AI Overview Optimization: What Actually Changes

The page ranking first in the United States for the phrase ai overview optimization is published by Google, and it tells you that AI Overview optimization is not a thing.

We read the live top ten for that term on 2026-09-10, along with the top ten for google ai overview seo, how to optimize for ai overviews and ai overview citations. That is 40 organic results. Google's own documentation holds position 1 on three of those four queries. A practitioner thread asking the question holds position 2 on the same three. Everything from position 3 down is a strategy listicle, and five of the eight are numbered: seven strategies, seven top strategies, ten tips.

Not one of the 40 reports a measured before-and-after against a control.

There is one more detail in that SERP worth pausing on. The result at position 2, on three of the four queries, is a practitioner thread asking how to optimize for AI Overviews. Google is ranking the question above nine pages that answer it.

That is the gap this post fills, and filling it turns out to be less about finding a better technique than about reading two things carefully: what Google actually writes down, and what the studies everybody quotes actually counted.

No, according to the only party with access to the systems. Google states that there are no additional requirements to appear in AI Overviews or AI Mode and no other special optimizations necessary, because the features run on the same core Search ranking and quality systems that decide ordinary results. Ordinary Search eligibility is the bar.

That is from Google's AI features and your website page, which carries a last-updated stamp of 2025-12-10. The dedicated AI optimization guide published later puts it more bluntly, and this is the sentence the entire industry has been arguing around.

Both quotations were matched against the text of the live pages fetched on 2026-09-10, not lifted from somebody else's summary of them. The same pass checked an invented control string against the same pages, which did not match, so we know the matcher was not simply returning true for everything put to it. Fourteen quoted statements matched; the control did not.

The mechanism behind the claim is stated too, and it is the part worth internalising.

Read the guide's own explanation and the reason for the "still SEO" position becomes structural rather than rhetorical. Google says its generative AI features are rooted in its core Search ranking and quality systems, and that they rely on retrieval-augmented generation against the Search index plus a technique it calls query fan-out, defined as a set of concurrent related queries the model generates to fetch additional relevant results. There is no second index. There is no separate ranking stack. The retrieval layer is the Search index, and the thing selecting from it is the Search ranking system.

If that is true, a separate optimization discipline could only exist in the gap between "ranked well" and "got cited". That gap is real, it is where all the interesting work is, and it is not where the listicles are pointing.

It is worth being precise about what that table is and is not. Every string in it was matched against the text of a page fetched over HTTP on 2026-09-10, stripped of script and style, and printed with the surrounding sentence so it could be read in context rather than as a fragment. That matters because quotations in this subject get passed hand to hand until they drift, and the most commonly circulated version of Google's position is a paraphrase of a paraphrase. Where Google's page exposes a last-updated stamp, it is reproduced, because two of these statements are in tension with each other and the only way to see which is current is to look at the dates.

This is the one genuinely novel mechanism in the stack, and it is the one most often turned into a product.

Google defines it plainly. Query fan-out is a set of concurrent, related queries generated by the model to request more information and fetch additional relevant search results to address the user's query. So a single question from a user becomes several retrievals behind the scenes, and the answer is assembled from what comes back across all of them. Google states that both AI Overviews and AI Mode may use the technique.

The obvious commercial move is to guess the fan-out queries and build a page for each. Two things stand in the way, and the first is simply that Google does not tell you what they were. There is no report, no export, and no field in Search Console that names the sub-queries run against your page. You would be optimizing against inference.

The second is that Google has named the tactic as a policy violation. Generating pages to cover anticipated fan-out queries, done primarily to manipulate rankings or generative AI responses, falls under the scaled content abuse spam policy by Google's own statement. It adds that this is an ineffective long-term strategy anyway, on the grounds that a high quantity of pages does not make a site higher quality or more relevant.

There is a legitimate version of the same instinct, and it predates all of this. If a subject genuinely has eight distinct questions underneath it, covering them properly is good work whether or not a model fans out. The difference between that and the spam version is whether the pages would exist if the fan-out mechanism did not, and that is a question you can answer honestly about your own site in about a minute.

Both AI pages route back to Google's general helpful-content guidance, which is the closest thing to an answer about content quality in this context.

That guidance is worth reading precisely because of what it does not say. Measured directly, the helpful-content page contains no mention of AI Overviews, AI Mode, or generative features at all, against a control count of nineteen occurrences of experience and E-E-A-T terminology in the same file. It is the general quality document, and both AI pages point at it rather than at anything AI-specific.

