AI search · 9 min read
How much traffic AI Overviews take, and why the published numbers disagree
You have probably seen a figure for how much AI Overviews reduce clicks. It was almost certainly produced by a company selling something, from a keyword set it chose, over a period it chose. Here is one study whose method is public, what Google says in response, and how to read the distance between them.
Written by Zubair Afzal, FounderUpdated
The state of the evidence
Five reasons two credible-sounding figures can be miles apart
There is no shortage of statistics on this subject. There is a shortage of statistics whose method you can inspect, and the ones that circulate hardest are usually the ones attached to a product launch.
- The samples describe different populations.
- A panel study measures what a group of people did across everything they searched. A keyword study measures click-through rates for a list of terms somebody selected. Those are not two estimates of one number. Quoting them in the same paragraph, or splitting the difference, gives you a figure that describes no group of searches that exists.
- Nobody controls for which queries get an AI Overview.
- These features appear far more often on question-shaped and informational queries than on navigational or transactional ones. So a comparison between searches with and without an overview is largely a comparison between two different kinds of search. Some of the observed difference is the feature; some of it is that people behave differently when looking something up than when looking something out.
- The keyword sets are chosen, not sampled.
- Most tool-vendor research starts from the keyword universe that vendor happens to track, which skews towards commercially valuable terms that agencies pay to monitor. That is a reasonable place to start and it is not a random sample of searches, so the resulting average is an average of that list rather than of search.
- The periods are short and the systems keep changing.
- A measurement taken over a few weeks is a snapshot of one configuration of a product that is being altered continuously — how often the feature triggers, how it is laid out, how prominent the links are. A figure from last spring may be describing a version of the interface that no longer exists.
- Only one party has the actual data, and it is interested.
- Google can measure this exactly and has not published the underlying data. Everyone else is inferring from outside with partial information. That does not make Google wrong, and it does not make the external estimates worthless. It does mean every number in this debate comes with an incentive attached, including the ones that suit an agency.
One study, stated properly
What Pew Research Center measured, with the caveats attached
Browsing data from 900 US adults, collected in March 2025 and published in July 2025. These are the three figures, and the limitations are set out immediately below them.
- Visits with an AI summary where a user clicked a traditional search result
8%
Visits with an AI summary where a user clicked a traditional search result
Source: Pew Research Center, July 2025 (900 US adults, March 2025 browsing data)
- Visits with no AI summary where a user clicked a traditional search result
15%
Visits with no AI summary where a user clicked a traditional search result
Source: Pew Research Center, July 2025 (900 US adults, March 2025 browsing data)
- Visits where a user clicked a source link inside the AI summary itself
1%
Visits where a user clicked a source link inside the AI summary itself
Source: Pew Research Center, July 2025 (900 US adults, March 2025 browsing data)
We use this study rather than a larger-sounding one because its method is published and its limits are stated by the people who ran it. Those limits are real and they matter: 900 adults, one country, one month, and a comparison of visits rather than of the same query under two conditions.
The third figure is the one worth sitting with. Where a summary appeared, a source link inside it was followed on roughly one in a hundred visits. Whatever value there is in being named in an answer, on this evidence it is mostly not arriving as a visit.
Reading it properly
What this evidence does and does not establish
It establishes that, for this panel in this month, visits accompanied by an AI summary produced fewer clicks to websites than visits without one. That is a genuine finding and it is consistent with the mechanism anyone would predict: if the question is answered above the results, fewer people need the results.
It does not establish a causal size. Because the feature triggers more often on informational queries, part of the gap between 8% and 15% reflects the kind of question being asked rather than the presence of the summary. A study that could separate those would need to show the same query to comparable people with and without the feature, which is something only the platform can do.
It says nothing at all about commercial searches specifically, which is the segment most businesses actually care about. Somebody searching for a supplier, a price or a location behaves differently from somebody looking up a definition, and no published research we would stand behind isolates that segment properly.
And it is one month, in one country, from one panel. Repeat measurement over time would tell you far more than a single well-reported snapshot, and that is exactly what nobody outside the platform is positioned to do.
The practical consequence is that the honest sentence for a business plan is narrower than the sentence people want. It is: informational queries appear to be losing clicks, the size of that effect is not settled, and our own exposure is measurable from our own data.
