The published estimates range from about 5% to about 58%, and the range is not noise. It tracks study design almost perfectly: the weaker the identification, the larger the number.

The two studies with genuine causal designs are also the two nobody quotes, because their findings are less alarming and they were not published by companies that sell search software. Meanwhile the platform itself has responded with adjectives.

This page sorts every available estimate by how it was produced, and says what actually determines your own exposure.

The two studies that can support a causal claim

Both are recent, neither is peer reviewed, and both are far better identified than anything else available.

The randomized experiment. Over 1,000 US participants recruited through a research panel, each installing a purpose-built browser extension, randomly assigned to a control group, a group with summaries hidden, or a group redirected to the conversational mode. Data collected 7 January to 10 February 2026, two weeks of observation per participant.

What it found, verbatim. That the appearance of a summary reduced users’ organic clicks to third-party sites by 39.8 percent and increased searches where the user clicked on no links at all by 34.5 percent. Citations within summaries generated relatively little referral traffic, about 8% of all clicks.

Why the design matters. Random assignment removes the query-type confound by construction. Whether you saw a summary was decided by the experiment, not by what you searched for.

The natural experiment. A difference-in-differences analysis using online encyclopedia clickstream data from December 2023 to December 2024, exploiting the staggered geographic rollout together with the multilingual structure of the encyclopedia. Two matched samples: 499,927 English-German article pairs and 530,873 English-French pairs.

What it found, verbatim. That default summary availability reduced English search traffic by 5.45% and 4.82% respectively, which the authors translate to roughly 100.27 million fewer search-originated visits per month.

Why that design is also strong. The same article in another language serves as its own control, which holds content and query type constant.

And why the two disagree. The authors of the second study address this directly, noting their effect is considerably smaller and attributing the gap to the panel’s nature and the opt-in design of the experiment. They also describe their own estimate as a reduced-form effect of the default regime rather than the effect conditional on a summary actually appearing, which would attenuate it.

Published estimates of the effect of generative search summaries on organic clicks, ordered by study designTable of the published estimates of how generative search summaries affect organic clicks, ordered by the strength of the study design rather than by the size of the headline figure, showing that the range across studies tracks identification strength. A randomized field experiment recruited more than one thousand United States participants through a research panel, each installing a purpose-built browser extension, randomly assigned to a control group, a group with summaries hidden, or a group redirected to conversational mode, with data collected between the seventh of January and the tenth of February two thousand and twenty-six over two weeks per participant; it reports that the appearance of a summary reduced organic clicks to third-party sites by thirty-nine point eight percent and increased searches with no link clicked at all by thirty-four point five percent, with citations inside summaries generating about eight percent of all clicks. A difference-in-differences analysis of online encyclopedia clickstream data from December two thousand and twenty-three to December two thousand and twenty-four exploited the staggered geographic rollout together with the encyclopedia’s multilingual structure, using two matched samples of four hundred and ninety-nine thousand nine hundred and twenty-seven English-German article pairs and five hundred and thirty thousand eight hundred and seventy-three English-French pairs; it reports reductions in English search traffic of five point four five percent and four point eight two percent respectively, equivalent to approximately one hundred point two seven million fewer search-originated visits per month. A behavioural panel of nine hundred United States adults captured sixty-eight thousand eight hundred and seventy-nine unique searches of which twelve thousand five hundred and ninety-three triggered a summary, finding link clicks on eight percent of visits where a summary appeared against fifteen percent where none did, one percent clicking a source cited within the summary, and sessions ending on twenty-six percent of pages with a summary against sixteen percent without. Commercial search tool analyses using aggregated search console data on three hundred thousand keywords report a thirty-four point five percent reduction in one version and approximately fifty-eight percent in a later version of the same analysis.Sorted by design, not by headlineCan support a causal claimRandomized field experiment1,000+ participants, Jan-Feb 2026−39.8%Random assignment removes the query-type confound by constructionDifference-in-differences~1M article pairs, Dec 2023-Dec 2024−5.45% / −4.82%The same article in another language is its own controlObserved behaviour, but not randomizedBehavioural panel900 adults, 68,879 searches8% vs 15%Real navigation, but summary queries are not randomAggregated console data, published by companies selling search softwareTool study, first version300,000 keywords, 2024 vs 2025−34.5%Same study, updatedSame keywords, 2023 vs 2025−58%The weaker the identification, the larger the number. That pattern is the finding.
The range tracks identification strength. The two studies that can support causation report the smaller effects. Source : Published studies, opened directly (2026)

What people actually did, on a real panel

Between the experiments and the tool studies sits one observational study of genuine navigation, and its numbers are the most concrete available.

