AI search CTR statistics

What this page covers, and what zero-click covers

These two pages are constantly conflated, so the line is worth drawing before any number appears. This page is about how many clicks survive — the rate at which a shown result still gets clicked when a generated answer is present, and how that rate should be segmented before it means anything. Zero-click search statistics is about whether a click happens at all — the share of searches that end without any onward visit, which is a property of the search session rather than of your listing.

The practical difference: a zero-click figure can rise while your CTR is unchanged, because the composition of sessions moved rather than anyone's click behaviour on a given result. And your CTR can fall while the zero-click rate is flat, because clicks redistributed to someone else. Mixing the two produces the single most confused genre of statistic in this field, and it is where the widely repeated 58% claim comes from.

The one well-sourced number in this category

It earns its place here because it is the number people reach for when arguing about clicks, and it is not about clicks. It says the overlap between ranking well and being cited is loosening — reported independently elsewhere and traced in full on the AI Overview statistics hub.

The best-documented public attempt at an actual click figure is Ahrefs' measurement of how much AI Overviews reduce clicks to top-ranking pages, which at least states its population. Most figures repeated in roundup posts do not, for the reasons set out in the provenance audit.

The four outcomes a single "CTR" number blends

A headline "CTR dropped 40%" almost never says which of the top two rows it describes, and the responses are opposite. The first is a citation-value question: if the value you are collecting is brand exposure rather than sessions, it belongs in the same bucket as brand mentions. The second is a pure-loss question: a page that used to get organic clicks now gets fewer, for a reason unrelated to its own ranking quality.

Full table, with populations

Classic SEO has a well-documented, widely replicated position-to-CTR curve, produced independently many times over many years. That maturity is exactly what AI-search CTR measurement lacks: no comparably rigorous, independently replicated curve exists for citation position within a generative answer. Vendor teardowns of which factors correlate with citation and of how ChatGPT selects sources describe selection, not the click that follows it. Google's own documentation of AI features describes the surface without publishing click data at all.

Where the click decision actually happens

A click is a decision made in about a second: the reader scans what is on screen and asks whether they already have what they came for. An AI answer changes the inputs to that decision in two separable ways. It puts a synthesised answer above the ranked list, so the reader reaches a stopping point sooner. And it pushes the ranked list further down the screen, so fewer readers ever see it.

Those two respond to different actions. The first is answer sufficiency — if the generated answer fully resolves the question, ranking first will not save the click. The second is screen space: the link is fine, but fewer eyes reach it. Only the first is something a publisher can act on by changing the page, and a page that merely restates the same short answer gives a reader no reason to click through. A calculator, a comparison table or a worked example does — the same lever the zero-to-cited log study ended up pulling.

Hypothesis A third effect is worth naming without a measurement behind it: seeing a source named inside an answer may make a reader more likely to click it, because the answer has already vouched for it. If real, it partly offsets the other two. Vendor claims that AI-search visitors convert far better point in that direction without isolating it; the conversion benchmarks page handles that claim properly.

Confounds that move CTR at the same time

None of these are exotic, and all are present in a typical site's data at once. Query mix drift is worse than it looks, because AI systems fan a single question out into many sub-queries — see the query fan-out corpus study for what that does to the population you think you are measuring.

How to measure your own CTR impact

Export query-level Search Console data for the trailing 90 days. Tag each query as AI-Overview-triggering or not — the AI Overview checker finds them. Within the triggering group, split by whether your page appears among the citations. You now have three comparable groups: cited-and-triggered, not-cited-but-triggered, and a non-triggering baseline. A meaningful gap between the first two suggests citation still carries click value. A similar drop across both, relative to baseline, suggests the presence of an overview reduces clicks regardless of citation status. A single blended average hides exactly that distinction.

Six mistakes recur. Averaging across queries with wildly different volumes — weight by impressions, or report the distribution. Assuming AI-answer presence is stable — it changes week to week and by location, so a query tagged once and treated as fixed will misclassify part of your data. Comparing groups that are not comparable — if your triggering queries are informational and your control is navigational, the difference you find is a query-type difference wearing a costume.

Ignoring position — match on average position, or at least report it beside the result. Reading a ratio when the denominator moved — always print raw clicks next to the ratio. And stopping when the number looks bad: a drop is the start of a question, not the answer to one. The visibility measurement standard sets out the reporting conventions in full.

Why a CTR figure does not cross a product boundary

The differences are structural: where citations appear, how many are shown, and whether links are visible without a hover — citation conventions differ per platform, as do the underlying selection patterns. Hypothesis Intent differs too — someone typing into a search box and someone mid-conversation have different tolerance for leaving the page — and we have no cross-surface measurement to point at. Ahrefs' work on how little these engines overlap in what they cite is the clearest evidence they are separate populations, and the cross-platform concordance study measures the same thing on a fixed panel.

