Google says AI Overview clicks are stable. Its own guidance says otherwise.

Executive summary

In August 2025, Google's VP and Head of Search, Liz Reid, published a blog post stating that organic click volume "has been relatively stable year-over-year," that average click quality "has increased," and that Google is sending "slightly more quality clicks to websites than a year ago." The post explicitly dismissed third-party decline reports as products of "flawed methodologies, isolated examples, or traffic changes that occurred prior to the roll out of AI features" (Confidence: High — primary source, verbatim quotes confirmed).

Every independently published study since then has found the opposite, using entirely different methods. Ahrefs' analysis of 300,000 keywords (via aggregated Google Search Console CTR data) found a 34.5% average CTR reduction on top-ranking desktop pages when an AI Overview is present (Confidence: High — primary source, corroborated by multiple outlets). Pew Research, working from real browsing-panel data on 900 US adults, found only 8% of visits with an AI summary present included a click to a traditional result, versus 15% without one (Confidence: High — primary Pew report). SparkToro's Rand Fishkin, using Similarweb panel data, found 68.01% of US Google searches in early 2026 ended with no click anywhere — up from 60.45% in 2024 (Confidence: High — primary source). None of these three teams talked to each other, used the same data source, or shares a commercial incentive to agree with one another. They agree anyway, on direction and on rough order of magnitude.

The more interesting story, though, isn't Google versus its critics — it's Google versus itself. In January 2026, Danny Sullivan — by then a Director within Google Search, having stepped down as Search Liaison the previous August — told the Search Off the Record podcast that Google does not want publishers "crafting anything for Search specifically," including breaking content into chunks for AI retrieval (Confidence: High — primary source, verbatim). Two months earlier, Ahrefs had published data showing that AI Overview citations are increasingly pulled not from a query's top-10 organic results but from sub-queries Google's own "fan-out" process generates — a mechanism that structurally rewards exactly the kind of atomized, individually-extractable content Sullivan says publishers shouldn't bother building (Confidence: High — primary Ahrefs source). Google is telling publishers not to optimize for a retrieval mechanic that Google's own engineering is visibly built around.

A timeline / comparison table

Part 1 — What Google's own people are saying, in their own words

Liz Reid's August 2025 post is the anchor document for Google's public position, and it is more carefully worded than most coverage of it suggests. It does not claim click quality "slightly increased" as a single assertion — it makes two distinct claims: that average click quality "has increased," and, separately, that Google is sending "slightly more quality clicks" to websites than a year ago (Confidence: High — primary source, precise wording matters here). The post frames third-party traffic-decline reports as methodologically unsound, citing "isolated examples" and pre-AI-rollout baseline comparisons as the likely culprits (Confidence: High).

That framing has been echoed, with variation, by other named Google staff. Nick Fox, SVP of Knowledge and Information, told the AI Inside podcast in December 2025 that optimizing for Google's AI experiences is "very similar, I would say the same, as how to perform well in traditional search" (Confidence: Medium — quote directly fetched, but not independently verified against a full transcript). Gary Illyes reportedly told an Asia-Pacific Search Central Live audience to "simply use normal SEO practices," adding "you don't need GEO, LLMO or anything else" (Confidence: Medium — the quote traces to one attendee's secondhand recap, corroborated by three independent outlets repeating it identically, but no primary Google transcript exists). John Mueller, asked whether structured data helps large language models understand a page, reportedly gave a characteristically qualified answer — "yes, no, and it depends," noting some SERP features (Shopping-type results, for instance) depend heavily on schema while ranking itself does not (Confidence: Medium — corroborated via two independent outlets after a primary source page returned an HTTP error on direct fetch).

The pattern across these statements is consistent: nothing new is required, existing SEO fundamentals transfer, and reports of harm are overstated or mis-measured. It's a coherent institutional line. Whether it's an accurate one is a separate question — and it's the one every outside measurement has answered differently.

Part 2 — What the independent measurements found, and why the disagreement in method matters more than the disagreement in numbers

Three organizations, three unrelated methodologies, one direction.

