Ground Truth.
AI, checked against the source.

News · 2026-11-18

A search experiment finds fewer publisher clicks in forced AI Mode

A field experiment found that forcing Google users into AI Mode reduced publisher click-through by 18.8 percentage points under the researchers’ search definition. The result adds evidence that answer-first search can change referrals, but it does not establish the loss for every publisher or isolate the effect of an ordinary AI Overview.

Key facts

A search engine has traditionally been a junction: a user asks a question, sees links and may visit a source. An answer generated on the results page changes that journey. It can save the user a click, but the site that supplied information may lose a visit that would otherwise support advertising, subscriptions or further reading.

The researchers call their paper “AI in Search Reduces Publisher Referrals Without Improving User Experience: Experimental Evidence.” Their study compares distinct product conditions. The reported 18.8-point reduction is specifically for forced AI Mode under their prompt-based definition of a search. It should not be presented as the result of showing an AI Overview within the usual results page. Google page changes also disrupted the extension used to hide AI features, complicating the no-AI condition.

A separate RevScope analysis reports that another randomized study found 39.8% fewer outbound clicks on searches eligible for an AI Overview. The underlying manuscript was not accessible during the dossier’s verification pass, so that statistic remains a secondary account, not an independently checked primary result. Even as reported, it is not a 40% fall in every publisher’s total traffic. RevScope gives a broader estimate of roughly 18.5% fewer organic clicks across searches in its comparison.

The financial stakes are concrete but difficult to attribute precisely. Nicolas Bouliane, who runs All About Berlin, reports a severe traffic fall and attributes it to AI Overviews. His post does not provide the counterfactual needed to separate that effect from other search changes, and it does not announce the site’s closure.

Google’s guidance describes AI features as a “jumping off point” to links and says clicks from such pages can be higher quality. The reviewed guidance does not publish the underlying quality comparison. A smaller number of more engaged visitors could still be valuable; whether that compensates for lost volume depends on the publisher’s business model and is not settled by the experiment.

The publisher discussion includes both the argument that search engines do not owe websites traffic and concern that cutting visits may weaken the supply of original information. Neither position resolves the empirical question. The strongest next evidence would connect randomized referral effects to publisher revenue and later visits, while keeping AI Mode and Overviews distinct.

These comparisons are easy to confuse because the numbers use different units. A percentage-point change subtracts one rate from another; a percentage reduction expresses a change relative to its starting level. A 18.8-percentage-point fall in one experiment therefore cannot be substituted for a 39.8% relative difference in another. The samples, products and definitions also differ.

The Overview study as described by RevScope compared groups using a browser extension across real searches. The dossier could not retrieve the underlying manuscript, so it cannot independently audit the study’s calculations or method. By contrast, the AI Mode paper’s full text was available. Those differences in source access do not make the secondary-reported study false, but they change how confidently its detailed numbers can be repeated.

There are also several possible paths between an answer and a publisher’s business. Some users might click less often but arrive better informed. Others might return later, subscribe directly, or never discover the source at all. The current evidence does not measure all of those paths. Google’s Search Console documentation says traffic from AI features is included in the overall Web search type rather than a separate line, making a clean site-level accounting more difficult. Independent studies that link referrals to outcomes over time would help distinguish a change in user convenience from a sustainable change in the economics of original reporting.


Primary source, verified: read the paper → (arXiv 2608.18352)

Key questions

Does the experiment prove AI Overviews cut all traffic by 40%?

No. Its 18.8-percentage-point result concerns forced AI Mode, not the separate Overview experiment or every publisher’s traffic.

Why are independent publishers concerned?

Fewer outbound clicks can make it harder for sites that depend on visits and advertising to fund original work.
Cite this

APA

Ground Truth. (2026, November 18). A search experiment finds fewer publisher clicks in forced AI Mode. Ground Truth. https://groundtruth.day/news/google-ai-mode-publisher-referrals-experiment.html

BibTeX

@misc{groundtruth:google-ai-mode-publisher-referrals-experiment,
  title  = {A search experiment finds fewer publisher clicks in forced AI Mode},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {nov},
  url    = {https://groundtruth.day/news/google-ai-mode-publisher-referrals-experiment.html}
}

Topics: search · publishers · research

Comments are replies to this story on Bluesky — reply with any Bluesky account to join in.