Its load-bearing sentence on E-E-A-T is more careful than the way the acronym usually gets used: while E-E-A-T itself is not a specific ranking factor, using a mix of factors that can identify content with good E-E-A-T is useful. That is a statement about proxies, not about a score, and anybody selling you an E-E-A-T optimization for AI Overviews is selling a metric Google says is not a metric.

The optimization guide contains a section that reads as a list of things people are currently wasting money on. Set against the advice on the pages ranking below it, the overlap is close to total.

Structured data. The guide is unambiguous.

It goes on to say structured data is still worth using as part of overall SEO because it helps with eligibility for rich results, which is a different benefit from the one being sold. The single structured-data instruction Google gives for AI features is that your markup should match the visible text on the page. That is an honesty requirement, not an optimization technique.

Chunking and answer-first rewriting. Google states there is no requirement to break content into tiny pieces for AI to better understand it, that there is no ideal page length, and that you do not need to write in a specific way just for generative AI search. Every one of those is contradicted by advice on the current SERP.

A page per fan-out query. This one carries a warning rather than a shrug. Generating pages to cover anticipated fan-out queries, done primarily to manipulate rankings or generative AI responses, violates Google's scaled content abuse spam policy by Google's own statement.

Inauthentic brand mentions. The guide says seeking inauthentic mentions across the web is not as helpful as it might seem.

None of this means the underlying activities are worthless. Clear writing, sensible structure and genuine mentions are all good for reasons that predate AI Overviews by two decades. What it means is that the AI Overview justification for them is invented, and a strategy sold on an invented mechanism cannot be evaluated, because there is no stated way for it to fail.

The register this produces is consistent and easy to recognise once you are looking for it.

Yes, exactly one, and it is administrative rather than editorial. A site must be included in Search generative AI features in Search Console to be eligible for display at all, which is the only documented gate in the whole chain that exists solely for these features and the only one no advice page on this SERP mentions.

That is from the guide stamped 2026-07-10. It is the only documented gate in the entire chain that exists solely for generative AI features, and it is absent from every advice page on the SERP.

The control it refers to lives in Search Console under property settings, named the Search generative AI control. Its help page carries a note that as of 2026-08-31 it was rolled out to all websites worldwide. It offers two options. The default includes your site's links and content in Search generative AI features. The other option is worded plainly: exclude my site's links and content from Search generative AI features.

Three things about it matter more than the toggle itself.

First, the scope is bounded and Google says so.

Second, the cost of excluding is stated without softening: you will not receive any traffic or impressions from these features. There is no partial mode where you keep the traffic and lose the summarisation.

Third, it is not a training control. Google's help page says so directly and points to Google-Extended for that purpose instead, which brings us to the most durable piece of wrong advice in this entire subject.

The week the control shipped, publishers started working out whether to use it, and the numbers they brought are the interesting part. One operator reported generative AI impressions tripling from around 5,000 to 15,000 daily with no traffic increase at all, which is exactly the shape that makes the toggle tempting and exactly the shape that makes it hard to evaluate.

The consensus in that thread ran against opting out, and the sharpest objection was a historical one: the same argument was had when featured snippets arrived, and hiding from the feature was not what worked then. The most useful suggestion was narrower, which is to exclude one informational cluster for thirty days and watch branded and direct traffic rather than organic click-through, since those are the channels an answer-without-a-click would show up in.

No, and this one has been wrong for a long time while being repeated constantly. Google-Extended is a training control governing whether crawled content may be used to train future Gemini models, and Google states plainly that it does not impact a site's inclusion in Google Search nor act as a ranking signal.

Google-Extended manages whether content Google crawls may be used for training future generations of Gemini models. The named products are Gemini Apps, the Vertex AI API for Gemini, and grounding with Google Search on Vertex AI. It is also not a crawler in the ordinary sense: it has no separate user-agent string, crawling is performed by existing Google agents, and the robots.txt token acts purely as a control.

So a site that added Google-Extended to robots.txt believing it had opted out of AI Overviews did not opt out of anything relevant. Its pages remained fully eligible for AI Overviews and AI Mode the entire time. Until 2026-08-31 there was no site-wide way to do what those teams thought they were doing, and now there is, in a completely different place.

This is the same shape of error we have written about before across our work on AI citations, where a control that governs training gets mistaken for a control that governs retrieval, because both are described as "AI" and the distinction is invisible unless you read the specification. The specifics differ across vendors. The confusion is identical, and it is worth reading our post on which OpenAI crawler actually decides citation eligibility alongside this one, because the two mistakes are structurally the same mistake.