The sample. 900 US adults on a panel that records real browsing through an installed application. 68,879 unique Google searches captured, of which 12,593 triggered an AI summary. Browsing observed through March 2025, with results pages scraped in April 2025 to identify retrospectively which queries had produced a summary.

The headline comparison. Users clicked a traditional link on 8% of visits where a summary appeared, against 15% where none did.

Clicks on sources inside the summary. 1% of visits.

Session endings. 26% of pages with a summary ended the session, against 16% without.

What it establishes. That in real navigation, the presence of a summary is associated with roughly half the link-clicking. That is a substantial, well-measured association.

What it does not establish. Causation. The queries that trigger summaries are not a random subset, and the study makes no causal claim.

Why the 1% matters more than it looks. The common reassurance is that citations inside summaries send traffic instead. On this panel they sent almost none, and the randomized experiment reached a compatible conclusion, putting citation traffic at about 8% of all clicks.

Deployment timeline of generative search summaries by marketTimeline of the staggered geographic deployment of generative search summaries, which forms the basis against which every published study measures its effect and explains why studies conducted in different markets are not directly comparable. The feature launched experimentally in May two thousand and twenty-three under an earlier name at the platform’s annual developer conference. On the fourteenth of May two thousand and twenty-four it was renamed and launched by default in the United States at the same conference. In August two thousand and twenty-four it extended to the United Kingdom, India, Japan, Brazil, Mexico and Indonesia. On the twenty-eighth of October two thousand and twenty-four it extended to more than one hundred countries. In early two thousand and twenty-five it became the default across the European Union. By May two thousand and twenty-five it covered more than two hundred countries and territories in more than forty languages. On the twenty-second of July two thousand and twenty-six it launched officially in France. This staggered rollout is what supplied the control group for the difference-in-differences study, which exploited the fact that English-language content was exposed by default while German and French language content was not yet, and it is also the reason that studies conducted in different markets and different periods report different magnitudes without necessarily contradicting one another. No official figure has been published by the platform regarding what proportion of queries display a summary; third-party estimates range from roughly twenty percent to roughly fifty percent depending on the tracking methodology, panel and geography used, and those estimates are not comparable with one another.The rollout every study measures againstMay 2023Experimental14 May 2024US, by defaultAug 2024Six more markets28 Oct 2024100+ countriesEarly 2025EU by defaultMay 2025200+ countries22 Jul 2026FranceThis stagger is what made the natural experiment possibleEnglish content exposed by default while German and French content was not yet, so the sameencyclopedia article could serve as its own control across languages.And it is why studies in different markets are not comparableA France study nine days after launch and a US study two years after it are measuring differentstages of the same thing. No official figure exists for what share of queries shows a summary;third-party estimates run from roughly 20% to roughly 50% on incomparable methods.
Staggered geography is what gave the natural experiment its control group. It is also why studies disagree by market. Source : Platform announcements, cross-checked against the academic timeline (2026)

Why the tool studies read higher

Not because anyone is lying. Because of a structural problem the design cannot escape.

The confound. 99.2% of keywords that trigger a summary are informational. Comparing queries that have summaries against queries that do not is therefore also comparing informational queries against everything else, and those have different click rates regardless of any AI feature.

What the better tool study does about it. Restricts both sides to informational keywords, which removes part of the problem. It still compares different queries rather than the same query before and after.

What the France study does about it. Names the issue and declines to resolve it. Its authors write that a slope is not a prediction, acknowledge late-July seasonality, and describe the summary rate as a risk factor rather than a cause.

What the client-data study does about it. States it plainly: that it cannot claim causation, and that higher-authority brands are also more likely to be cited.

So the honesty is mostly present. The caveats are in the source documents. They fall off in transmission, which is how a carefully hedged correlation becomes a headline percentage.

And the commercial position matters. Every one of these studies was published by a company selling search software or search consulting. That does not make the data wrong, and their willingness to publish samples and caveats is better than most of the industry. It belongs in the citation.

