Click-through on AIO queries fell roughly 65%, then partly rebounded — but the gap versus non-AIO queries is now the baseline to plan around.
Most of what gets said about AI Overviews and click-through rests on a single measurement taken at a single moment. This study does not: it tracks the same brands across billions of impressions over time, which lets it show something a snapshot structurally cannot — that the initial collapse in click-through was not the steady state. Clicks fell hard, then partly came back. The rebound is the finding most people miss, and the gap that remains is the number worth planning against.
The research tracks click-through rate on queries that display an AI Overview against queries that do not, across a large panel of brands, over a sustained period. Its scale — billions of impressions across dozens of brands — matters less than its shape: because the same properties are observed repeatedly over time, the data can distinguish a temporary disruption from a permanent shift, which single-period studies cannot do.
That distinction is the reason this study is worth reading closely. The question everyone actually wants answered is not what happened in the week AI Overviews expanded, but what the new normal looks like once users and interfaces have settled. A longitudinal design is the only kind that can speak to that, and its answer is more nuanced than the headline number that circulated at the time.
When AI Overviews appeared at scale on a query set, click-through to organic results on those queries fell sharply — by roughly two thirds at the trough. That is a severe drop by any standard, and it is the figure that dominated industry coverage, because it arrived first and it was alarming enough to be repeated widely.
The mechanism is not mysterious. A synthesised answer placed above the organic results satisfies a meaningful share of the queries it appears on, and a user whose question has been answered has no reason to continue. Add the vertical displacement — organic links pushed further down a page that now leads with a summary — and the effect compounds: fewer people need to click, and those who might have are further from the link.
What happened next is the part that rarely travels. Click-through did not stay at the trough. It recovered part of the lost ground and stabilised at a level well below the pre-AIO baseline but materially above the worst point. The disruption was real; the collapse was not the destination.
Several things plausibly contribute. Interfaces changed as the feature matured, altering how prominently sources are displayed. User behaviour adjusted as people learned what these answers are good for and where they still want a source. And the query mix on which Overviews appear shifted as coverage was recalibrated. The study does not adjudicate between these, but it does establish that the endpoint differs from the shock — which is the practically important fact.
If the trough was temporary and the rebound is real, the operationally relevant figure is neither — it is the persistent difference between AIO and non-AIO queries once things settled. That gap is what your forecasts, your traffic models, and your expectations should be built on, because it describes the environment you are actually operating in rather than the moment of maximum disruption.
Planning off the trough overstates the damage and can trigger overcorrection: abandoning content that still performs, or reallocating budget away from search on the basis of a number that no longer holds. Planning off the pre-AIO baseline understates it and leaves forecasts that will simply not be met. The stable gap is the honest middle, and it is the figure this study exists to establish.
A snapshot study can only tell you what a moment looked like. A longitudinal one shows the trough was not the destination — which is why the persistent gap, not the peak drop, is what you should be forecasting against.
The strength of this study is that repeated observation of the same properties controls for a great deal that cross-sectional comparisons cannot. If you compare AIO and non-AIO queries at one instant, differences in query type, intent, and competitiveness contaminate the result — the queries that trigger Overviews are not a random sample. Watching the same set over time removes much of that noise.
What it cannot do is prove causation for the rebound. Observing that click-through recovered does not tell you why, and the candidate explanations have quite different implications: an interface change might reverse, whereas genuine behavioural adaptation probably will not. Understanding this limit is why the study should be read as establishing a pattern to plan around rather than a mechanism to exploit.
What it cannot tell you: why click-through partly recovered, whether the rebound holds, or how the gap varies for your particular vertical and query mix. Magnitudes in studies of this kind swing substantially with keyword selection, so treat the direction as robust and the exact figure as indicative.
Any industry-level figure is an average across query types, verticals, and intents that behave very differently. Informational queries with easily-summarised answers lose more click-through than queries where the user needs to reach a specific destination, transact, or compare in detail. A brand whose demand skews toward the latter will see a smaller effect than the headline suggests; one whose content is largely definitional will see more.
This is not a criticism of the research — a study covering many brands necessarily reports a central tendency. It is a caution against importing the number directly into your own forecast. The reported gap tells you the direction and rough magnitude of an effect that is real; your own data tells you what it means for you, and the two should be used together rather than one substituting for the other.
The practical follow-up to this study is to establish your own version of it. Identify which of your important queries currently trigger AI Overviews, then compare click-through on those against structurally similar queries that do not, over a period long enough to be meaningful. Search Console supplies the click and impression data; the AIO presence has to be established separately, since it is not reported there.