If you want less of your content used, and you do not want the all-or-nothing exclusion, the robots meta tags are the documented lever. Google's robots meta tag specification, stamped 2026-03-24, states that nosnippet will also prevent the content from being used as a direct input for AI Overviews and AI Mode, and that max-snippet will also limit how much of the content may be used as a direct input.

That is genuinely useful and it is more granular than the Console toggle. It also has two holes that the same specification documents, and anyone relying on these controls should know about both.

The first: the max-snippet limit does not apply in cases where a publisher has separately granted permission for use of content. So the cap is a default rather than a hard ceiling, and a separate agreement overrides it.

The second is the more interesting one. Structured data remains usable for search results when declared within a data-nosnippet element. Google states that robots meta tag limitations do not affect the use of structured data, with the exception of article.description.

Put those together and you get a specific, avoidable failure. A publisher who wraps a passage in data-nosnippet to keep it out of AI features, while shipping that same passage as structured data in a JSON-LD block, has not withheld the content. They have withheld one copy of it and volunteered another. This is where Google's instruction that structured data should match the visible text stops being a style note and becomes load-bearing, because the two rules interact: markup that mirrors your page is exactly the markup that survives your snippet controls.

Rank. Consistently, across every published study, and with the strength of the relationship being the only thing in dispute.

This is where the field's evidence base lives, and it is also where it falls apart under inspection, because the numbers everybody quotes disagree with each other far more than anyone quoting them admits.

Six studies. One apparently simple question: what share of AI Overview citations come from pages ranking in the organic top 10. Reported answers of 16.7, 32, 37.9, 52, about 60, and 76.1 percent.

That is a 3.6x spread, and it is not because Google changed by 3.6x. It is because those six studies are not measuring the same quantity.

Work through the differences and each number becomes legible.

The Ahrefs March 2026 figure of 37.9 percent counts URLs appearing within the first ten blocks, and the study says explicitly that ads, featured snippets, People Also Ask boxes and video packs were each tracked as separate blocks. On a feature-heavy SERP, ten blocks may contain four organic results. Their blue-links-only control returns 37.10 percent, so the block definition is not the whole explanation, but it means the number is not answering the question most readers think it is answering.

The Originality.ai figure of 52 percent is computed after excluding citations outside the top 100, and the study states that 52 percent of all citations fell outside the top 100. So the headline describes a share of the remaining subset, not a share of all citations.

The seoClarity figure of 32 percent comes from a single day of collection, 2025-10-12, across 362,000 US desktop queries and 5.1 million citations. That is a large sample of a narrow moment.

The BrightEdge figure of 16.7 percent is the outlier at the bottom and the least auditable: a 16-month, nine-industry claim published with no query count, no citation count and no market stated.

And the Authoritas figure of about 60 percent is expressed inversely, as roughly 40 percent of top-10 organic pages not being cited, which is a different question again.

There is a further wrinkle worth knowing about if you go and read these sources yourself, because we hit it and it cost us a correction. The seoClarity page's own summary at the top advertises different figures than its body carries. The summary says roughly 50 percent top-10 overlap and roughly 49 percent of citations from the top 20. The body says 32 percent and 56 percent. The most likely explanation is a summary left over from an earlier revision that was never re-cut when the study was updated, which is an ordinary publishing accident rather than anything sinister.

It has a consequence, though. Two figures widely attributed to that publisher in 2026 write-ups, that 97 percent of AI Overview citations come from the top 20 and that position 1 appears in AI Overviews more than half the time, are not on the current page at all. The live page says 56 percent and 43 percent. Those look like earlier numbers still circulating as though they were current, which is the same failure mode as the click-through figure discussed below and the same fix applies: open the page, read the body, note the date.

What the one academic source says

There is one source in this area with no product attached to it and a released dataset, which makes it worth separating from the rest.

An April 2026 benchmark built on 11,500 real user queries, published as a public dataset, compared Google organic results, AI Overviews, and a generative model directly on the same queries. Three of its findings bear on this post.

It puts AI Overview prevalence at 51.5 percent of those queries, far above the keyword-database estimates, for the sampling-frame reason discussed above.

It reports that the sources consulted across different engines have very little overlap, at under 0.2 average Jaccard similarity. That is a formal statement of something practitioners describe anecdotally all the time, which is that being cited by one assistant tells you almost nothing about whether another will cite you.

And it reports that sites blocking Google's AI crawler are significantly less likely to be retrieved by AI Overviews despite the content being accessible. That is the one finding in this post that points at a concrete technical action with evidence behind it, and it lines up with the documented behaviour of the snippet controls rather than contradicting it.

