Coverage settled near a quarter of queries after Google recalibrated in late 2025 — though figures swing widely by method.
Ask how often AI Overviews appear and you will get answers ranging from a small minority of queries to a substantial majority, all from credible sources measuring in good faith. They are not contradicting each other so much as measuring different things. This study puts coverage at roughly a quarter of queries following Google’s late-2025 recalibration — and the more valuable contribution is the demonstration of why any single headline prevalence number is close to meaningless without its methodology attached.
The research tracks how frequently AI Overviews appear across a very large query sample, observed after Google adjusted the feature’s coverage in late 2025. The scale is what gives the figure stability — prevalence measured across a small keyword set swings wildly depending on which keywords happen to be in it.
The reported figure settles near a quarter of queries. That is a materially different number from the peaks reported during the feature’s aggressive expansion, and it reflects a deliberate pullback rather than a measurement change. Coverage went up sharply, then came down, and this study measures where it landed rather than where it peaked.
The variation between published prevalence figures is not sloppiness. It is a direct consequence of the fact that Overviews appear on some kinds of query far more than others, which means the answer depends almost entirely on which queries you count. A sample weighted toward informational, question-shaped queries produces a high figure. One weighted toward navigational, branded, or transactional queries produces a low one.
Add further variation from the geography sampled, the device, the language, the time window, and whether the tracker records a feature that appears intermittently — and two rigorous studies can differ by a factor of several while both being correct about the sample they measured. The lesson is that prevalence is not a property of search; it is a property of a query set.
The more interesting fact behind the figure is that coverage moved — expanded aggressively, then was pulled back. That tells you something about the feature that a static prevalence number does not: Google is actively tuning where it considers a generated summary appropriate, which means coverage is a policy decision rather than a fixed property.
For planning, that matters more than the current level. A number that reflects a deliberate, revisable choice can change again, in either direction, on a timescale shorter than most content strategies. Building a plan that assumes today’s coverage is stable is building on something the study itself demonstrates is not stable.
Published figures range widely because Overviews appear far more on some query types than others. The only prevalence number that means anything for your planning is the one measured on your own queries.
Whatever the correct market-wide figure, it is close to useless as a planning input, because no business competes across a representative sample of all queries. You compete on a specific set determined by your category, your customers, and your commercial priorities — and the prevalence on that set could plausibly be double the market figure or a fraction of it.
A brand in a definitional, explanatory category will see Overviews on most of its important queries. A brand whose demand is overwhelmingly branded and navigational will barely encounter them. Both would be badly misled by planning off a market average, in opposite directions. The average describes a web nobody actually operates in.
What it cannot tell you: prevalence on your query set, which is what actually matters. Figures across published trackers differ substantially because coverage is highly sensitive to query composition, geography, and observation window — a spread that reflects genuine variation rather than measurement error.
The productive response to this study is to run its method on your own portfolio. Take your priority queries, check which currently return an Overview, and record it — producing your own prevalence figure, segmented by content type and by commercial stage. That number is directly actionable in a way no published average can be.
The segmentation is where the value sits. Prevalence at the top of your funnel will typically differ sharply from prevalence at the decision stage, and knowing which parts of your portfolio are exposed tells you where citation work is urgent and where traditional ranking still carries the visibility. A single site-wide figure would obscure exactly that distinction.
Since the study establishes that coverage is actively tuned, a prevalence measurement is a reading rather than a constant. Queries that show no Overview today may show one next quarter, and the reverse also occurs. A measurement taken once and treated as fixed will quietly diverge from reality.
The practical routine is to re-check prevalence on your tracked query set periodically, alongside your citation tracking, and to treat a change in coverage as a distinct event from a change in your own performance. Without that, an expansion of coverage into your category looks like a sudden inexplicable decline in click-through, and the diagnosis goes astray.
Knowing an Overview appears is step one. DUNkē tracks whether you are cited inside it across eight AI engines — per prompt, against competitors — so coverage and presence are measured together.
Prevalence is only meaningful when multiplied by impact. A high coverage rate on queries that matter little to your business is a smaller problem than moderate coverage on your highest-intent terms. The exposure calculation that matters is coverage weighted by commercial value, not coverage alone.
Doing that arithmetic properly frequently reorders priorities. A category with modest prevalence but where every affected query sits at the decision stage may warrant more urgent attention than one where Overviews appear constantly on early-stage informational questions that were never going to convert directly. Prevalence tells you where the feature is; value tells you where it hurts.
This figure describes a particular sample, at a particular time, in the aftermath of a deliberate change by Google. It is not a constant of nature, and the study’s own framing acknowledges that figures swing widely by method.
