What changed when AI Overviews launched

May 2023 was not the launch

Three boundaries matter more than everything else in this chronology, and they are the three most often got wrong. Search Generative Experience (SGE) was announced at Google I/O on 10 May 2023 as an opt-in Search Labs experiment, US English only. AI Overviews launched at I/O in May 2024 to all US users, with the SGE branding retired — the first time ordinary American searchers saw a generated answer without having chosen to. AI Mode was announced on 5 March 2025 and became generally available in the US on 20 May 2025, and it is a separate surface rather than a rename: Ahrefs measured only 13.7% citation overlap between the two across 540,000 query pairs in December 2025, despite 86% semantic similarity in the answers themselves.

The consequence is arithmetic rather than editorial. A before-and-after study anchored to 2023 is anchored to an opt-in experiment almost nobody used. A study treating AI Mode dates as AI Overviews dates is comparing systems with different default models and largely different citation sets. One imprecision is worth carrying honestly: sources give either 14 May 2024, the keynote date, or the week following, and Wikipedia gives the month without a day. Use "May 2024" unless you can cite the keynote.

The rollout, dated

AI Overviews reached the United Kingdom, India, Japan, Brazil, Mexico and Indonesia in August 2024, a hundred-plus countries by October 2024, and 200-plus countries and 40-plus languages by May 2025. AI Mode followed its own path: India in July 2025, the United Kingdom on 28 July 2025, the European Union on 8 October 2025 with France excluded over neighboring rights, and France finally on 22 July 2026 — the last major withheld market for both surfaces. By 27 July 2026, Similarweb data reported by TechCrunch put AI Overviews on 43% of US searches, up from 15% a year earlier.

Underneath the geography, the models changed repeatedly and quietly. AI Mode moved to Gemini 3 Flash on 17 December 2025 and to Gemini 3.5 Flash on 19 May 2026; AI Overviews took Gemini 3 as its global default on 27 January 2026. None of those swaps was announced as a change to citation behavior, and all of them plausibly are one.

Two rules follow. A single date is never a rollout date, because announcement, availability and default are three different events. And a window straddling a model change measures two systems — which matters because no behavior changelog exists for either surface, so the public record of when these systems changed is a patchwork assembled after the fact.

What became impossible to measure

Clicks and impressions from AI Overviews have always been folded into the main Search Console performance report under the Web search type, undifferentiated from ordinary organic results. Google counts those clicks and has never broken them out. Every click-through-rate study in this field therefore works by comparison — keywords with an overview against keywords without, or the same keywords before and after — and none of them measures AI Overview clicks directly.

The counting rules also differ. Google confirmed in August 2024 that an AI Overview impression registers only when the link is scrolled or expanded into view, while a classic organic impression is counted when the result set is generated. The two are not on the same scale, and their ratio is not a click-through rate.

The Search Console generative AI performance report, live on 3 June 2026, did not close the gap: impressions only, AI Overviews merged with AI Mode, broken down by pages, countries, devices and dates, with no query dimension, no clicks and no average position. Carolyn Shelby put it accurately in Search Engine Journal on 4 August 2026 — Google "has given us a new diagnostic lens, not a new scoreboard."

The click studies, one at a time

Six bodies of work are quoted as "the AI Overviews traffic studies." They differ enormously in method, sample and who paid, and treating them as one literature is why the public numbers look incoherent.

Pew Research Center, 22 July 2025

Nine hundred US adults on the Ipsos KnowledgePanel who consented to browsing-data collection; 68,879 unique Google searches during March 2025, of which 12,593 produced an AI summary; result pages re-scraped by a third party between 7 and 17 April 2025 to identify which had carried one. Users clicked a traditional result on 8% of visits where a summary appeared against 15% where it did not, clicked a link inside the summary on 1% of visits, and ended the session on 26% of pages with a summary against 16% without. Funded by Pew Research Center, a nonpartisan nonprofit subsidiary of The Pew Charitable Trusts — the only large study here with no commercial interest in the answer. Its weaknesses, fairly stated: observational, so some of the gap is query mix; the result pages were reconstructed weeks later; and tracking-consenting panel members are not a random sample. Pew published it in full.

Agarwal and Sen, SSRN, 3 April 2026

The only randomized evidence in the field. Agarwal at the Indian School of Business and Sen at Carnegie Mellon's Heinz College recruited 1,065 US desktop Chrome users through Prolific and ran a browser extension that randomized whether the AI Overview was hidden on each search, across 68,089 searches, with overviews appearing on roughly 41% of them. Outbound organic clicks fell 39.8%, zero-click searches rose 34.5%, and removing the overview from the top position produced an 88% relative increase in outbound clicks. Randomization removes the query-mix confound that limits Pew. The finding that makes it read as honest is the one cutting against publishers: no difference in satisfaction, perceived information quality or ease of finding information. Users were fine; the sites were not. Limits: a working paper rather than peer-reviewed at the time of writing, US desktop Chrome only, and a paid self-selected panel.

Ahrefs, 17 April 2025

Three hundred thousand informational keywords split 150,000 with an AI Overview and 150,000 without as a control, comparing position-one click-through rate in March 2024 against March 2025 and subtracting the control group's natural decline. The raw rate fell from 0.073 to 0.026; the residual attributed to AI Overviews is 34.5%. Ahrefs sells SEO software and has a commercial interest in search looking difficult; its method here is published, control-grouped and reproducible in outline, which makes it the best-constructed vendor study in the set. Ahrefs published the method with the result.