We should be clear about our own limits here: we read the abstract and the reported findings, not the full paper, so the internals are unexamined by us and should be treated that way.

This deserves its own answer because it is the most repeated trend line in the subject and it is not a trend line.

Both figures are real and both come from the same publisher. The 76.1 percent is from July 2025. The 37.9 percent is from March 2026. Set side by side they look like a dramatic decline in how much ranking matters, and that framing has been reproduced widely.

Three things changed between the two studies.

The July 2025 study analysed only the three most visible citations in each AI Overview. The March 2026 study analysed all citations, 4 million URLs across 863,000 keyword SERPs, which works out to roughly 4.6 per SERP. Sampling only the most prominent citations biases toward the strongest-ranking sources, so the earlier figure was always going to be higher.

The publisher also states it improved its parsing methodology so it could see more of the citations. A better parser finding more citations will find more of the marginal ones, which again pushes the number down without anything at Google changing.

And the comparison set changed from ranking positions to SERP blocks.

Two instruments, one trend line. The honest summary is that we do not have a clean time series on this quantity at all, from anyone. What we have is two snapshots taken with different equipment.

Also stale, also superseded, and superseded by the very publisher that produced it.

There is a better source for the underlying question, and it is the one least likely to be selling you something. Pew Research Center ran a browsing-behaviour study on 900 United States adults who agreed to share their activity, covering 68,879 unique Google searches in March 2025, of which 12,593 produced an AI summary. It found that users who encountered an AI summary clicked a traditional search result on 8 percent of visits, against 15 percent for visits with no summary. It found that a link inside the summary itself was clicked on just 1 percent of visits to pages carrying one. And it found sessions ending on 26 percent of pages with a summary against 16 percent without.

That study has no product attached to it, measures humans rather than aggregated console data, and states its own limitation plainly: technical constraints meant only Google searches could be analysed, and it covers a single month.

It is also not comparable with the agency CTR studies, and the two are routinely stacked as though they corroborate each other. Pew measures per-visit click behaviour across all query types on a probability panel. The agency studies measure impression-weighted click-through on informational client keywords pulled from Search Console. Different populations, different denominators, different query mixes. Both can be true and neither confirms the other.

There is no single answer to this, and the range is wide enough that quoting one number is close to meaningless.

A keyword-database study covering more than 10 million keywords put prevalence at 15.69 percent in November 2025. An academic benchmark of 11,500 real user queries, released as a public dataset, put it at 51.5 percent in April 2026.

The gap is a sampling-frame artifact, not a discovery. Keyword-tool databases are built around commercial head terms, which trigger AI Overviews far less often. A representative sample of what people actually type skews toward questions, and questions trigger the feature heavily.

The Semrush series carries one further detail that most 2026 write-ups drop: the direction is not monotonic. Prevalence rose from 6.49 percent in January 2025 to a peak of 24.61 percent in July 2025, then fell back to 15.69 percent by November. Anyone describing a smooth upward march is describing a curve they truncated.

Google's own framing is consistent with a variable rate. Its documentation states AI Overviews are only shown when its systems determine the feature is additive to classic Search, and adds that they often do not trigger. That sentence is the answer to one of the most common searches in this cluster, which is some version of "why is my AI Overview not showing". It is not a bug and it is not something you failed to optimize for.

Google says it is not required. The interesting question is whether it helps anyway, and here there is exactly one study designed to answer it rather than assert it.

The design is careful in a way that is rare in this field. 730 AI citations across 75 commercial queries and 1,006 unique pages, comparing ChatGPT and Gemini citations against Google's top-10 organic as a control set, analysed with generalised estimating equations and query-clustered standard errors, with domain authority carried as a covariate.

The naive pooled model showed schema negatively associated with citation. The author identified that as an artifact rather than publishing it as a finding: Google's own ranking enriches top-10 results for schema-bearing pages, which inflates schema prevalence in the control set. His within-Google diagnostic makes the point cleanly, finding schema prevalence among cited and non-cited Google pages statistically indistinguishable at 43.1 percent against 44.8 percent.

The corrected model returns a null for schema presence, at an odds ratio of 0.678 with a p value of 0.296, and nulls for entity richness and schema-to-query alignment as well. The dominant predictor is organic rank, with citation odds falling roughly 24 percent per position, and cited-in-43-percent-of-queries at position 1 falling to 5 percent by position 7.