Audience size compounds it. Search share, Bing's user base, ChatGPT's monthly users and Perplexity's differ by orders of magnitude — the market share hub keeps those in one place, with the caveat that they count different things. A CTR effect on a small surface is a small effect on your business whatever the percentage says. And several engines lean on a conventional index underneath: Seer found 87% of SearchGPT citations matching Bing's top results. Never carry a percentage across a product boundary; if a figure does not name the surface it was measured on, it is not usable.

Common misreadings of a CTR drop

"Our CTR fell, so AI Overviews are taking our traffic." A CTR fall with stable clicks and rising impressions means something else entirely: you are being shown for more queries, including ones you do not win.

"We got cited, so the click loss does not matter." That is a business judgement, not a measurement. Decide explicitly what a citation without a click is worth, and write it down before you look at the data.

"The drop is X%, matching the published figure." Two numbers can agree by coincidence, especially when the published one blends populations unlike yours. Agreement across different methods is meaningful; agreement between one measurement and one headline is not.

"Traffic is flat, so nothing changed." Flat totals hide compensating shifts. Segment before concluding nothing happened.

Acting on a mis-measured figure has real costs. It means deprioritising content that was working, or starting a rewrite treadmill where teams shorten pages so there is less to summarise and lengthen them so they feel substantial — both guesses, neither with a replicated study behind it. The opposite error is just as expensive: concluding citations are worth nothing because they produce no measurable clicks. The gap between mentions and measurable visibility is real rather than an excuse, and hard to measure is not the same as worth zero.

Who this applies to, and who it does not

Most useful to sites whose traffic concentrates in informational queries. Reference content, how-to guides, definitional pages and comparisons sit closest to what a generated answer can summarise. It applies less where the query is a step in a transaction: booking, checkout, account access and branded navigation are poorly served by a summary. Hypothesis We would expect these least affected, with no segmented public data confirming it.

It applies unevenly by size, too. A site with a handful of high-volume queries can inspect them one by one; one with tens of thousands of long-tail queries needs the statistical approach. Local queries and YMYL topics behave differently enough that a general figure is close to useless for either. And if organic search is a small share of your total traffic, this analysis may not earn its cost.

Reporting it upward, lead with raw clicks rather than the ratio, show the three segments side by side, and state the confounds you could not control. "Our average position also fell during this window, so part of this is not about AI answers" costs nothing and buys the rest of the report credibility.

A field precedent: the last time clicks moved

Two things were true across all the earlier episodes. Some queries genuinely stopped producing clicks. And the headline figure was almost always more dramatic than the segmented reality. A selection effect explains the gap: the people who measure and publish are the people who were affected, and sites that saw no change had nothing to write up. Hypothesis We think the same skew is operating now. That is not a claim that nothing is happening; it is a claim about which stories get written down.

The second lesson is more testable. The queries that lost clicks in earlier episodes were mostly the ones whose answer fitted in a sentence; deeper questions kept producing clicks because a sentence could not satisfy them. Whether that pattern repeats here is the version of this question we would most like to see measured.

A null result is a real possibility. If we ran it and found no detectable effect of citation status once position and query type are controlled — or an effect far smaller than the headlines — that goes to the null results registry. So does the most common outcome of all: a failure to measure, because AI-answer detection was too unreliable.

Open questions

Open question Does being cited inside an answer raise click probability, relative to an equivalent uncited link at the same position? Nobody has isolated this cleanly.

Open question Does the length of an AI answer predict the click rate of the links beside it?

Open question Does the source you displace matter? Answers lean on a small set of frequently cited domains, and whether a click behaves differently when the answer cites a forum rather than a publisher is unmeasured.

Open question Do readers who click through after an AI answer behave differently once they arrive? They may be better qualified, having self-selected past a summary.

Open question How stable is AI-answer presence for a given query over weeks? Everything here assumes it is stable enough to tag. If it is not, much published measurement in this area has a problem nobody is discussing.

Verification status

The one strong figure here is the AI Overview rank-overlap trend — Traceable , same organisation, disclosed samples both times, and a citation-composition measure rather than a click rate. Bare "CTR dropped by X%" claims are treated as Broken chain as a category until a population-qualified, disclosed-method version exists. A platform change exposing AI-answer presence in first-party reporting would be the largest possible improvement, since it removes the tagging step that makes this analysis expensive and error-prone for everyone doing it independently. The same grading scheme runs across every statistics hub here, including GEO statistics and AEO statistics; the standards behind it are set out in the method notes.