Position-1 CTR drops about 58% when an AIO is present; YouTube and branded mentions correlate most strongly with AI brand visibility.
Ranking first has always been the objective, and this study quantifies what it is now worth when an AI Overview sits above it: roughly forty per cent of the click-through it used to earn. That is the headline. The more useful part is buried further in — an analysis of what correlates with appearing in AI answers at all, which points somewhere most brands are not looking: video presence and being mentioned by name across the web, rather than anything on the page itself.
The research pairs a large keyword panel with click data to compare position-one click-through on queries where an AI Overview appears against queries where it does not. Working from search performance data rather than modelled estimates gives it a solid foundation, and the scale of the keyword set makes the central finding reasonably robust to the vagaries of any particular vertical.
Alongside the click analysis, it examines which brand-level signals correlate with visibility inside AI answers. These are two quite different investigations sharing one dataset, and they deserve to be read separately — the first is a measurement of harm, the second is a hint about remedy, and conflating them produces confused conclusions.
When an AI Overview is present, the click-through rate of the first organic result falls by well over half. The position is unchanged; its value is not. This is the clearest available statement of a shift that rank tracking cannot express, because a rank tracker reports where you sit and says nothing about what sitting there now earns.
The finding is intuitive once stated. A summary that answers the query occupies the top of the page and satisfies a large share of the people who would otherwise have clicked the first link. What remains is the residual: users who want more detail, want to verify, want to transact, or simply prefer choosing their own source. That residual is substantial but it is not what position one used to deliver.
It is worth noting that the top position has the most to lose in absolute terms, because it captured the largest share of clicks to begin with. A result that received a modest share of a query’s traffic loses less in absolute terms when the query is partly answered above it, simply because it had less to lose.
This has an uncomfortable implication for the highest performers: the pages that benefited most from strong rankings are the pages most exposed to answer displacement. Brands with dominant organic positions in categories where Overviews are common are precisely the ones whose traffic models are most disrupted — which is the opposite of the usual assumption that market leaders are the most insulated.
The second half of the analysis is the part worth acting on. Looking at what distinguishes brands that appear in AI answers from those that do not, the strongest correlations are not on-page factors at all: they are video presence and the frequency with which a brand is mentioned by name across the web.
This points at something the click finding does not. The remedy for displacement is not to optimise the page that is being displaced — it is to become the kind of brand the answer draws on, and the signals that predict that turn out to live mostly off your own site. It reframes the problem from a page-level optimisation task into a presence-building one, which is a much larger and slower undertaking.
The click loss is the headline, but the correlation analysis is the actionable part: video presence and branded mentions track most closely with appearing in AI answers — neither of which is an on-page fix.
There is a coherent mechanism behind the mention finding, even though the study establishes correlation rather than cause. Engines composing grounded answers cross-check across independent sources, and a brand mentioned frequently and consistently across the web supplies exactly that corroboration. A brand that appears only on its own property gives an engine nothing to verify against.
This also aligns with what is known about how these systems assemble entity understanding: repeated independent description is what allows a model to hold a confident picture of what a brand is and what it is authoritative on. Mentions are the raw material of that picture, which is why their frequency would plausibly track with being cited — the engine needs to know who you are before it will name you.
The video correlation is initially more surprising, and several explanations are plausible. Video platforms host an enormous volume of instructional, review, and comparison content that answer engines demonstrably draw on. Video presence also tends to accompany broader marketing investment, so it may partly proxy for overall brand prominence rather than acting directly.
Both readings have the same practical consequence, which is convenient: building genuine video presence where your topics are discussed is worth doing either because it directly supplies citable material or because it is part of the broader presence that does. What would be a mistake is to treat it as a mechanical lever — publishing video for its own sake, disconnected from the questions your audience actually asks.
What it cannot tell you: whether video presence or branded mentions cause AI visibility, or merely accompany the kind of brand prominence that does. Correlational findings of this kind are directionally useful and should not be read as a mechanism.
The temptation with a correlation finding is to treat it as a recipe: publish video, generate mentions, appear in answers. That reasoning is unsound, because both signals plausibly proxy for something broader — overall brand prominence — which is what may actually drive the citation. Manufacturing the proxies without the underlying substance is unlikely to reproduce the effect.
The defensible reading is narrower and still useful. These signals mark out the kind of brand that gets cited: widely discussed, present where its topics are covered, recognisable as an entity. Building that genuinely is a sound strategy on its own merits, and the correlation suggests it also serves AI visibility. Building the appearance of it is not the same thing and is unlikely to work.
This study measures what the Overview costs you. DUNkē measures whether you are inside it — citations across eight AI engines, per prompt, benchmarked against competitors, tracked over time.
