the same five checks, in order, against the underlying source — the procedure documented in the provenance audit. Pass all five and a figure earns traceable. Pass one through three but fail four and it is partial. Fail check one, two or five and it is broken.

The order matters as much as the checks. A figure that fails check one is not worth running through the other four, and running them anyway is how a plausible-sounding number acquires an air of scrutiny it never survived. Check four — is the method described well enough to repeat? — is where most of this field lands: a vendor can name itself, state a sample size, and still never describe how the measurement worked. That earns partial, not traceable. Applying the same procedure to a claim you want to be true is the only genuinely difficult part.

What each page found, and why it carries that grade

AI Overviews. The share of AI Overview citations coming from Google's top-10 results fell from roughly 76% to roughly 38% year over year, per Ahrefs' measurement of AI Overview citations against the top 10, covered independently by Search Engine Journal. Graded traceable: the same organisation ran both measurements with a comparable method and disclosed sample sizes both times. That makes it one of the few genuine trend claims in the field rather than two unrelated snapshots. Semrush's AI Overviews study is the closest independent comparison, though its method differs enough that the two should not be treated as replications of each other.

AI Mode and query fan-out. Google's own patent US11663201B2 documents eight sub-query type categories, and Google's later comments on fan-out in AI Mode confirm the mechanism without quantifying it. The widely repeated "8-16 sub-queries per question" figure could not be traced to any disclosed-method study. Graded partial: the mechanism is documented, the specific count is folklore.

ChatGPT. Citations skew heavily toward Wikipedia and encyclopedic sources. Google ranking predicts ChatGPT citation weakly — around 12% overlap, far below AI Overviews — a figure that traces to Ahrefs' overlap study with a disclosed sample. The source-skew figure is vendor-reported from a non-public corpus, which is why the page carries partial. Seer Interactive's finding that most SearchGPT citations match Bing's top results complicates the picture usefully: low overlap with Google does not mean low overlap with every conventional index.

Claude. The least-measured major surface. No large-scale citation study exists publicly, and the comparative work that does exist — such as Discovered Labs' comparison of how each assistant chooses sources — is qualitative rather than measured at scale. That absence is itself the finding.

Perplexity. Citations skew heavily toward Reddit. Ranking overlap sits around 33%, notably higher than ChatGPT's, making it the engine where classic SEO transfers most directly. The overlap figure is traceable; the Reddit-skew figure is vendor-reported, from the same class of corpus described in Profound's platform citation patterns.

Gemini. Graded broken for a structural reason rather than a sourcing one. Gemini, AI Overviews and AI Mode are three distinct Google surfaces that coverage routinely conflates, so figures attributed to "Gemini" frequently describe something else.

Cross-platform domain skew. Each engine has a distinct centre of gravity: encyclopedic for ChatGPT, community for Perplexity, multimodal for AI Overviews. All figures come from one vendor corpus that is not independently public. Partial throughout. The concordance page puts the reported domain overlap between ChatGPT and Perplexity at roughly 11% — vendor-reported, two engines only, and if it holds up, the strongest argument in the field for measuring each engine separately.

Crawler economics. AI crawlers take dramatically more pages per referral than classic search crawlers do, spanning three orders of magnitude between the most and least extractive. The underlying observations come from Cloudflare's crawl-to-click work and its breakdown of who is crawling the web, with Search Engine Journal's summary the version most people actually read. Reproduced via secondary aggregators rather than pulled directly, hence partial.

Market share. AI referrals remain around 1% of the average site's traffic while growing fast — a share consistent with Similarweb's generative-AI usage panel and dwarfed by conventional search in StatCounter's market share series — and ChatGPT's share of the AI category fell as competitors gained. The page's real contribution is the denominator warning: a within-category percentage quoted as a whole-traffic figure is the most common error here. Referral traffic carries the same caveat, plus the likelihood that standard analytics undercount it because much AI-sourced reading never produces a click at all.

Zero-click. Graded traceable, and the most useful disambiguation page on the site. Three separate legitimate studies produce numbers near 58% that measure entirely different things. The figures are fine. Collapsing them into one is the problem.

Conversion. Graded broken, and deliberately so. Published claims range from roughly 4.4x to 23x organic conversion. That is not a range, it is different orders of magnitude measuring different things, and no source reconciles them.

