What Wikipedia Reveals About AI Overviews And Web Traffic

The Only Independent Measurement Of AI Overviews' Damage To Wikipedia Just Got Smaller — And Shakier

A University of Washington working paper by Mehrzad Khosravi and Hema Yoganarasimhan, last revised on September 2, now estimates that Google's AI Overviews cut monthly external-search referrals to English Wikipedia by roughly 5% after the feature became the US default in May 2024. Using 499,927 matched English-German article pairs and 530,873 English-French pairs as controls — since AI Overviews weren't yet default in Germany or France — the authors calculate a 5.45% decline relative to German Wikipedia and 4.82% relative to French, working out to about 100.27 million fewer referrals a month and 1.20 billion a year. An earlier version of the same paper, built on daily pageviews, had reported a much bigger drop of 15%. The paper isn't peer-reviewed, and Google disputes that the underlying data can isolate its own feature at all.

Why should anyone buying AI-search tools care about a Wikipedia study?

Because Wikipedia's public clickstream data lets researchers build a real control group — something almost no other publisher's analytics can do.

Most sites measuring "AI Overview impact" are stuck reading Google's own aggregated numbers, the same reporting this site has already flagged as inadequate for isolating AI-driven search. Wikimedia releases article-level referral data monthly, which let outside researchers compare markets with and without AI Overviews rather than trust a single vendor's dashboard. That's a rare setup, and it's why this paper — however rough — is being treated as the closest thing to independent evidence available.

How solid is the 5% figure, really?

Not very: the estimate has already shifted from 15% to under 5% across revisions and still can't cleanly separate Google from other search engines.

The original February draft used daily pageviews and produced the 15% figure; the August 26 and September 2 revisions switched to monthly referrals with German and French controls, landing on 5.45% and 4.82%. A separate English-Japanese comparison in the current version shows a 16.53% decline, but over a shorter time frame and with a different control group — the authors themselves call that result "directional support only." Wikimedia's clickstream files also lump every search engine together, which is exactly the gap Google pointed to when it disputed the metric to UOL Tilt. The paper's author countered that non-Google engines make up only a small slice of the traffic being measured, but that's an assumption, not a verified breakdown.

What should you ask your own analytics or AI-visibility vendor?

Ask how "referral" is defined, what counts as a control group, and whether the reported number has already moved between versions.

Every AI-visibility platform selling you a clickthrough or citation percentage is making choices similar to the ones this paper had to revise in public. If a vendor can't show you the equivalent of a German-Wikipedia control — a comparable market or query set unaffected by the feature you're measuring — treat the headline number as directional, not audited, the same way these researchers now describe their own Japanese comparison.

Does a Wikipedia-specific traffic drop even apply to other publishers?

Only partly — Wikimedia says many readers now reach its content through AI tools without visiting wikipedia.org at all, unlike ad-funded sites.

Wikipedia is donation-funded and less exposed to a referral drop than a publisher monetizing pageviews directly. For sites relying on click-through traffic rather than brand exposure, a comparable referral decline would hit revenue in a way it doesn't hit Wikimedia — which is a reason to treat this 5% as a floor for the AI-answer-engine traffic question, not the whole story.

Frequently asked questions

Did Google confirm the 5% number?

No. Google told UOL Tilt that the combined search-engine data can't isolate its own feature, while the paper's co-author said other engines account for only a small share of the traffic in the sample.

What time period does the study cover?

December 2023 through December 2024, with May 2024 — when the US AI Overviews rollout began — treated as the first post-treatment month.

Why use Wikipedia instead of a broader web dataset?

Wikimedia publishes monthly, article-level clickstream files that most publishers don't have access to, which is why researchers picked it as a rare natural test case.

Sources: Search Engine Journal.