How each study of generative search summaries handles the query-type confoundTable recording how each published study of generative search summaries handles the structural confound arising from the fact that ninety-nine point two percent of keywords triggering a summary are informational, which means any comparison between queries with and without a summary is also a comparison between query types that have different click rates independently of any artificial intelligence feature. The randomized field experiment handles the confound natively through random assignment, since whether a participant saw a summary was determined by the experiment rather than by the query, eliminating selection by query type by construction. The difference-in-differences study handles it through design as well, comparing the same encyclopedia article in an exposed language against the same article in an unexposed language, which holds both content and query type constant. The behavioural panel of nine hundred adults observes real navigation but does not randomize, so it compares visits where a summary appeared against visits where none did without controlling for why the summary appeared. The commercial search tool study restricts both sides of its comparison to informational keywords, which removes part of the problem, but still compares different queries rather than the same query before and after. The France study of nine hundred and sixty-three domains names the issue without resolving it statistically, its authors writing that a slope is not a prediction, acknowledging late July seasonality and describing the summary rate as a risk factor rather than a cause. The client-data study of fifty-three brands states plainly that it cannot claim causation and that higher-authority brands are also more likely to be cited. The caveats are therefore present in the source documents themselves and are lost during transmission, which is the mechanism by which a carefully hedged correlation becomes a headline percentage.99.2% of summary keywords are informationalSo comparing “with summary” against “without” also compares query types.StudyHow it handles the confoundVerdictRandomized experimentRandom assignment removes itSolvedDifference-in-differencesSame article, other language, as controlSolvedBehavioural panelObserves real navigation, does not randomizePresentTool study, USRestricts both sides to informational keywordsPartlyTool study, FranceNames it, declines to resolve itAcknowledgedClient-data study”We cannot claim causation.”AcknowledgedThe caveats are all in the source documents”a slope is not a prediction”. “We cannot claim causation. Higher-authority brands are alsomore likely to be cited.” They fall off in transmission, not in the research.
The caveats are in the source documents. They fall off in transmission. Source : Study documents, opened directly (2026)

Google’s answer contains no numbers

The platform has responded twice at senior level. Both responses are worth reading closely for what they do not contain.

August 2025, from the head of Search. That overall, total organic click volume from Google Search to websites has been relatively stable year-over-year.

The quality claim, verbatim. That average click quality has increased and that Google is sending slightly more quality clicks to websites than a year ago, defining quality clicks as those where users do not quickly click back.

The attack on the studies. That this data is in contrast to third-party reports which inaccurately suggest dramatic declines in aggregate traffic, often based on flawed methodologies, isolated examples, or traffic changes that occurred prior to the rollout of AI features.

What the post contains in support. No figure. No percentage. No chart. No sample, period or denominator.

July 2026, from a senior vice president. That Google is now sending billions of clicks to websites every week through AI features in Search alone. Again no denominator, which is the number that would make it meaningful.

How to weigh this. Google is the only party that can see the whole picture, and its assertion may well be right. But an assertion criticising others for flawed methodology, while publishing no methodology of its own, cannot settle the question. It is a claim, at the same evidential level as the studies it dismisses, minus the samples.

Content of the platform’s two public responses on search traffic and the supporting data providedAnalysis of the two senior public responses issued by the search platform regarding the effect of generative summaries on website traffic, together with what each contains by way of supporting evidence. In August two thousand and twenty-five, the head of Search stated that overall total organic click volume from Google Search to websites has been relatively stable year over year, and separately that average click quality has increased and that the platform is sending slightly more quality clicks to websites than a year ago, defining quality clicks as those where users do not quickly click back. The same post criticised third-party research, stating that its data is in contrast to third-party reports that inaccurately suggest dramatic declines in aggregate traffic, often based on flawed methodologies, isolated examples, or traffic changes that occurred prior to the rollout of artificial intelligence features in Search. The post contains no figure, no percentage, no chart, no sample size, no observation period and no denominator. In July two thousand and twenty-six, a senior vice president stated that the platform is now sending billions of clicks to websites every week through artificial intelligence features in Search alone, again providing no denominator, which is the quantity that would render the statement interpretable. The appropriate weighting is that the platform is the only party able to observe the complete picture and its assertion may be correct, but an assertion which criticises others for flawed methodology while publishing no methodology of its own cannot settle the question, and stands at the same evidential level as the studies it dismisses while lacking their disclosed samples.What Google published, and what it did notAugust 2025, head of Search”total organic click volume from Google Search to websites has been relatively stable year-over-year""average click quality has increased… those where users don’t quickly click back”Same post: third-party reports are “often based on flawed methodologies, isolated examples”July 2026, senior vice president”billions of clicks to websites every week through AI features in Search alone”. No denominator.What the posts contain in supportNo figure. No percentage. No chart.No sample, period or denominator.How to weigh it fairlyGoogle is the only party that sees everything,and may well be right.But an assertion criticising others for flawed methodology, published without one, cannot settle it.
An assertion criticising others for flawed methodology, published without a methodology. Source : Google blog, August 2025 and July 2026 statements (2026)

Your exposure is your query mix

The global percentage is not your percentage, and the thing that determines yours is knowable from your own data.

The mechanism. Summaries appear overwhelmingly on informational queries. If your organic traffic comes from brand searches and transactional queries, your exposure is small. If it comes from guides, definitions and comparisons, it is large.

Which describes B2B unusually well. B2B organic strategy is built on informational content: what is X, how does Y work, X versus Y. That is precisely the query class where summaries appear.