GEO tactics. Graded per-tactic rather than as a page. Four of fourteen tracked tactics carry fact-grade evidence, all from the single peer-reviewed GEO benchmark study. Most of the rest are widely recommended and untested, including several supported only by correlational write-ups such as Zyppy's citation ranking factors.

Commerce. Reported growth in AI-driven store traffic is real and substantial, from what was likely a very small base. The much-quoted forecast about agent-influenced transactions is a projection, not a measurement — and the shift it anticipates, described in Previsible's account of agentic shopping, would change what a "visit" even means.

Local search. Graded broken as a structural gap. Local intent is a large share of all search behaviour and rigorous AI-search research on it is almost entirely absent. The page states that plainly instead of borrowing a general-purpose number.

Which figures survive scrutiny

The traceable-graded figures on this site have three things in common, and together they describe what good measurement looks like in this field.

The AI Overview ranking-overlap trend. Same organisation, two points in time, comparable method, disclosed sample sizes. That combination is what supports a genuine trend claim rather than two disconnected snapshots. Most "X changed to Y" claims here fail precisely because the two numbers came from different methods.

The Princeton GEO study's effect sizes. Peer-reviewed, disclosed method, a stated benchmark of roughly 10,000 queries, and a built-in negative control that actually came back negative. That last detail matters more than most readers notice: a study that finds an effect everywhere it looks has not demonstrated it can detect a failure.

The SparkToro zero-click figure. A named clickstream panel, a stated geography, a stated date. It is frequently misused, but the underlying measurement is sound. The failure is in how people quote it, not in how it was produced.

None of these are the most exciting numbers in the field. The exciting ones tend to be the vendor-reported ones. That inverse relationship — between how quotable a figure is and how well it is sourced — is the single most useful thing to internalise from this page.

Which figures do not, and for which of three reasons

Conversion benchmarks: irreconcilable spread. Claims ranging from 4.4x to 23x cannot all describe the same quantity. Some measure session-level conversion, some a modelled revenue effect, some do not say. Until someone publishes definitions alongside the numbers, quoting any single figure from that range is picking one at random and presenting it as consensus.

Gemini and local search: structural gaps. Neither is broken because a specific number failed a check. They are broken because the research needed to produce a reliable number has not been done at all. Publishing a borrowed figure from an adjacent surface would fill the gap dishonestly.

Terminology adoption: no disclosed-method dataset. Every "GEO vs AEO, which is winning" page eventually shows a trends chart. We could not find one with a stated method behind it, so that page states the gap instead of fabricating a chart to fill the section.

Naming these plainly is uncomfortable, and it is the point. A statistics hub that grades everything favourably is a marketing asset. One that flags its own weakest categories is a reference.

Five failure modes that recur across the field

Grading sixteen categories one by one surfaces patterns no single page shows on its own. Recognising these is more durable than memorising any individual figure — every grade above is ultimately an assessment of how many of the five a figure has survived.

One: a small number of original sources, endlessly re-cited. Trace enough figures and the same handful of names appear underneath nearly all of them. A number that appears in forty articles can still be one measurement. Ubiquity is not corroboration.

Two: the measuring parties have a commercial stake. The organisations best positioned to measure AI citation at scale are visibility vendors whose data is the product they sell. Not an accusation of dishonesty — a structural reason why full methods rarely get published, and why so many rows here carry partial.

Three: populations get dropped in transmission. A figure is measured on a specific, stated subset. Each time it is repeated the qualifier gets shorter. Several steps later it describes everything. This is the single most common way an accurate number becomes a false claim.

Take a bare claim like "68% of searches now end without a click." Follow the link and it often reaches another blog post rather than a study. Check the population and it may cover one platform rather than all searches. Check the date and it may predate the product it is being used to describe. Three checks, a few minutes, and most roundups skip all three.

Four: correlation gets written up as causation. Almost every finding in this field observes that cited pages tend to have some property. Very few show that adding the property causes citation. The language used to report them rarely preserves that distinction.

Five: nothing gets re-measured. Figures are published once and quoted for years, against products that change every few months. A study accurate when run may describe a system that no longer exists. Very little here gets a second measurement, which is why the AI Overview trend figure stands out so sharply for having one.

Which statistics this site refuses to publish

A statistics hub is defined as much by what it leaves out. Several widely circulated figures are absent here on purpose, and naming them is more useful than silently omitting them.