This study and the longitudinal CTR research point the same direction with different emphases, and reading them together is more informative than either alone. The longitudinal work establishes that the effect moved over time — a steep drop followed by partial recovery. This study measures the effect at position one specifically, where the absolute stakes are highest.
The two are consistent: a severe effect, concentrated hardest on the top position, somewhat less severe than the initial trough suggested. Where they differ is in what they can support. The longitudinal design speaks to trajectory; this one speaks to magnitude at a specific rank and adds the correlational hint about remedy. Neither is a substitute for the other, and citing one while ignoring the other produces an incomplete picture.
The click measurement is fundamentally a planning input rather than an instruction. It tells you to revalue position one on Overview-triggering queries, adjust traffic forecasts for the segment of your portfolio that is exposed, and stop treating rank as a sufficient proxy for visibility — because on these queries it demonstrably is not.
It also argues for a specific diagnostic habit: when a page holding a strong position sees click-through fall, check whether an answer feature has appeared before investigating the page itself. A great deal of effort is currently being spent rewriting titles and descriptions in response to displacement that no metadata change can address, and the first check would prevent most of it.
The correlational half translates into a longer-horizon programme rather than a task list. If being mentioned across the web and being present in video are markers of the brands that get cited, then the work is to become genuinely more present and more discussed — through earned coverage, authentic community presence, useful content in the formats your audience actually uses, and the entity clarity that lets engines connect it all to you.
This is slower and less tractable than on-page optimisation, which is precisely why it is a durable advantage once built. It is also work most organisations underinvest in relative to on-page effort, because its returns are diffuse and hard to attribute. The finding is a reasonable argument for rebalancing that allocation.
Both halves of this study come with real limits. The click measurement is an average across 300,000 keywords spanning intents and verticals that behave very differently — your own exposure could reasonably be half or double the headline figure depending on your query mix.
The correlation analysis is correlational. It identifies signals that accompany AI visibility; it does not establish that producing those signals produces the visibility. Treat it as a description of what cited brands look like, not as a mechanism to operate.
Put the two findings together and they say something many teams would rather not hear. The damage lands on the page, and the remedy mostly does not live there. Losing more than half your click-through at position one is a page-level symptom of a brand-level condition — and the signals that predict who gets cited are the accumulated results of presence, coverage, and recognition built over years.
That is an inconvenient shape for a problem, because it means the fastest available responses are the least likely to help. But it is consistent with everything else known about how grounded answer engines choose sources, and recognising it early is better than spending a year optimising metadata against an effect that metadata does not touch.
The finding exposes a structural blind spot in conventional measurement. A rank tracker resolves a query and records the position of your URL in the organic list. It does this accurately and it will keep doing it accurately while the value of that position halves, because position and return are different quantities and the tool only measures one.
The consequence is a reporting failure that looks like success: rankings stable or improving, traffic falling, and no obvious explanation in the rank data. Teams in this situation frequently conclude that click-through optimisation is the answer and spend months on metadata. The first diagnostic should instead be whether an answer feature has appeared, which requires looking at the results page rather than at the rank report.
A percentage drop in click-through is a relative measure, and translating it into business impact requires the absolute numbers behind it. A large relative fall on a query with modest volume may be immaterial; a smaller relative fall on your highest-volume commercial term may dominate the total effect. The headline figure describes intensity, not consequence.
This is why exposure analysis has to work in absolute terms — sessions lost, weighted by conversion value — rather than in percentages. It frequently reorders priorities substantially, because the queries with the most severe relative penalty are often not the queries that matter most commercially. Doing the arithmetic properly is what turns an alarming statistic into a defensible list of where to act first.
The study focuses on position one, but the implications extend downward with a twist. Lower positions had less click-through to lose in absolute terms, so their absolute losses are smaller — but they also sit further down a page that has grown taller, which means the vertical displacement affects them more severely in relative terms.
The practical upshot is that being pushed from first to third on an Overview query is a compounded loss: less click-through available overall, and a smaller share of what remains. It also raises the return on winning the citation rather than the ranking, since a cited source in the summary appears above every organic position regardless of where the page itself sits. For a page ranking fourth, citation is worth considerably more than a two-place ranking gain.
If branded mentions and video presence are what distinguish cited brands, the implied time horizon for improvement is long. Neither can be manufactured quickly — mention frequency accumulates through coverage, discussion, and reputation over years, and genuine video presence requires sustained production rather than a campaign.
This is worth stating clearly when setting expectations, because it contradicts the hope that AI visibility is a technical fix. The on-page work covered elsewhere in this Academy — answer-first structure, extractability — is fast and genuinely helps. The signals this study identifies as most predictive are slow. A programme that promises rapid citation gains is either doing the fast work only, which has a ceiling, or overpromising.