One narrow positive result survives: Product and Review schema carrying concrete populated attributes such as pricing and aggregate rating was cited at 61.7 percent against 41.6 percent for generic Article, Organization and BreadcrumbList markup, most pronounced on lower-authority domains. That is a specific, plausible, testable claim about attribute-rich commercial markup, and it is a much smaller claim than "add schema to win AI Overviews".

Three caveats belong with this study every time it is cited, and we would rather state them than have somebody else find them.

It measures ChatGPT and Gemini, not Google AI Overviews. It is a preprint and has not been peer reviewed. And its author runs an AI search optimization agency, which he declares in the paper, along with the observation that the null result contradicts his own commercial interests. That last point cuts in favour of the finding rather than against it, but it should be visible either way.

There is no equivalent study for AI Overviews specifically. That absence is worth sitting with: the single most recommended technique in this entire field has never been isolated as a variable against the surface it is recommended for.

It is not for want of asking. In August 2026 somebody put the question to a technical SEO community in exactly those terms, asking whether anyone had added or fixed schema on a page and watched something change in AI citations, separate from everything else they were doing at the same time.

Thirty-three replies, twenty upvotes, an upvote ratio of 0.96, and not one before-and-after. The answers ranged from "table stakes" to "frankly speaking, it does nothing", with one reply quoting Google's structured-data line verbatim. The closest anyone came to a positive claim was a reference to studies showing a 20 to 30 percent increase in citations for pages using schema, immediately followed by the same commenter conceding they had not seen any specific test where someone added it and tracked the result.

We went looking, using the same standard we would want applied to us: a stated baseline, an intervention, and a way for the result to come out negative.

Across twenty-one practitioner threads from the last three months, spanning technical SEO, general SEO, digital marketing and several answer-engine communities, we found no controlled before-and-after experiment on AI Overview citations. Not a small number. None.

What we found instead falls into three groups, and the pattern is consistent enough to be worth naming.

The positive claims are uncontrolled, and several were withdrawn in their own threads. A widely upvoted post claiming a technique made an assistant the site's top referrer turned out, when asked directly for method, to have been built by copying whatever pattern the already-cited pages had. That is selecting on the outcome, and the community said so. The same author later downgraded one of the claims to a cheap bet rather than a proven lever. Another claim that adding markdown files increased assistant signups came with no baseline, no dates and no site named when asked.

The negative findings are the well-evidenced ones. The strongest measurement in the set found that when an engine did not run a live search, it produced zero citations across all 406 such answers, while still naming brands at roughly the same rate as when it did search. Being named and being cited turn out to be independent. Another found that a brand's own product pages were near worthless for citation compared with roundup and article content on the same domain. A third found that the pages that got cited were, on average, slightly worse written than the ones that did not.

The methodology warnings are the most useful output of the whole corpus, and two of them invalidate a lot of published work retroactively. One practitioner running an automated tracker found a page cited for one query and completely absent for the same query with one extra word in it. Different source list, same page. Another described generating a fresh question set on each run early on and producing what they called beautiful, entirely fake regressions, which were then explained with confident stories about algorithm updates. A rotating prompt set does not measure a trend, it manufactures one.

Set against that, the thread that keeps coming back is the one from a site that does everything right and is still never pulled into the panel.

The operator outranks the pages being cited, has higher authority than them, and has the full technical checklist in place. Forty-seven comments, and the replies split between the standard three-step advice and people saying plainly that AI Overviews do not follow organic rankings and that there is no proven way to optimize specifically for them. When one commenter asked the author of the top-voted standard advice for the research behind it, no answer was given.

That is the honest state of the field. It is not that the advice is malicious. It is that almost none of it has been tested against the surface it names, the people giving it mostly know that, and the few who ran real measurements mostly found out what does not work rather than what does.

The findings that survive scrutiny are all about ceilings

Read the practitioner measurements together and a shape emerges that nobody is selling, because it does not convert.

The strongest single result in that corpus came from running 50 buyer questions across 10 national brands and 4 engines, 2,000 live answers, reading each engine's own disclosure of whether it had run a live web search. Where a live search happened, the brand was cited about 17 percent of the time. Where no search happened, citations were zero, across all 406 such answers. Not low. Zero. Mentions were almost unaffected, at 69 percent against 68 percent, which means being named by an assistant and being cited by it are close to independent events.

The counterintuitive part is which questions did not trigger retrieval. Broad category questions searched about half the time. The buyer-shaped ones, brand A against brand B, or a situational recommendation, mostly did not. So the queries closest to a purchase were the ones most likely to be answered from memory, where no amount of on-page work can reach.

