AI click loss now has a causal number

A causal click-loss estimate is the share of clicks a page loses because an AI answer appeared, measured by removing that answer for a randomly assigned group of users, rather than by comparing traffic before and after a rollout.

Almost every AI traffic number in circulation is the second kind. Panels observe that pages carrying an AI summary get fewer clicks than pages without one, and the gap gets quoted as the cost of AI. The Pew Research Center study that anchored most of 2025's coverage worked this way: across 900 U.S. adults and 68,879 unique Google searches, users clicked a traditional result on 8% of visits with an AI summary versus 15% without. Useful, but confounded. Google decides which queries get a summary, and the queries it picks are systematically different from the ones it skips.

Two preregistered field experiments have now closed that gap. Both randomly assigned real users to see or not see Google's AI features, then watched where the clicks went.

#What the experiments did

The first, by Saharsh Agarwal and Ananya Sen, ran a custom Chrome extension across 1,065 U.S. desktop Chrome users recruited on Prolific between January 7 and February 10, 2026, collecting 68,089 unique searches. One arm saw normal Google. One arm had AI Overviews silently suppressed. A third, smaller arm was redirected into AI Mode. The paper is on SSRN.

The second, from Stephanie T. Wang, Jeffrey Gleason, Yakov Bart, Christo Wilson and Danaé Metaxa, used the same basic instrument with a different design: three days of baseline, then seven days of assigned treatment, with N=1,100 completing the experiment across a no-AI arm, a normal-Google control, and an arm where every search was redirected to AI Mode.

Same platform, same country, different research groups, different months. That makes the places where they agree worth taking seriously.

#AI Overviews: real, and narrower than the headline

Suppressing AI Overviews raised outbound organic clicks from 0.37 to 0.62 per search, a 39.8% reduction attributable to the summary, while zero-click searches rose 34.5% when the feature appeared. That is close to the observational estimates, which is itself informative. The confounding did not turn out to be doing most of the work.

The breakdowns are where the planning value sits. In that experiment, 87% of overviews rendered at position zero, and removing those produced an 88% relative click increase while overviews in lower positions had no measurable effect. The damage is a placement phenomenon, not a presence phenomenon.

Intent mattered even more. The query mix ran 71% informational, 18% navigational and 12% transactional, and only the informational queries moved: 0.26 more clicks per search when the overview was removed, with navigational and transactional effects statistically indistinguishable from zero. Sponsored clicks and clicks into Google's own properties did not shift meaningfully either.

If you run a blended sitewide forecast off a single click-loss percentage, you are applying an informational-query effect to a whole inventory. Segment first. The pages at risk are the definitional and explanatory ones. Product, pricing, comparison and branded-navigation pages were not measurably affected in this experiment.

#AI Mode is a different regime, not a bigger version of the same one

The AI Mode arm is where the two studies get uncomfortable. Forcing searches into AI Mode cut external click-through by 18.8 percentage points against normal Google (95% CI -22.2 to -15.3), while removing AI entirely gained only 8.8 percentage points (95% CI 2.3 to 15.3). The asymmetry is the story. The distance from today's Google to a no-AI Google is much shorter than the distance from today's Google to an AI Mode Google.

The destination breakdown cuts against a common assumption. Under AI Mode, click-through fell 12.5 points for news sites, 21.2 points for Reddit and 9.9 points for Wikipedia. The sources people describe as AI-proof because they get cited constantly are the ones losing the most visits. Citation volume and visit volume are separate quantities, and AI Mode widens the distance between them.

Both AI Mode results carry caveats worth stating. The Agarwal and Sen arm was exploratory, with the authors noting that higher dropout likely biased its results upward. The Wang arm redirected every single search into AI Mode, which is heavier usage than any real user has today. Read them as an upper bound on a scenario, not a forecast of next quarter.

#Nobody is getting a better search experience

The finding both studies share, and the one that gets the least attention, is a null. Removing AI Overviews produced no measurable change in user-reported satisfaction, information quality, or ease of finding information, with the authors describing precisely estimated nulls, statistically indistinguishable from zero across all three dimensions. Downstream engagement (back-button rate, sub-ten-second bounces, time on page) did not move either.

AI Mode did worse than nothing. It reduced trust by 0.34 points on a seven-point scale, satisfaction by 0.73 standard deviations and perceived agency by 0.66, added 0.43 minutes per session, cut sessions per day by 0.92, and raised the fraction of users searching on a competitor engine by 11.2 percentage points.

That last number is the one to keep. A feature that costs publishers a fifth of their click-through while pushing users toward rival engines is not a settled equilibrium. Plan for the AI Mode regime, but do not treat its current shape as permanent.

#What this changes in practice

Three things.

Stop quoting a single blended AI click-loss figure. The causal estimate applies to informational queries where a summary occupies position zero. Tag your inventory by intent and apply the loss where the experiments actually found it.

Model AI Overviews and AI Mode as separate scenarios. They have different magnitudes and different victims. A reference-heavy content library that survives AI Overviews comfortably is exactly the profile that AI Mode hits hardest.

Track citations and visits as two metrics, never as a proxy for each other. Under AI Mode, the correlation between being the source and receiving the visitor gets weaker, and a citation-only dashboard will show a healthy trend while the traffic line falls.

The measurement got better this year. The honest summary is that AI Overviews cost real clicks in a narrow, identifiable slice of queries, AI Mode costs considerably more across a wider slice, and neither is currently buying users a better experience.