What you are looking for is your own gap, and how it varies across your content types. That number is worth far more than any published average, because it reflects your actual query mix and audience. It also gives you a baseline to measure against as coverage expands, which is the only way to tell whether a future change in traffic is your doing or the environment’s.
Search Console shows the clicks you lost. It cannot show whether you are cited in the Overview that took them. DUNkē tracks your citations across eight AI engines — per prompt, against competitors — so you can see both halves.
This study explains a pattern that has become common in Search Console reporting: impressions rising while clicks stay flat or decline, with click-through falling as a result. Read without context, that looks like a listing problem — a weak title, a poor description, a mismatch with intent — and teams respond by rewriting metadata that was never at fault.
The correct reading is frequently that an answer feature has appeared above the result. You are still being shown, hence the impressions; fewer people are continuing, hence the flat clicks. The remedy is entirely different: rewriting the title will not recover clicks absorbed by a summary, whereas being cited within that summary might. Recognising the signature is one of the more valuable diagnostic skills in current practice.
It is worth stating plainly what the study does not support. It does not show that organic search has stopped mattering: a substantial share of clicks persists on AIO queries, and many queries display no Overview at all. It does not show that content investment is wasted — the content being displaced is frequently the same content the Overview is built from. And it does not show that the decline continues indefinitely, since the observed pattern is a drop followed by partial recovery and stabilisation.
Overreading it in any of those directions leads to poor decisions, and the industry has produced examples of all three. The disciplined reading is narrower and more useful: on the subset of queries where an Overview appears, expect materially fewer clicks than you would have received before, and plan accordingly — while continuing to compete for both the ranking and the citation.
The figures reported here are drawn from a specific keyword panel and a specific observation window. Studies measuring the same phenomenon with different query sets have produced materially different magnitudes, and that variation is genuine rather than a sign that one of them is wrong — they are measuring different slices of a heterogeneous web.
Treat the direction as well-established and the exact percentage as indicative. Any figure quoted to the decimal point about AI Overview impact across the whole of search should be read with suspicion, including this one.
The practical translation into planning is a segmented forecast rather than a blanket adjustment. Split your query set by whether Overviews currently appear, apply a click-through expectation to each segment informed by your own measured gap rather than an industry average, and model the expansion of coverage as a separate variable, since the share of queries showing Overviews has itself been moving.
This is more work than applying a single percentage to total organic traffic, and it produces a forecast that survives contact with reality. It also surfaces which parts of your query portfolio are most exposed, which is directly actionable: those are the queries where citation work matters most and where traffic-dependent goals should be set most cautiously.
Step back from the number and the study says something simple. A portion of the value that organic ranking used to deliver has moved into the answer layer, and it has not come back. The rebound shows the shift is smaller than the initial panic suggested; the persistent gap shows it is real and durable. Both halves of that are important, and most coverage reported only one.
The strategic response follows: continue to compete for rankings, because they still deliver, while adding the work that makes you a source the answer draws on, because that is where the displaced share went. This is not a hedge — it is what the data actually supports. A brand that does only the first is optimising for a shrinking portion of the page, and one that does only the second is abandoning traffic that still arrives.
The average conceals enormous variation between businesses, and understanding where you sit on that spread matters more than the headline. The most exposed are those whose organic demand is concentrated in definitional and explanatory queries — what something is, how it works, why it happens — because those are exactly the questions a generated summary answers completely. Content built to capture that demand is competing directly with the thing displacing it.
The least exposed are businesses whose demand is navigational, transactional, or requires the user to reach a specific destination to accomplish something. A summary cannot complete a purchase, open an account, or deliver a tool. Between those poles sits comparison and evaluation content, which is partly summarisable and partly not — an Overview can list the options but rarely resolves a decision, so click-through erodes without disappearing. Locating your own portfolio on that spectrum is the first step in translating this study into a forecast you can defend.
An easily overlooked implication is that a substantial share of queries display no AI Overview at all, and on those queries nothing has changed. Click-through behaves as it always did, ranking returns what it always returned, and the entire discussion of displacement is irrelevant. For many businesses this is the majority of their query portfolio.
This matters because the industry conversation has tended to generalise a finding about a subset into a claim about search overall. The disciplined framing is that a portion of queries now carry a substantial click penalty, that portion is growing but is not the whole, and the unaffected remainder still rewards conventional optimisation exactly as before. Treating every query as affected leads to under-investing in ground that is still entirely winnable.
The residual traffic on AIO queries is worth examining rather than merely counting, because there is reason to think it differs in character from what preceded it. Users who click through despite having been given an answer are self-selecting: they want more depth, they want to verify a claim, they are ready to act, or they distrust the summary. Those are not casual visitors.
