The AI Overviews engagement drop, measured

If you have read three articles about the AI Overviews engagement drop, you have probably seen three different numbers: 34.5%, 47%, 58%, 61%. All of them are real measurements. None of them should be your planning number.

This guide lays out what the major studies actually found, explains the four methodological differences that produce the spread, and gives you a procedure for measuring your own exposure — which is the only figure that should drive a decision.

The numbers on the table

Four bodies of evidence carry most of the weight in this debate.

The most careful of these is Pew’s, because it observed real behaviour rather than modelling it from ranking data. Observed studies tend to produce more conservative and more trustworthy figures, and Pew’s sits comfortably in the middle of the published range — which is itself a useful signal about the outliers on either side.

The Ahrefs pair is the most useful for trend, because the two measurements use a comparable method ten months apart. That progression — 34.5% to 58% — is the strongest available evidence that the effect deepened rather than plateaued through 2025 and into early 2026.

Why the studies disagree

Four variables explain nearly all of the spread between a 34.5% figure and a 61% one.

That last row causes more confusion than the other three combined. When someone says “AI Overviews cut clicks in half,” they usually mean the relative figure. When you model revenue impact, the absolute figure is what actually changes your forecast. Being sloppy about which one you are using produces business cases that are wrong by a factor of six.

A 7-point absolute drop and a 47% relative drop describe the same event. Only one of them belongs in your revenue model.

How often AI Overviews actually appear

Effect size only matters in proportion to exposure, and exposure is more stable than most commentary suggests. SERP feature tracking through 2026 puts AI Overview presence at roughly 32.5% of monitored queries, fluctuating only between about 31.0% and 34.4% across measurement periods.

Some datasets report substantially higher prevalence — figures near 48% circulate widely. Those are almost always samples weighted toward informational queries, which is legitimate for some purposes and badly misleading for others. If a prevalence figure does not come with a description of the query mix, it is not comparable to anything.

The narrowness of that 31–34% band is the more interesting finding. It indicates Google has moved past the experimental phase into systematic, settled deployment. AI Overviews are not expanding to swallow every query, and they are not being rolled back.

Signs of stabilisation

Early 2026 data suggests several metrics are levelling off rather than continuing to deteriorate. Alongside the steady prevalence band, ranking volatility dropped sharply from late March 2026 — by roughly 65% in one measurement period against the preceding turbulent phase.

That matters strategically more than it looks. A surface that is stable can be optimised for. The correct posture in 2025 was defensive triage; the correct posture in 2026 is systematic optimisation, because the target has stopped moving. This is the argument developed at length in the AI Search 2026 report.

Measuring your own exposure

You can do this in an afternoon with Search Console and a spreadsheet, and the result is worth far more than any published average.

1. Export your top 200 queries by impressions for the last 16 months from Search Console.
2. Tag each by intent — informational, commercial investigation, transactional, navigational, branded.
3. Check AI Overview presence for a representative sample of each group, logged out and incognito so personalisation does not distort the sample.
4. Compare CTR year over year within each group. If informational CTR fell while transactional held, you have confirmed AI Overview exposure rather than a general ranking problem.
5. Segment by position. A CTR decline concentrated at positions one to three is a strong AI Overview signature, because that is where vertical displacement bites hardest.

If the pattern does not appear, your decline has a different cause — an algorithm update, a technical regression, seasonality, or a competitor. Establishing that before you rebuild a content strategy is worth the afternoon. Our SERP analysis guide covers the competitor side of that check.

Turning it into a planning number

Once you have your own segmented CTR data, three rules keep the forecast honest.

• Model the absolute change, not the relative one. Your revenue depends on how many clicks you lost, not on the percentage that sounds most alarming.
• Model by segment, never in aggregate. A site with 70% transactional traffic and one with 70% informational traffic face completely different exposures. A single blended number hides the entire actionable finding.
• Separate substitutable from non-substitutable pages. The pages bleeding are the ones whose value fits in a paragraph. They are usually a minority of the site producing a majority of the loss, which is good news — it makes the problem tractable.

Then act on the substitutable set specifically: consolidate, deepen, add original data nobody else has, and shift the target from ranking below the overview to being cited inside it. The mechanism breakdown covers which fix matches which signature, and the 2026 playbook orders them by leverage.

What the data does not tell you

Three limitations are worth holding in mind, because every study above shares them.

None of them measure queries that never reached Google. When someone asks ChatGPT instead of searching, no Google-side dataset records it. Every figure in this article measures the click decline within Google, which means the total shift away from websites is larger than any of these numbers imply — by an amount nobody has credibly quantified.

None of them measure downstream value. A lost click from someone who wanted a one-line fact is not equivalent to a lost click from a buyer. Several teams reporting steep CTR declines have simultaneously reported flat or improved conversion rates on remaining traffic, because the visits that disappeared were the least commercially valuable ones. Traffic decline and revenue decline are different measurements and should never be used as proxies for each other.

None of them are stable enough to quote a year from now. Ahrefs’ own figure nearly doubled in ten months using a comparable method. Any number in this space carries an implicit expiry date, which is the strongest possible argument for measuring continuously rather than citing a study once and planning against it for a year.

That is also the honest case for tracking your own prompts on a schedule. A published average tells you what happened to a sample of the web at a moment in time. A stored history of your own answers tells you what is happening to you now — which is the only thing you can act on.

Sources

• Pew Research Center — approximately 68,000 real search queries; clicks fell from 15% to 8% of visits when an AI summary was present.
• Ahrefs — CTR impact of AI Overviews; 34.5% decline for position one (April 2025), approximately 58% (February 2026).
• SERP feature tracking, 2026 — AI Overview presence averaging ~32.5% of monitored queries within a 31.02%–34.40% band; ranking volatility down roughly 65% in one measurement period from late March 2026.
• Publisher traffic reporting, H1 2025 — median publisher down ~10% year over year; non-news content sites down ~14%.
• Google — 2.5 billion monthly active users for AI Overviews and 1 billion for AI Mode, stated 19 May 2026.

Frequently asked questions