That result is about a different surface from AI Overviews, and we are not going to pretend otherwise, since AI Overviews are grounded in Google's index rather than answering from weights in the same way. It is included because the structure of the finding generalises: if retrieval did not happen, page-level optimization is inert, and almost nothing in the popular advice distinguishes the case where you lost the citation from the case where there was never a retrieval to win.

A second audit, of 158 published articles tracked across several engines including AI Overviews, found something similarly awkward for the content-quality pitch. Articles that got cited scored slightly lower on writing quality than articles that did not, 60.4 against 62.5. The variable that moved was where the piece was published, not how well it was written: the same author's material was cited around half the time on an established industry publication and not at all on their own company blog. Age mattered too, in the opposite direction to the freshness advice, with older articles cited far more often than fresh ones.

That study was published the same day we ran this research, has had no scrutiny from anyone, and is vendor-run on the vendor's own corpus. Treat it as unrefuted rather than as validated. We include it because the direction it points, that placement dominates polish, is corroborated by the third-party-source finding below rather than standing alone.

The mentions correlation, and why it is not a lever yet

The one correlation consistently reported as stronger than links is brand mentions.

An analysis across roughly 75,000 brands reports correlations against AI Overview visibility of 0.664 for branded web mentions, 0.392 for branded search volume, 0.326 for domain rating and 0.218 for backlinks. The distribution underneath is the interesting part and it is not linear: the top quartile of brands by mention volume averaged 169 AI Overview mentions, the next quartile 14, and the bottom half between zero and three.

Two cautions before anybody reallocates a budget on that. It is correlational by construction, and the people publishing it say so. And a threshold-shaped distribution is exactly what you would expect if a third variable, such as simply being a large well-known company, drives both mentions and citations. That does not make the finding useless. It makes it a description of who gets cited rather than an instruction for how to become them.

The practical version of this that does hold up is narrower and comes from a case rather than a correlation. A team spent three months publishing pricing pages on their own blog to correct what an assistant was saying about them, and nothing moved. The actual sources were two comparison articles from 2024 on sites they had never heard of. Getting one corrected, and a newer piece placed on a site the models already cited, changed the answers about five weeks later.

That is a single case with no holdout, and the three months of publishing is arguably its own null arm. But the diagnostic underneath it is free and takes ten minutes: before writing anything to fix an AI visibility problem, find out what the engine is actually reading, because it is usually not you. Ask the question with search on and with search off in a fresh session. If the wrong answer only appears with search off, it is in the weights and publishing will not touch it. If it appears in both, something live and wrong is being retrieved, and that you can go and fix.

No, and this one is settled by documentation rather than by evidence.

An AI Overview holds one position in the results. Every link inside it is assigned that same position. There is no first slot to win, no ordering to climb, and no rank-within-the-panel to report.

Any tool showing you a rank inside an AI Overview is showing you a number Google does not define, derived from the visual order of the panel at the moment of a scrape. Given that the same panel changes roughly every two days, that number would be unstable even if it meant something.

This is the part of the research that is best established and least discussed, and the reason for the imbalance is not subtle: it does not lead to a product recommendation.

The larger study analysed more than 43,000 keywords, each carrying at least 16 recorded AI Overviews, over a month. It reports a 70 percent chance of an AI Overview changing between one observation and the next, an average persistence of 2.15 days, and only 54.5 percent of cited URLs overlapping between consecutive responses, meaning roughly 45 percent of the sources credited are new each time. Entity overlap runs at 54 percent. Semantic similarity of the answers themselves runs at 0.95.

The authors add a caveat that pushes the finding further rather than softening it: because their checks were not daily, the real change rate is likely higher than measured.

An independent study using different methods across 11,203 keywords and roughly five months found about 70 percent of the pages ranking in AI Overviews changing over two to three months, with AI Overview volatility running consistently above organic volatility.

Two publishers, different instruments, different windows, converging figure. That is what corroboration looks like, and this is the only place in the subject where we have it.

The implication is the interesting part. The answer Google gives is stable at 0.95 semantic similarity, while the set of sources credited for that answer turns over by nearly half between observations. Google is not changing its mind. It is re-picking who gets credit for the same conclusion, repeatedly, on a two-day cycle.

That has a direct consequence for anyone measuring their own visibility. A single observation of your citation state is close to worthless. If you check on Monday and you are cited, and check on Thursday and you are not, you have not learned that something broke. You have sampled a process with a 70 percent change rate twice.