Given that the strongest correlates are slow, it is worth being precise about what can move quickly. Restructuring existing high-authority pages so their answers are extractable is fast and frequently effective, because the credibility already exists and only the form is wrong. Correcting entity ambiguity is fast. Ensuring AI crawlers are not blocked is immediate.
These do not substitute for the presence-building the correlations point at, but they remove the obstacles that prevent existing authority from converting into citations. The sensible sequence is therefore to clear the fast blockers first, which sometimes produces surprising immediate gains for brands that were already well-regarded but structurally unciteable, then commit to the slow work that lifts the ceiling.
This study offers a good opportunity to establish a general habit, because correlational findings about AI visibility are proliferating and most are being reported as instructions. The question to ask of any of them is what else could explain the association — and in most cases the answer is overall brand prominence, which plausibly drives both the measured signal and the citation.
That does not make such findings useless. Knowing what cited brands look like is genuinely informative for strategy, and it points investment in a defensible direction. What it does not license is a mechanical programme to reproduce the signal in isolation. The discipline is to treat these findings as descriptions of a profile to build toward, and to be sceptical of anyone converting them into a checklist of things to manufacture.
Because the remedy is slow and diffuse, measurement discipline matters more than usual. The chain to watch runs from presence-building activity, to mention frequency and video reach, to citation rate on your tracked prompts, to branded demand. Each link takes time, and skipping straight to the last one produces a verdict long before the mechanism has had a chance to operate.
The practical arrangement is to track citation rate as the leading indicator you can actually influence, with mention volume as an input measure and branded demand as the downstream corroboration. Reviewed over quarters rather than months, that chain tells you whether the investment is working. Judged on monthly conversion data, it will look like failure regardless of whether it is succeeding.
If the signals that best predict AI visibility are branded mentions and video presence, then the finding has an uncomfortable budgetary implication: some of what determines search visibility now sits in budgets that search teams do not control. Brand marketing, PR, and content production for platforms outside your site all contribute to the correlates identified here.
That argues for treating AI visibility as a cross-functional objective rather than a search deliverable. In practice it means the search team needs visibility into what PR and brand are doing and a say in where it is directed, since coverage aimed at the topics you want to be cited for is worth considerably more for this purpose than coverage aimed anywhere else. Organisations where those functions operate independently will underperform ones where they are pointed at the same targets.
Because the findings are correlational and drawn from a broad panel, the useful follow-up is to check whether they hold for you. If mention frequency genuinely tracks with citation, you should be able to observe it in your own portfolio — comparing the topics where you are most discussed against the topics where you are most cited, and seeing whether they align.
Where they do, you have local confirmation and a clearer case for investment. Where they do not, that is informative too: it may indicate that your mentions are concentrated in places engines do not draw on, or that a different constraint — entity ambiguity, poor extractability, technical access — is binding before the mention signal can matter. Either result is more useful than accepting a published correlation on faith.
If this study reduces to a single operational statement, it is that ranking and being cited have become separate achievements requiring partly separate work. Position one still earns clicks, but far fewer when an Overview sits above it, and the things that get you into that Overview are not the things that got you to position one.
Holding both objectives simultaneously is the practical consequence. Continue the ranking work, because it still returns; add the presence and structural work that earns citations, because that is where the displaced share went. A team optimising only for rank is defending a diminishing asset, and one that abandons rank for citations is discarding traffic that still arrives.
If one working habit comes out of this research, it is checking the results page before diagnosing a page. When click-through falls on a query where your position is unchanged, the first question is what else is now on that page — not what is wrong with your title, your description, or your content.
That single check would prevent a substantial amount of misdirected work currently being done across the industry, because the symptom of displacement looks exactly like the symptom of a weak listing in every report that does not show the results page. Building the check into the diagnostic routine costs nothing and reliably distinguishes an environmental change from a performance problem, which is the distinction most worth getting right.
When an AI Overview is present, position-one click-through falls by well over half, which means ranking first now returns roughly forty per cent of what it used to on those queries. Because the top position captured the most clicks to begin with, it absorbs the most damage in absolute terms — making the brands with the strongest organic positions the most exposed rather than the most insulated.
The more actionable half of the study is the correlation analysis, which finds that video presence and branded mentions across the web track most closely with appearing in AI answers. That points the remedy off-page: becoming the kind of widely-discussed, recognisable, present brand that answer engines draw on. Read the correlations as a description of what cited brands look like rather than as a lever to pull, and rebalance effort toward the presence-building that most teams currently underfund.
“The clicks are lost on the page. The reason they went elsewhere mostly isn’t on the page at all — which is why the fastest fixes are the least likely to work.” The Age’X Research Team
Get a free GEO audit — the same analysis behind every article here.