If that holds, the correct expectation is fewer sessions with better qualification — which changes how the loss should be assessed. A third of the traffic delivering the same absolute number of conversions is a materially different outcome from a third delivering a third of the conversions, and only your own conversion data can distinguish them. Measuring conversion rate on AIO-exposed queries before and after is the check that tells you which situation you are actually in.
With hindsight, a permanent two-thirds collapse was improbable, and understanding why builds intuition for reading future disruptions. New features launch in their most aggressive configuration, get tuned in response to how users and publishers respond, and settle somewhere less extreme. Users simultaneously learn what the feature is good for and route around it where it is not.
Both adjustments push in the same direction, which is why the shape observed here — sharp shock, partial recovery, new plateau — is a common pattern rather than a surprise. The practical lesson is to be sceptical of the first measurement of any new disruption. The initial number captures the moment of maximum dislocation, before either side has adapted, and it is almost never where things end up.
Click-through impact and coverage are two variables that have to be multiplied rather than considered separately. A severe penalty on a small share of queries and a mild penalty on most queries produce very different aggregate outcomes, and citing either figure alone tells you almost nothing about total exposure.
Your real exposure is the share of your commercially important queries that display an Overview, multiplied by the click-through penalty on those queries, weighted by what each query is worth. That calculation is straightforward once you have both measurements for your own portfolio, and it produces a number that means something — unlike either input on its own. The companion prevalence study in this hub covers the coverage half of that equation.
The finding makes an unusually clean business case for citation work, because it quantifies what is being lost. If a defined segment of your queries carries a substantial click penalty and the displaced attention went to the cited sources in the summary, then the value of becoming one of those sources is directly comparable to the value of the traffic that left.
This is a stronger argument than the usual case for AI visibility, which often rests on anticipated future importance. Here the loss is measured, present, and attributable to a specific mechanism, and the remedy addresses that mechanism directly. For teams struggling to justify investment in something whose returns are hard to attribute, the displacement figure is the most concrete number available.
Pulling the threads together produces a specific modelling exercise. Segment your query portfolio by whether an Overview currently appears. Apply your own measured click-through gap to the exposed segment rather than a published average. Model coverage expansion as a separate variable, since the exposed share is itself moving. Then weight by commercial value so the output reflects revenue rather than sessions.
The result is a forecast that survives scrutiny and identifies where to act, rather than a blanket percentage applied to a total. It also makes the assumptions explicit, which means they can be revisited as conditions change — and given that this study demonstrates conditions do change, a model whose assumptions are visible is considerably more useful than one with a single number baked into it.
Bringing a finding like this to people who fund search work requires framing, because the raw number invites the wrong conclusion. Presented alone, a sixty-five per cent drop reads as an argument for abandoning the channel, and some organisations have drawn exactly that inference. The rebound and the persistent-gap framing are what make the picture accurate rather than alarming.
The sequence that works is to establish the structural change first, then the measured effect on your own queries, then the segment of your portfolio actually exposed, and only then the response. Leading with the industry headline concedes the argument before you have made it. Leading with your own segmented exposure gives stakeholders a proportionate problem and a specific plan, which is what distinguishes a case for reallocation from a case for retreat.
The most important thing this research cannot settle is whether the plateau holds. A pattern of shock, partial recovery, and stabilisation is consistent with an effect that has run its course — and equally consistent with a pause before coverage expands further and the cycle repeats on a new set of queries.
That uncertainty is a reason to keep measuring rather than to discount the finding. A brand tracking its own click-through gap and its own coverage will see a second wave arriving; one that took a reading in 2025 and filed it will experience the next expansion as an unexplained decline. The study establishes where things landed, and the only way to know whether they stay there is to keep watching.
Click-through on queries showing an AI Overview fell steeply — by roughly two thirds at the trough — and then partly recovered, stabilising well below the non-AIO baseline but materially above the worst point. Because the study observes the same properties over time rather than at a single moment, it can establish that the collapse was not the destination, which is a finding that snapshot research structurally cannot produce and which most coverage of the topic omitted.
The number to plan against is therefore the persistent gap rather than the peak drop. Segment your queries by whether an Overview appears, measure your own gap rather than importing an average built from other people’s query mix, read the impressions-up-clicks-flat pattern as displacement rather than a metadata failure, and continue competing for the ranking while adding the work that gets you into the answer. Both halves of the finding matter, and most of the industry only repeated one.
“The collapse made the headlines and the rebound didn’t. But the number you should be forecasting against is neither — it’s the gap that was still there once everything settled.” The Age’X Research Team
Get a free GEO audit — the same analysis behind every article here.