This is the question the entire SERP skips, and it is a measurement question rather than an SEO question.

Start with what tracking is actually available. Google states that clicking a link to an external page in an AI Overview counts as a click, and that these are reported in the Search Console performance report within the Web search type rather than broken out separately. The dedicated generative AI insights surface, announced in June 2026 and rolled out to all sites, reports impressions.

So the honest position on attribution is that it is coarser than anybody would like, and that a large share of the impact of AI Overviews on your traffic is not separable from ordinary Search in the console you already have.

Which pushes the measurement burden onto direct observation of the panel itself, and that is where the volatility finding above becomes the binding constraint rather than a curiosity.

Practitioners who track this seriously have arrived at the same conclusion independently, and the thread below is the clearest statement of it we found.

Two findings in it are worth carrying away. The first is that one check is not a reading: the recommendation that recurs is to run a fixed prompt set in fresh sessions, log domain appearance alongside the cited URL, and compare share over time rather than treating any single answer as a rank. The second is sharper, and it is a first-hand observation rather than advice. One tracker reported a page cited for the exact query supabase error 42501 row level security fix and completely absent for supabase error 42501 row level security policy fix. One extra word, a different source list, the same page. If your prompt set is built from paraphrases of your own product description, it will hand you a reassuring number that means nothing.

Here is our own position, published with its numbers, including the number that is unflattering.

We ran a step-zero measurement: 12 money queries, 5 identical repeats each, one model, one day. All 60 calls succeeded. Seven of the 12 queries flipped outcome between repeats that differed in no way at all. Not between Monday and Thursday. Between identical calls.

We then ran a permutation test, 20,000 random splits of the same query set with no intervention applied whatsoever, to see what a difference looks like when nothing has happened. The null difference centred on zero, mean plus or minus 0.0016, with a standard deviation of 0.215.

From those two runs the minimum detectable lift for that design comes out at 51.1 percentage points. Roughly four times underpowered for any effect anyone would plausibly claim. A design of 40 queries by 40 samples reaches 9.9 points.

We publish that because it is the number that decides whether any of the rest of this is knowable, and because we would rather be the people who said the instrument was not good enough than the people who reported a lift it could not have seen.

It also explains a decision that might otherwise look like modesty. We have no causal lift figure anywhere on this site. Not a small one, not a hedged one. A design whose smallest resolvable effect is 51.1 percentage points cannot honestly report a 12 point improvement, because 12 points is indistinguishable from the noise the same design produces when nothing at all has happened. Publishing one anyway is how most of the case studies in this field get written, and the arithmetic that would catch it is never shown.

The other reason to publish a detection floor is that it converts an argument into a calculation. Someone can disagree with our interpretation of Google's documentation. Nobody can disagree that a 12 by 5 design cannot see a 10 point effect, because that follows from the spread of the null distribution, and we published the null distribution. The whole point of putting the unflattering number in public is to make the next claim checkable by somebody who does not trust us.

It also sets the standard we would apply to anybody else's claim, including every study quoted in this post. If someone tells you a technique moved their AI Overview citations, the first question is not what the technique was. It is how many queries, how many repeats, and what the spread looked like when they changed nothing. Our own answer to that third question was a standard deviation of 0.215, which is wide, and we had to run 20,000 splits to find out.

For a fuller treatment of how sample size interacts with this, our post on prompt tracking sample size works through why the intuitive number of repeats is usually far too small, and the measuring answer engine optimization lift page describes the design we use when the question is causal rather than descriptive.

Mostly no, and the places it does change are worth knowing because they affect measurement rather than content.

Google's documentation treats AI Overviews and AI Mode together for nearly everything discussed above. The same eligibility applies, the same Search Console control governs both, the same snippet controls limit both, and both may use query fan-out. If you were hoping for a separate lever here, there is not one.

The measurement side is different in one specific way. Google states that a follow-up question asked inside AI Mode is essentially a new query. So a conversation that starts with one question and continues through four refinements is five queries, and your visibility across that conversation is not one state but five. Any tool reporting a single AI Mode visibility number for a topic is flattening something that has real structure in it.

There is also a reporting boundary worth being precise about, because it is easy to overclaim in both directions. Clicks on external links inside an AI Overview count as ordinary clicks and land in the standard performance report under the Web search type. The dedicated generative AI insights surface reports impressions. Measured directly, that help page mentions impressions ten times and contains no mention of clicks, click-through rate or position at all. The honest reading of that is not that clicks from these features do not exist, because the performance report documentation says plainly that they do. It is that the dedicated report is an impressions surface, and anyone building a click-based AI Overview dashboard on top of it is building on a metric it does not carry.