Ahrefs' April 2025 study used aggregated Search Console CTR data across 300,000 keywords — 150,000 with an AI Overview present, 150,000 without — comparing March 2024 to March 2025, and found a 34.5% average CTR reduction on top-ranking desktop pages when an AI Overview appears. The study explicitly contrasted its finding with Sundar Pichai's public claim that AI Overview citations receive higher CTR than a standard top organic ranking (Confidence: High — primary source; contrast with Pichai's claim confirmed verbatim).

Pew Research took a completely different approach: real browsing-panel data from 900 US adults, covering 68,879 unique Google searches in March 2025, of which 12,593 returned an AI summary. Only 8% of AI-summary visits produced a click to a traditional result, against 15% for non-AI-summary visits. Users were also more likely to abandon the search session entirely after an AI-summary page — 26%, versus 16% without one (Confidence: High — primary Pew report, methodology and figures confirmed).

SparkToro's Rand Fishkin used Similarweb's US desktop-and-mobile panel and found that in January–April 2026, 68.01% of Google searches ended with no click to any destination — the open web, Google's own properties, or paid results — up from 60.45% in 2024. That is a broader "zero-click" measure than "no click to the open web," and should be read precisely: it means fewer than a third of all Google searches now send a click anywhere at all (Confidence: High — primary source; publication date June 9, 2026, covering Jan–Apr 2026 data).

Laurence O'Toole, CEO of UK measurement firm Authoritas, added a fourth data point in the same week as Reid's post: his firm's research found a 47.5% desktop and 37.7% mobile CTR drop when AI Overviews appear (Confidence: High — quote and figures from Press Gazette, primary article fetched). Four instruments — GSC aggregation, real-panel browsing behaviour, Similarweb panel data, and Authoritas' own crawl-and-measure methodology — landed on the same sign despite diverging magnitudes (more on why that matters below).

Part 3 — The named, on-record pushback

The day after Reid's post went up, Press Gazette collected on-record reactions from five named SEO practitioners, none of whom pulled punches.

"Every single study done on the impact of AI Overviews, from every single source, shows the exact same thing: A significant drop in clicks." — Barry Adams, Polemic Digital (Confidence: High — primary Press Gazette article, verbatim)

Glenn Gabe, an independent SEO consultant, said Google "needs to stop the BS and just explain that AIOs and AI mode are causing drops in traffic" (Confidence: High). Laurence O'Toole asked, pointedly, "If they were genuinely sending more traffic to websites then surely Google would show data and charts" (Confidence: High). Aleyda Solis called the framing "half-baked half-truths" (Confidence: High). David Buttle, formerly of the Financial Times and now at DJB Strategies, noted that Reid's post "claims it is to 'share some insights' but it contains no data whatsoever," calling the "quality clicks" framing "desperate conjecture and spin" (Confidence: High).

Notably, one quote attributed to Buttle in earlier reporting — that Google's behaviour is "demonetising segments of the media" — does not appear in the Press Gazette article and should not be used; it is dropped here for lack of a verifiable source.

Part 4 — The contradiction Google is having with itself

This gets less coverage than the Reid-post backlash, but it matters more for anyone doing the work rather than just reading about it.

In January 2026, Danny Sullivan — again, by this point a Director within Google Search, not the sitting Search Liaison, a title he left in August 2025 — told the Search Off the Record podcast: "So we don't want you to do that… We don't want people to have to be crafting anything for Search specifically. That's never been where we've been at and we still continue to be that way," specifically addressing the practice of chunking content into discrete, LLM-retrievable segments (Confidence: High — primary source, verbatim).

Set that next to what Ahrefs published seven weeks later. Its March 2026 study of roughly 4 million AI Overview citation URLs across 863,000 keywords found that the share of citations pulled from a query's top-10 organic results fell from approximately 76% in July 2025 to 37.9% in March 2026 (counting all SERP-feature blocks; 37.1% counting organic blue links only). Ahrefs attributes the fall to "query fan-out" — Google generating sub-queries whose own separate SERPs, increasingly populated by YouTube results, supply citations independent of the primary query's top-10 ranking (Confidence: High — primary Ahrefs source, precise dual figure).