The independent study and Google’s public position, side by side
These are usually presented as contradicting each other. Read carefully, they mostly do not — they describe different quantities.
| Dimension | Pew Research Center | |
|---|---|---|
| What is claimed | Clicks to search results were less frequent on visits where an AI summary appeared than on visits where none did. | Total organic click volume from Search to websites has been broadly stable year over year, with clicks of higher quality. |
| The quantity measured | Share of visits producing a click, by whether a summary was present. | Aggregate click volume across all queries, without a breakdown by query type. |
| Sample and period | Browsing data from 900 US adults, March 2025, published July 2025. | Google’s own complete data. The scope, definitions and figures behind the statement were not published. |
| Independence | A non-partisan research organisation with no product in the category. | The operator of the product being assessed, responding to critical coverage. |
| What it cannot tell you | Causal size, behaviour on commercial queries, anything outside the US, or whether the effect has grown since. | Whether any particular segment, site type or query class is down, since a stable total can hide a large redistribution. |
| How to use it | As a citable data point, quoted with its sample and period in the same sentence. | As evidence against the strongest catastrophe claims, not as evidence that your informational pages are fine. |
A filter
Five questions to put to the next statistic you are shown
You do not need to read a whole study to know whether to quote it. These five questions dispose of most of what circulates, in about two minutes.
- Who produced it, and do they sell a product whose value rises if the number is alarming?
- What is the sample — how many queries, sites or people, and were they selected or sampled?
- What period does it cover, and how long ago was that in the life of a product that changes monthly?
- Does it compare the same query with and without the feature, or two different sets of queries?
- Does it separate informational from commercial queries, and if not, which is your traffic made of?
The industry argument about the global figure is less useful than it looks, because the only number that can change what you do is your own, and it is sitting in Search Console.
Segment your queries by shape, compare impressions with clicks over a long enough window to rule out seasonality, and look at where the two curves separate. That gives you an exposure profile specific to your business rather than an average of somebody else’s keyword list.
If the separation shows up on definitional and how-to queries and not on product, location or brand terms, you have the common pattern, and the response is a reallocation rather than an emergency.
Questions
What people ask when they need a number they can defend
What is the most reliable figure available for AI Overviews click loss?
The Pew Research Center analysis published in July 2025 is the most inspectable one we are aware of. It used browsing data from 900 US adults collected in March 2025 and found that people clicked a traditional search result on 8% of visits where an AI summary appeared, against 15% of visits where none did.
That is a real, replicable measurement with its method published. It is also 900 people, one country and one month, and it compares visits rather than identical queries, so it should be quoted with those qualifications attached rather than as a headline percentage.
Why do the published numbers differ so much?
Because they measure different things and say so, and then get quoted as though they measured the same thing. One study counts what a panel of people did across all their visits; another compares average click-through rates for a chosen keyword set; another looks at one industry.
None of those is wrong on its own terms. Putting them in the same sentence, or averaging them, produces a figure that corresponds to no population at all.
Can Google’s statement and the studies both be true?
Yes, and that is the most likely reading. Google has said publicly that total organic click volume to websites has been broadly stable year over year. Total volume is dominated by navigational and commercial searches, which are far less affected than question-shaped ones.
So a large decline concentrated in informational queries is compatible with a stable overall total. It is not compatible with the reassurance a publisher of definitional content would want.
Should I quote any of these numbers in a business case?
Quote your own. Search Console gives you impressions and clicks by query for your site, and the pattern that matters — impressions holding while clicks fall on question-shaped queries — is visible without reference to anybody’s study.
If an external figure is genuinely required, use one whose sample and period you can state in the same sentence, and state its limitations in the next one.
Does Search Console tell me when an AI Overview appeared?
Not as a separate breakdown. Impressions and clicks from AI features are included in the overall Search performance data rather than split out, so you cannot filter to them directly.
What you can do is segment by query shape and page template and look at where clicks and impressions separate. That is inference rather than measurement, and it should be described as such internally.
Get your own figure instead of somebody else’s average
We will take your Search Console data, split it by query shape and template, and show you where clicks and impressions have actually separated. You get the working, not a score.
Related
Where to go next
Last updated