Shorter than any listicle on this SERP, which is the point. Everything below is documented rather than asserted, and where something is merely correlational it says so.

Check the Search Console generative AI control first. It is the one genuinely AI-specific eligibility gate that exists. It defaults to include, and child properties inherit from the nearest configured parent, so an inherited setting on a parent property can exclude a subdomain nobody has looked at. If it says exclude, nothing downstream matters.

Confirm ordinary Search eligibility. Google's stated requirement is that a page must be indexed and eligible to be shown in Google Search with a snippet. That is the bar. Not a special one, but a real one.

Audit your snippet controls for accidents. A nosnippet rule prevents content being used as a direct input for AI features. If a template applies one broadly, you may be withholding the passage you most want quoted. Check the interaction with structured data described above, because the two controls do not compose the way people assume.

Make structured data match the visible text. This is Google's only structured-data instruction for AI features. It is also the rule that governs whether markup survives your snippet controls.

Rank better, with the usual caveat. Every published study agrees rank correlates with citation, and the one controlled experiment finds rank the dominant predictor while its schema variables come back null. Correlation is the right word here. None of these studies randomised anything.

Do not build pages against fan-out queries. Google names this as a scaled content abuse violation when done to manipulate generative AI responses. It also does not expose which fan-out queries it ran, so you would be optimizing against a guess.

Establish your noise band before you test anything. Given a roughly 70 percent inter-observation change rate, a before-and-after with one reading on each side cannot distinguish an effect from the ordinary churn. This is the step that is almost always skipped, and skipping it is what makes the resulting case study unfalsifiable.

Treat every statistic in this field as an instrument reading. Ask for the sample size, the collection window, and whether the publisher sells the tool that produced it. Three of the six overlap studies quoted above are published by companies selling AI visibility products, and one of those six publishes no denominator at all.

Find out what the engine is actually reading before you write anything. The ten-minute search-on and search-off diagnostic above will tell you whether a wrong or missing answer is coming from live retrieval, which you can act on, or from training data, which you largely cannot. Skipping this step is how a quarter disappears into publishing pages that were never in the retrieval set.

Two of those items are worth expanding for anyone whose buyers are technical, because the question shape differs sharply by audience. We keep a running note on the questions developer tool buyers ask AI, and the pattern there is that the highest-intent queries are also the most specific, which is exactly the class most likely to be answered without a fresh retrieval. The same holds for AI infrastructure buyers. If that is your market, the retrieval question is not a technicality, it is the whole game.

If you want the fuller treatment rather than the summary above, everything we have published on measurement works through the instrument side, and the rest of our original research on AI citations sits alongside it.

We have no customers. There is no client engagement behind this, no case study, and no aggregate of anybody's data but our own two published runs. If this post contained a chart showing citation lift from a technique, that chart would be invented, so there is not one.

We also cannot tell you that any of the documented steps above will get you cited. Google documents eligibility gates, not outcomes, and the difference between the two is the whole subject. The gates are necessary. Nothing in the public record establishes that anything is sufficient.

And we cannot give you a causal estimate of our own, because our instrument's detection floor is 51.1 percentage points and we said so above rather than burying it. When we have run a design large enough to resolve something smaller, we will publish that with its numbers too, including if the answer is that the effect is not there.

What we can tell you is that a field whose most-quoted overlap statistic varies by a factor of 3.6 across six studies, whose headline CTR figure has been superseded by its own publisher, and whose most recommended technique has never been isolated as a variable against the surface it is recommended for, is a field where the confident answers are the ones to distrust first.

That includes ours. The difference we are aiming for is that ours come with the number that says how wrong they could be.

One last note on how to read this post in six months, because everything above has a shelf life and some of it is short. The documentation quotations carry Google's own update stamps, and two of the pages we quote already disagree with each other because one was revised and the other was not. The Search Console control is two weeks old at the time of writing. The click-through picture reversed once already and could reverse again. The volatility finding is the most likely to hold, because it has been reproduced by two publishers using different instruments, and the schema null is the most likely to be overturned, because it rests on one preprint measuring an adjacent surface.

If you are reading this after those things have moved, the part worth keeping is not any individual figure. It is the habit of asking, of every number in this field, what was counted, against what denominator, on what date, and by somebody selling what.

If you want to see how we approach this for a specific product surface, our work with developer tools and APIs starts from the same place: establish what moves on its own, then ask what is left.