Fan-out retrieval, by construction, rewards content that can be sliced into a self-contained, directly-answerable unit and matched against a narrow sub-query — which is functionally what "chunking" means. Sullivan is telling publishers not to bother optimizing for a mechanism that Google's own citation data shows is doing exactly that kind of matching. Jeff Dean, Google DeepMind's Chief Scientist, gave the clearest public description of the underlying retrieval funnel on the Latent Space podcast: "You're going to attend to trillions of tokens, but you're going to want to identify what are the 30,000-ish documents that are with the maybe 30 million interesting tokens. And then how do you go from that into what are the 117 documents I really should be paying attention to in order to carry out the tasks that the user has asked me to do?" (Confidence: High — primary source, "117 documents" figure confirmed verbatim). That is a description of a system built to isolate small, well-defined, individually retrievable units of information from a vast corpus — which is not obviously compatible with "don't craft anything for Search specifically."

Note for design/editorial: a diagram is warranted here — a funnel visual mapping Dean's stated retrieval cascade (trillions of tokens → ~30,000 documents → ~30 million tokens → 117 documents) against Ahrefs' fan-out citation-share collapse (76% → 37.9%), to make visually explicit that the "don't optimize for us" messaging and the actual retrieval architecture are pulling in opposite directions.

There is also a precedent for treating Google spokesperson statements as provisional rather than settled. Gary Illyes said at PubCon Austin in September 2023 that links are not a top-three ranking factor and haven't been "for some time" (Confidence: High — primary source). Seven months later, at SERP Conf in Bulgaria, he went further — "We need very few links to rank pages… Over the years we've made links less important" — only to walk it back on X after Ahrefs' Patrick Stox quoted it back at him: "I shouldn't have said that… I definitely shouldn't have said that" (Confidence: Medium — event and walkback corroborated across multiple outlets, not independently transcript-verified). That is not evidence that current statements about AI Overview clicks are wrong. It is evidence that Google's own public statements on ranking mechanics have, at least once on the record, needed retracting within the same year they were made — which argues for treating any single quote as a snapshot, not a settled policy.

Part 5 — Where this is (and isn't) playing out on stage

Google Search Central Live held its first Canadian edition in Toronto on April 21, 2026 — a half-day, in-person event (not recorded or livestreamed) with sessions on AI in Search, crawling and indexing, Search Console, and structured data. Confirmed speakers included Danny Sullivan, Martin Splitt, Daniel Waisberg, Annanya Raghavan, and Ryan Levering (Confidence: High — confirmed via Google's own Search Central blog and multiple independent recaps). Sullivan's presence confirms he remains an active Google Search spokesperson well after his title change.

BrightonSEO's 2026 programme is set to close with Crystal Carter, Head of AI Search and SEO Communications at Wix, presenting "How to get your website ready for AI Agents" (Confidence: Medium — session title confirmed via two independent pages, though a direct fetch of the brightonseo.com listing itself did not surface the session, likely due to JavaScript rendering). SMX Advanced 2026 in Boston closed on June 5 with an "Ask the Experts" panel featuring Duane Forrester (UnboundAnswers.com, ex-Microsoft), Michael King (iPullRank), and Martha van Berkel (Schema App), covering AI visibility, entities, schema, and brand discoverability (Confidence: Medium — panel, date, and speakers confirmed; no independent source verifies a specific panel title or any on-record disagreement among panelists, and none should be implied).

None of these three events staged a direct confrontation between "Google says clicks are stable" and "every independent study says they aren't" — a gap the next section takes up directly.

What everyone is missing

The methodological diversity is the strongest evidence, and it's underused. Commentators keep arguing over whose percentage is more accurate — 34.5% versus 47.5% versus a zero-click rate crossing 68% — when the more load-bearing fact is that four measurement approaches with no shared data source, no shared incentive, and no coordination all point the same direction. A single study can be an artifact of its method. Four independently-designed studies converging on direction, even while diverging on magnitude, is closer to a finding than a dispute.