One tracking vendor’s sector figures. Summary coverage on B2B technology queries rising from 36% to 82% in a year, with health at 72% to 88%, while commerce queries were held far lower.

The caveat on those figures. They come from a proprietary rank-tracking parser rather than observed behaviour, published by a company selling search software. Treat them as an order of magnitude and a direction.

A widely circulated B2B statistic I am not going to repeat. A specific comparison of summary frequency on B2B versus consumer keywords circulates everywhere. I could not trace it to any study with a disclosed sample, so it is not in this article.

The number you can actually compute. Open Search Console, split your organic clicks into brand, informational and transactional, and look at the informational share. That fraction is your exposure. Everything above is context for interpreting it.

Determining an individual site’s exposure to generative search summaries from its own query mixDiagram describing how an individual website can determine its own exposure to generative search summaries from its own query data rather than from any published global percentage. The underlying mechanism is that summaries appear overwhelmingly on informational queries, with ninety-nine point two percent of summary-triggering keywords being informational according to one commercial study, so a site whose organic traffic derives from brand searches and transactional queries carries small exposure while a site whose traffic derives from guides, definitions and comparisons carries large exposure. This distinction describes business-to-business organic strategy with unusual precision, since business-to-business content programmes are built on informational material of exactly the class where summaries appear. One tracking vendor reports summary coverage on business-to-business technology queries rising from thirty-six percent to eighty-two percent within a year, with health rising from seventy-two to eighty-eight percent, while commerce queries were held substantially lower; those figures derive from a proprietary rank-tracking parser rather than from observed user behaviour and are published by a company selling search software, so they should be treated as an order of magnitude and a direction rather than a measurement. A further widely circulated statistic comparing summary frequency on business-to-business against consumer keywords could not be traced to any study with a disclosed sample and is therefore excluded. The calculation an individual site can actually perform is to open its search console data, divide organic clicks into brand, informational and transactional categories, and examine the informational share, since that fraction constitutes the site’s exposure and every published global figure serves only as context for interpreting it.The global percentage is not your percentageSplit your own organic clicks three waysBrand queriesYour company name, yourproduct names.Exposure: lowInformational queriesWhat is X, how does Y work,X versus Y.Exposure: this is the partTransactional queriesPricing, demo, quote,supplier, near me.Exposure: lowWhy B2B sits badly hereB2B organic strategy is built on informational content. That is exactly the query class summaries cover.One vendor’s sector figuresB2B technology coverage 36% to 82% in a year.With the caveat attachedRank-tracking parser, not observed behaviour.Your informational share is your exposure. Everything else on this page is context for reading it.
Split your own organic clicks three ways. The informational share is the part at risk. Source : Method, applied to the studies' own finding on query type (2026)

What to do, given genuine uncertainty

Five moves that are correct under any of the published estimates.

Compute your informational share first. Before reacting to a global figure, find out how much of your organic traffic is even in the exposed category.

Stop treating rankings as the outcome. Position is now a weaker proxy for traffic than it was, because the click rate attached to a position has changed. Measure clicks and conversions, not positions.

Judge content on assisted revenue, not sessions. A page that gets summarised and cited may still influence a purchase. If your only measure is sessions, you will delete pages that are working.

Write for the question, not the keyword. Content that answers a specific question with something a summary cannot reproduce, your own data, your own method, a named case, is what earns the click that remains.

And keep the channel in proportion. If organic informational traffic was 15% of your pipeline, a 40% reduction on part of it is a manageable problem. Sizing it before reacting is the whole discipline here.

Where to go next

You want the technical requirements underneath. Technical SEO: three requirements.

You are moving the site. Website migration without losing rankings.

You are wondering about page speed. Website speed: what matters.

You want the paid side of acquisition. Paid advertising vs organic.

You are building the measurement layer. What GA4 does not show.

You want the demand model behind content. Demand generation vs lead generation.

In short

  • The published range is 5% to 58%, and it tracks study design: the weaker the identification, the larger the number.
  • The randomized experiment on 1,000+ participants found a 39.8% reduction in organic clicks and a 34.5% rise in searches with no click at all.
  • The natural experiment on about a million article pairs found 5.45% and 4.82%, roughly 100 million fewer visits a month.
  • The behavioural panel of 900 adults found link clicks on 8% of summary visits against 15% without, and 1% clicking a cited source.
  • Citation traffic is small in both: about 1% on the panel, about 8% of all clicks in the experiment.
  • 99.2% of summary keywords are informational, which is the confound the tool studies cannot fully escape.
  • Google says traffic is “relatively stable” and click quality has risen, in posts containing no figure, percentage or chart.
  • None of this is peer reviewed. The two strongest studies are 2026 preprints.

Compute your informational share before you react to anyone’s percentage. Book a diagnostic, or see how we approach B2B websites.