Use it to calibrate expectations about the general environment. Do not use it as an input to a forecast — for that, measure your own queries, segment by commercial stage, and re-measure on a cadence, because coverage has already been shown to move in both directions.
Since prevalence depends almost entirely on query composition, it is worth being specific about which queries attract Overviews. The pattern is consistent: questions with a stable, summarisable, generally-agreed answer draw them most reliably — definitional questions, how-things-work explanations, straightforward comparisons, and procedural queries.
Coverage thins markedly on queries where a summary adds little or risks being wrong: navigational queries where the user wants a specific destination, transactional queries, highly local queries, queries about rapidly changing situations, and queries where the honest answer is genuinely contested. Mapping your own portfolio against that distinction predicts your exposure surprisingly well before you measure anything.
A market-wide figure of roughly a quarter invites a reading that three quarters of search is unaffected, which is true only in a sense that does not help anyone. The affected quarter is not randomly distributed — it is concentrated in exactly the informational, explanatory queries that content marketing has spent two decades learning to capture.
So a business whose organic strategy is built on educational content may find that most of its important queries fall inside that quarter, while a business competing mainly on branded and transactional terms may find almost none do. The market average is arithmetically correct and strategically useless for both. Only the distribution across your own portfolio tells you anything actionable.
The recalibration this study captures is instructive beyond its specific numbers. Coverage was expanded aggressively, then pulled back — a pattern consistent with a platform testing the boundaries of where a generated answer serves users well and retreating where it does not. That is ordinary product development, and it should be expected to continue.
The planning consequence is to treat coverage as a variable rather than a constant, and to avoid strategies that depend on it staying where it is. A content programme predicated on Overviews never reaching a particular query category is fragile; one built to compete for both the ranking and the citation is robust to movement in either direction. Given the demonstrated volatility, robustness is worth more than optimisation against the current state.
A complication that partly explains disagreement between trackers is that Overviews do not appear reliably for a given query. The same search can return one on one occasion and not on another, which means prevalence is not a binary property of a query but a probability — and how a tracker handles that materially affects its reported figure.
A tracker recording a single observation per query produces a different result from one sampling repeatedly and recording any appearance, or one requiring appearance in a majority of observations. None is wrong; they are answering slightly different questions. For your own measurement, sampling each query more than once and recording the frequency is more informative than a single check, particularly for queries near the boundary.
Coverage has consistently varied by market, with rollout and calibration differing across regions and languages. A figure measured predominantly in one market can substantially misrepresent conditions in another, which matters for anyone operating internationally or outside the markets where most published research is conducted.
For a business operating in a market that published trackers under-sample, the only reliable option is measuring locally. This is one of the clearer cases where imported industry figures are actively misleading rather than merely imprecise, and where a modest amount of first-party measurement produces far better information than any amount of reading.
The practical build is straightforward. Take your tracked prompt and query set — the same one used for citation measurement — and record for each whether an Overview appears, sampling more than once where results are inconsistent. Segment the output by content type, funnel stage, and commercial value, and re-run it on a regular cadence.
What this produces is a coverage map of your own portfolio that updates over time, which serves several purposes at once: it identifies where citation work is urgent, it explains click-through changes that would otherwise look inexplicable, and it gives early warning when coverage expands into a category that was previously unaffected. It is a small addition to an existing tracking programme and one of the higher-value ones.
Because coverage moves and its movement directly affects traffic, a change in prevalence on your queries is worth alerting on rather than discovering in a quarterly review. An expansion into a previously unaffected category produces a click-through decline with no corresponding change in your rankings or content — the classic pattern that gets misdiagnosed as a site problem.
Having the coverage measurement running continuously turns that from a mystery into a labelled event. It also distinguishes an environmental change from a performance change, which is one of the more useful distinctions in search reporting and one that is nearly impossible to make without prevalence data of your own.
A final caution: coverage says nothing about whether you are cited. A query showing an Overview is a query where citation is possible and where clicks are depressed; whether you are inside that summary is an entirely separate measurement. Prevalence describes the environment, citation describes your position within it.
Conflating the two produces a common analytical error — treating high coverage as a problem in itself, when high coverage on queries where you are consistently cited is a substantially better position than moderate coverage where you never appear. The two measures belong together, and the combination is what describes your actual exposure. Either alone is half a picture.
Coverage data changes content planning in a specific way: it tells you which planned pieces are being commissioned into a contested environment. A guide targeting a query where an Overview reliably appears faces a different economic proposition from one targeting a query where none does, and knowing which before commissioning is genuinely useful.
This does not mean avoiding covered queries — the content that gets cited in those summaries is frequently exactly this kind of material, and being absent guarantees absence from the answer. It means setting expectations correctly at the point of commissioning, so the piece is judged on citation presence rather than session volume, and so the format and structure are built for extraction from the outset.