Google's real contradiction isn't with its critics — it's internal. Reid, Fox, and Illyes are all making a coherent, defensible case that fundamentals haven't changed. But Sullivan's anti-chunking guidance and Ahrefs' fan-out citation data describe two different realities of how content actually gets pulled into an AI Overview. One message says don't bother restructuring for retrieval; the documented mechanics say retrieval is increasingly happening at the level of narrow, atomized answers to sub-queries Google itself generates. Both statements can be sincere and still be in tension.

"Just do normal SEO" is true and nearly useless as guidance. Mueller's "yes, no, and it depends" on schema is the more honest version of the same message Illyes gave at Search Central Live APAC. The instruction not to build anything special is defensible in the sense that there's no secret trick — but it tells a publisher nothing about which existing practices (structured data for certain SERP features, clear entity definitions, extractable answer formatting) actually matter more now than they did two years ago.

The conference circuit is managing the disagreement, not resolving it. Search Central Live Toronto, BrightonSEO, and SMX Advanced all kept the same senior Google-adjacent voices on stage through 2026, each restating variations of the "nothing has fundamentally changed" line, without a confirmed session that put a Google speaker and an independent measurement author on the same panel to work through the gap. That absence is itself informative about how contained this disagreement is being kept.

Future predictions

• Expect Google to publish at least one more "clicks are healthy" post before the end of 2026, timed to a period of heightened press coverage of AI Overview traffic complaints, following the same pattern as the August 2025 post (Confidence: Medium — based on observed pattern, not a confirmed announcement).
• Expect Ahrefs, SparkToro, or a comparable independent measurement firm to publish an updated dataset later in 2026 tracking whether the top-10-citation share (37.9% as of March 2026) continues to fall as fan-out retrieval matures (Confidence: Medium — extrapolation from an established publishing cadence, not a confirmed future study).
• Expect continued qualification and walk-back from individual Google spokespeople on specific mechanics (in the mold of Illyes' 2024 links retraction), even as the institutional "nothing has changed" message holds (Confidence: Low — pattern-based inference from one prior instance, not a prediction of a specific future statement).
• Expect schema and structured-data guidance from Google staff to remain deliberately non-committal ("it depends") rather than harden into explicit AI-retrieval recommendations, given Mueller's and Sullivan's stated reluctance to have publishers build for Search specifically (Confidence: Medium — consistent with two independent 2025–2026 statements).
• Expect at least one 2026–2027 conference (BrightonSEO, SMX, or a Search Central Live edition) to eventually feature a Google speaker and an independent-data researcher on the same panel discussing AI Overview traffic impact directly, given the topic's persistence across three straight conference seasons (Confidence: Low — no such session is currently confirmed; this is a prediction, not a report).

Practical takeaways

1. Treat "clicks are stable" and "clicks are down 34–47%" as compatible statements about different populations, not a resolved dispute — Google's figure is an aggregate across all query types, while Ahrefs, Pew, and Authoritas measured specifically AI-Overview-present queries. Build your own before/after CTR comparison from Search Console data segmented by query type before assuming either figure describes your site.
2. Do not restructure existing pages into disconnected "chunks" on the strength of Sullivan's framing alone — his guidance says not to bother, while Ahrefs' fan-out data suggests atomized, directly-answerable content is being cited more, not less. Resolve this tension with structural clarity rather than fragmentation: see Technical Foundations for how to make a single well-structured page extractable without breaking it apart.
3. Apply Mueller's "yes, no, and it depends" test to structured data feature-by-feature rather than site-wide — schema investment should be prioritized toward SERP features it demonstrably affects, which is covered under Entity & Knowledge Architecture.
4. Stop relying on any single vendor's zero-click or CTR figure as your baseline; track your own AI Overview citation presence and click impact directly, which is the core discipline behind LLM Visibility Monitoring.
5. Given that citation share is shifting toward sub-query fan-out results rather than primary top-10 rankings, prioritize coverage of the adjacent questions a user's query implies, not just the head term — this is the practical core of Generative Engine Optimization.
6. Design pages so a human landing after an AI Overview click — the 8-15% of visits Pew documents — gets something the summary didn't already give them; thin duplicate value is why session-abandonment rates run higher on AI-summary visits, a problem addressed under Search Experience Design.