MarketLensMost of the AI Overview debate runs on aggregate platform data — impression counts, CTR curves, rank distributions. All of it is inferred from the server side. It tells you what was rendered. It cannot tell you what a person did with it.
The Pew Research Center study is different, and that difference is the whole point. Pew did not query an API. It recruited real people, installed consented browser tracking, and watched 68,879 unique Google searches happen on their own devices between 1 and 31 March 2025. That is a behavioral log, not an impression log — and behavioral logs are the only place you can observe the thing SEO teams actually care about: where the click went when it stopped coming to you.
Everything below is drawn from the primary source: Pew Research Center, “Google users are less likely to click on links when an AI summary appears in the results”, published 22 July 2025. Read it yourself before citing it; this study is quoted secondhand more often than it is opened.
This article breaks down what the panel measured, why its numbers deliberately disagree with your Search Console, and how to reproduce the same cohort split on your own property before you reallocate a single dollar of content budget.
What a browser panel measures that Search Console cannot
Search Console is a server-side ledger. It records an impression when your URL is present in the rendered results payload, and a click when a user follows it. It has no concept of scroll depth, viewport position, dwell, or session termination. If Google renders a 1,200-pixel AI summary above your listing, GSC still logs your impression at position 3 — the user simply never saw position 3.
The Pew panel closes that blind spot. Because the instrumentation sits in the browser, it can distinguish four outcomes that GSC collapses into one:
- The user clicked a standard organic or paid result.
- The user clicked a citation link inside the AI summary.
- The user issued another query without clicking anything.
- The user ended the browsing session entirely.
That fourth outcome is the one nobody has been counting, and it is where the traffic goes.
The headline behavioral splits
The figures Pew published describe a consistent pattern across the panel: on search visits where an AI summary appeared, panelists clicked a traditional result on roughly 8% of visits, compared with roughly 15% of visits with no summary present. Clicks on the source links embedded inside the summary were rare — on the order of 1% of visits.
The session-termination number is arguably more consequential than the CTR number. Roughly 26% of visits with an AI summary ended the user’s browsing session, against roughly 16% without. A user who ends the session does not run a refined query, does not open a comparison tab, and does not enter anyone’s funnel.
| Behavior | AI summary present | No AI summary | Practical reading |
|---|---|---|---|
| Clicked a traditional result | ~8% of visits | ~15% of visits | Roughly half the click opportunity disappears |
| Clicked a link inside the summary | ~1% of visits | n/a | Citation placement is visibility, not traffic |
| Ended browsing session | ~26% of visits | ~16% of visits | Query resolved at the SERP; no second chance |
The one-percent citation click rate deserves a hard look, because a lot of GEO positioning quietly assumes otherwise. Being cited in an AI Overview is a brand impression with a hyperlink attached. That is worth something — the reader arrives already primed by the summary — but it is a different unit from a blue-link click, and one percent of visits is the volume you should plan around. Model citations as high-quality and low-volume. Anyone selling citation counts as a traffic strategy is conflating two units that Pew measured separately for a reason.
Why Pew’s percentages will not match your site
Panel studies measure a population. You operate a property. Four adjustments separate the two, and skipping them produces bad forecasts:
Query mix. The panel’s searches are ordinary consumer queries — navigational, trivia, local, shopping. Across the whole sample, an AI summary appeared on 12,593 of the 68,879 searches: a minority of them, roughly one in five. Your own rate could be far higher or far lower depending on what people ask you, and the only way to know is to check your own queries rather than inherit an average built from someone else’s.
Device split. The mobile viewport amplifies displacement dramatically. The same summary that pushes result one below the fold on desktop can push it two full screens down on a phone.
Intent depth. Definitional and “what is” queries are the most summarizable and therefore the most vulnerable. Queries requiring a decision, a price, a booking, or a tool are structurally resistant.
Time. March 2025 is a snapshot of a moving system, and both AI Overview coverage and the interface itself have changed since. A single-month panel explains the mechanism; it does not track the drift. Pair it with your own period-over-period data, and with an understanding of how these summaries differ structurally from the featured snippets that preceded them, covered in AI Overviews versus featured snippets.
Reproducing the cohort split on your own property
You do not need panel software to run this analysis. You need a labeled query list and two months of Search Console data.
- Export 16 months of query-level GSC data for your top pages. Use the bulk BigQuery export if you have it; the UI export truncates and will bias your sample toward high-volume queries.
- Label each query as AI-triggering or not. Sample your top 200 queries, check each in a clean browser session or via a rank tracker that reports AI Overview presence, and store the label. Two hundred queries labeled by hand takes about ninety minutes and is worth more than any estimated dataset.
- Compute the displacement ratio per query: the percentage change in impressions divided by the percentage change in clicks, period over period. A ratio above 1 with falling clicks is the displacement fingerprint.
- Segment by intent, not just by label. Within the AI-triggering cohort, separate definitional queries from comparison and transactional queries. You will usually find the damage is concentrated in the definitional tail, which changes what you rewrite.
- Check position stability. If average position held steady while CTR fell, you have a viewport problem, not a ranking problem. That distinction is the entire difference between a content rewrite and a link-building program — and it uses the same Search Console mining technique as finding striking-distance keywords in positions 4 to 20.
Run this once and you will have something the aggregate studies cannot give you: the size of the effect in your own vertical, on your own pages, with your own query mix.
What the behavioral data implies for content strategy
If roughly a quarter of AI-summary visits end the session outright, then optimizing purely for rank on summarizable queries is optimizing for an audience that no longer exists in the numbers you expect. The strategic response has three parts.
Stop competing on summarizable ground. Any page whose entire value proposition can be compressed into 60 words is a page an AI Overview will replace. Definitional content still has a job — it earns citations and topical coverage — but it should not carry your conversion targets.
Build the answer the summary cannot produce. Proprietary benchmarks, interactive calculators, pricing you actually publish, first-party case studies with named outcomes, and tooling all survive summarization because the summary must send the user somewhere to get them. This is the practical form of information gain — content carrying something a model could not have assembled from everything else it already read — which we treat as a measurable property in measuring information gain and content novelty.
Re-baseline your reporting. If your dashboard still treats impressions as a proxy for visibility, it is now measuring something closer to noise. Report clicks per thousand impressions per cohort, not total impressions, or you will keep reporting growth into a declining funnel.
The honest limitation
One panel, one country, roughly 900 participants, one month. The confidence intervals on the sub-segments are wide, self-selected panels skew toward more technically comfortable users, and browser-level instrumentation cannot see in-app searches. Pew is careful about all of this; secondhand citations of the study usually are not.
The correct use of this data is directional. The direction is unambiguous and has been replicated by every independent methodology that has looked: when a generative summary occupies the top of the results page, fewer people click anything, and more people leave. Calibrate the magnitude yourself.
Your next step
Label 200 of your queries this week and compute the displacement ratio. If the AI-triggering cohort shows impressions flat-to-up while clicks fall more than 15%, you have quantified your own exposure — and you now know exactly which pages need to stop explaining and start offering.
If you would rather have that cohort analysis delivered rather than built, MarketLens runs it as part of the Premium Generative Content Audit, using your connected GSC and GA4 properties across 100% of indexed pages.