Having warned against the industry average, it is worth noting that a single site-wide prevalence figure of your own carries the same defect at smaller scale. If a quarter of your queries show Overviews, that could mean uniform light exposure or it could mean total exposure across your entire informational library and none elsewhere.
Those are completely different situations demanding different responses, and the aggregate cannot distinguish them. The segmentation that matters is by content type and by funnel stage, because those are the axes along which both coverage and commercial consequence vary. A prevalence report that produces one number for the site has reproduced the error it was built to avoid.
There is a further use for prevalence tracking that is easy to miss: coverage expansion into a category is a signal about how the platform assesses that category. Overviews appear where a generated answer is judged safe and useful, so their arrival on a set of queries indicates the platform now considers those questions summarisable.
For a business, that carries information beyond the click impact. It suggests the questions concerned have become sufficiently well-documented across the web that a confident synthesis is possible — which is itself a statement about how commoditised that information has become. Categories where this is happening are ones where differentiation has to come from something other than explaining the basics well.
Prevalence tracking surfaces an uncomfortable category: queries where an Overview appears reliably, you are not cited, and the competitive position makes citation unlikely in any reasonable timeframe. The disciplined response is to stop investing there rather than to keep producing content that competes for a diminished click on a query you cannot win.
Reallocating that capacity to queries where you can be cited, or to parts of the portfolio where no Overview appears, produces more return from the same effort. This is one of the clearer cases where good measurement enables a decision to stop, which is usually harder and more valuable than a decision to start. A prevalence map that never leads to anything being abandoned is probably not being read honestly.
This study answers how often, and the click-impact research answers how much. Neither is useful alone: severe impact on rare queries and mild impact on ubiquitous ones can produce identical aggregate outcomes, and only the combination distinguishes them.
The complete exposure picture requires both measured on your own portfolio — the share of your commercially weighted queries showing an Overview, multiplied by the click-through penalty you actually observe on them. That single derived figure is more useful than either input and is the number worth putting in front of stakeholders, because it expresses the problem in terms of business consequence rather than platform behaviour.
Of the research collected in this hub, prevalence is the finding most likely to be out of date soonest, because it measures a setting rather than a behaviour. Click-through effects reflect how people respond to an interface and change slowly; coverage reflects a decision that can be revised at any time.
That is not a reason to discount it — knowing where coverage settled after a deliberate recalibration is genuinely informative about how the platform is thinking. It is a reason to treat the number as perishable and the method as durable. The lasting value of this study is the demonstration that prevalence is query-set-dependent and actively managed, which will remain true whatever the current figure happens to be.
One underused application of prevalence tracking is timing. Coverage tends to arrive in a category rather than on isolated queries, so watching adjacent query groups gives some warning about where it is heading next — and content published into a category before coverage arrives has time to establish authority and citation presence before the click economics change.
That is a meaningful advantage. A page that is already the established, well-cited source when an Overview appears is positioned to be one of the sources the summary draws on; a page published afterwards is competing for citation from a standing start in a category that now delivers fewer clicks to fund the effort. Coverage tracking used prospectively rather than only retrospectively is one of the few ways to get ahead of the change rather than responding to it.
The lasting value of this research is not the quarter-of-queries figure, which is a reading taken at a moment in a landscape the study itself shows is actively managed. It is the demonstration that prevalence is a property of the query set being measured, which reframes every published coverage statistic as an answer to a question about a particular sample.
Once that is understood, the correct response to any prevalence claim becomes automatic: ask which queries, in which market, over what window, sampled how often. Applied consistently, that habit turns a confusing body of contradictory industry figures into a set of readings that can be compared sensibly. And it points at the only measurement that actually informs your decisions, which is the one you take on your own portfolio and repeat.
AI Overview coverage settled near a quarter of queries following Google’s late-2025 recalibration, measured across a very large sample. But the more useful contribution of this research is the demonstration that prevalence figures vary enormously by method — because Overviews appear far more on some query types than others, making any headline number a property of the query set that produced it rather than of search itself.
The practical consequence is to stop importing industry averages into planning. Measure prevalence on your own priority queries, segment it by content type and commercial stage so you can see which parts of your portfolio are exposed, weight it by commercial value rather than treating all coverage as equally costly, and re-measure periodically — because the study establishes that coverage is a revisable policy decision, not a stable feature of the landscape.
“Every credible tracker reports a different prevalence figure, and they are all right. Coverage is a property of the queries you counted — which is why the only number worth planning with is your own.” The Age’X Research Team
Get a free GEO audit — the same analysis behind every article here.