A single composite "AI visibility score." Every major vendor has one. Every one is unfalsifiable by design, because the weighting between inputs is proprietary. Two scores of 78 from different tools are not comparable, and neither can be checked from outside.

Bare "AI Overviews cut clicks by X%" claims. Not because the underlying studies are bad, but because the figure is meaningless without its population attached. The disambiguated version lives on the zero-click page.

Precise-sounding percentages with no named source. A figure like "73.2% of marketers say" fails check one. Suspicious precision is itself a signal: a number stated to one decimal place, from a method that could not support that resolution, usually indicates the precision was invented rather than measured.

Forecasts presented as measurements. Projections about what share of transactions AI agents will influence can be reasonable and still are not data. Where this site references one, it is labelled a projection rather than folded into a statistics table.

Anything from a source we could not identify. Several genuinely interesting figures circulate with no traceable origin. They may well be accurate. Without a source there is no way to know, and repeating them would add to exactly the problem this hub exists to document.

How to quote a figure from this site

Cite the original source, not this site. Where a figure comes from Ahrefs or a peer-reviewed paper, name them. For most rows this site is a grading layer over other people's data, and citing the intermediary when the primary is available weakens your claim. Carry the qualifier with the number. "Vendor-reported from a non-public corpus" is four words; dropping them turns a partial-graded figure into something that reads as settled fact.

State the population and the date. Every percentage describes some specific set of things, and these products change fast enough that an undated figure has no shelf life. Check the linked page before you publish — grades move in both directions.

Every grade here should be contestable. Take any row, open the linked page, find the figure and follow its citation. Reach a dataset with a stated sample and the row should be traceable; a named organisation with no published method, partial; another blog post, or two figures that will not reconcile, broken. Where your answer differs from ours, ours is the one that needs defending. Who maintains this and on what basis is set out on the about page — a grading system is only as trustworthy as the disclosed interests of whoever assigns the grades.

The five biggest measurement gaps

Reading sixteen graded categories together also makes the holes visible. These are the questions where a reliable number would change how people work, and where none currently exists.

One: does anything a site controls causally change citation rate? The most important open question, and the least answered. Nearly everything published is correlational. Until randomised tests exist, most tactical advice rests on association. Two: how long does a citation last? Every study is a snapshot; none follow the same citations forward, so a citation that vanishes in two weeks and one that persists a year are currently indistinguishable.

Three: do retrieval bots execute JavaScript? Binary, foundational, cheap to test, still untested publicly. Four: what does Claude cite? The least-measured major surface, with no large-scale study at all. Not technically hard — simply undone. Five: how much do engines actually disagree? The ~11% concordance figure covers two engines from one vendor corpus. Whether it holds across more engines, or varies by query type, is unmeasured.

The first-party studies meant to replace these rows

A hub that only grades other people's numbers is a critique, not a contribution. The point of grading is to identify which rows deserve to be rebuilt on measurement this site controls. That programme sits under the AI Citation Index, with predictions published before data collection begins — a prediction written after seeing results is not a prediction. Everything running or planned is listed on the studies index.

Three routes in, depending on what you need:

• Mechanism questions — the citation half-life study and the JavaScript rendering test address gaps two and three above.
• Tactic questions — the schema markup study tests a recommendation that circulates constantly on thin evidence; the closest existing public work is Ahrefs' look at schema and AI citations, and the freshness claim usually rests on pieces like this account of freshness in AI search.
• Null results — the null results registry exists so findings that come back flat get published rather than shelved. The field's best existing example is Ahrefs' study across 300,000 domains, which found no clear llms.txt effect on citations, as Search Engine Journal reported; our log-based replication is the llms.txt page.

A negative result published plainly is worth more to this hub than three positive ones without a method.

Changelog and update cadence

This hub updates in substance whenever a linked page changes, because it holds none of the numbers itself. The common failure of a statistics roundup — where the roundup keeps quoting a figure its source has since revised — is structurally impossible here.

Grades get reviewed whenever new evidence lands, and can move in either direction. A partial figure becomes traceable if the vendor publishes its method. A traceable figure becomes partial if a re-measurement uses an incomparable method. A broken figure becomes partial or traceable if someone reconciles the conflicting sources. That downgrade direction is the one most reference pages quietly avoid: a reference that only ever upgrades its own confidence is not tracking evidence, it is tracking optimism.