Seer Interactive, November 2025 and April 2026

Agency client data rather than a random sample: 3,119 informational queries across 42 organizations, 25.1 million organic and 1.1 million paid impressions, June 2024 to September 2025. Organic click-through rate on AI Overview queries fell from 1.76% to 0.61%, paid from 19.7% to 6.34%. The caveat Seer states and the headlines drop is that queries without overviews fell 41% over the same period; when the control declines almost as fast as the treatment, most of the headline is not attributable to it. Seer's later data, across 5.47 million queries and 53 brands, complicates it further — between September and October impressions went from 15.8 million to 33.1 million while clicks went from 398,798 to 400,271, halving the rate arithmetically. Only in November did clicks fall.

Authoritas, Similarweb and Enders, 7 August 2025

Press Gazette collated three at once. Authoritas reported clickthrough down 47.5% on desktop and 37.7% on mobile — figures submitted with a complaint to the UK Competition and Markets Authority, the underlying study unpublished. Figures produced for a regulator by a complainant are advocacy documents; that does not make them wrong, but they were not selected neutrally. Similarweb reported zero-click searches across the top hundred global news and media sites rising from 50.5% to 52.7% year over year, and zero-click news searches from 56% in May 2024 to almost 69% in May 2025 — a proprietary panel, and two figures with different denominators that should not share a trend line. Enders Analysis, using Sistrix data across six UK publishers, found "some correlation" and did not claim six sites generalize.

Google's answer, and what is not in it

Liz Reid, Vice President and Head of Google Search, published Google's rebuttal on 6 August 2025: "total organic click volume from Google Search to websites has been relatively stable year-over-year"; "we're actually sending slightly more quality clicks to websites than a year ago"; and that this contrasts with third-party reports which "inaccurately suggest dramatic declines in aggregate traffic — often based on flawed methodologies."

The post contains no numbers. It refers to billions of clicks twice without quantifying them, and gives no baseline, no window beyond "year-over-year", no tolerance defining "relatively stable" and no breakdown by site type. "Quality clicks" is defined only as clicks where users do not quickly click back — under which cutting low-quality clicks registers as improvement even if the total falls. Two weeks earlier a Google spokesperson had told Search Engine Land that Pew's work used "a flawed methodology and skewed queryset that is not representative of Search traffic", without saying which queries would be representative.

The correct handling is narrow, and it is where most coverage fails in one direction or the other. A first-party assertion with no disclosed data is a position, not a finding, and belongs in a report attributed and dated rather than laundered into fact. Equally it has not been disproved: an aggregate can hold steady across all of Search while informational publishers take severe damage, because a total survives a shift in its own distribution. Google holds the only data that would settle it, which makes the claim unfalsifiable rather than false.

The numbers that get confused with each other

Five confusions account for most of the incoherence in public discussion of this subject.

• The two 34.5% figures. Ahrefs found a 34.5% reduction in position-one click-through rate; Agarwal and Sen found a 34.5% increase in zero-click searches. Different quantities, different samples, coincidentally identical digits, cited interchangeably.
• A rate change read as a traffic change. Seer's later data shows the rate halving while absolute clicks held, because impressions doubled. Publishers are paid in the numerator.
• The control group dropped from the headline. Minus 61% on AI Overview queries against minus 41% on queries without them describes a far smaller attributable effect than the first figure alone.
• Switched denominators. Ahrefs found AI Overviews on 9.46% of its indexed keywords globally in May 2025; Similarweb found them on 43% of US searches in July 2026. Both can be true, because keywords and searches are different populations.
• "AI Overviews reduced traffic 70%." In wide circulation, traceable to no study, and apparently a headline compression of the Seer or Authoritas figures.

The lesson underneath all five: there is no site-agnostic impact figure. Measured effects vary by query type, position, vertical and whether the site was cited, and the numbers in circulation answer five different questions on five different samples.

What the launch changed, and what it did not

The damage was concentrated rather than general. Ahrefs found 97.7% of AI Overview keywords to be informational; queries that are transactional, or aimed at reaching a specific site, barely trigger the feature. Publishers of explanatory and how-to content absorbed the measured decline, and commentary generalizing from them to all of search is generalizing from the worst-affected group.

Being cited does not restore what was lost, or not by the route it is sold. Seer's finding that cited brands earned 35% more organic and 91% more paid clicks compares cited brands to uncited ones rather than the same brand before and after citation, so selection is uncontrolled: brands that get cited may be brands that were already winning. Pew's 1% in-summary click rate caps the direct route. The plausible mechanism is brand exposure feeding clicks elsewhere on the page.

The famous failure cases of late May 2024 were real, and Google published a response and narrowed triggering. No study established what share of AI Overviews contained comparable errors, then or since; those screenshots circulated because they were funny, not because anyone sampled.

What the launch reliably produced was vocabulary. Generative engine optimization (GEO) — changing content in the hope that a generative system retrieves, cites or repeats it — moved from a November 2023 academic preprint to a venture-funded category label by May 2025. Google's position on the vocabulary arrived on 28 August 2025, when Danny Sullivan told the keynote audience at WordCamp US, the annual United States conference of the WordPress community, that "Good SEO is good GEO, or AEO, AIO, LLM SEO, or LMNOPO."

Frequently asked questions