AI Overviews cut English Wikipedia traffic by about 15 percent

Why Wikipedia's language editions work as a control group

A traffic decline is easy to observe and hard to attribute. Ranking changes, news cycles and the ordinary seasonality of an encyclopedia all move a page view count, and any of them can be running at the same time as a product launch. What a causal estimate needs is a version of the same content that the feature did not reach, observed over the same months.

Wikipedia supplies one. The same article exists in many languages, written about the same subject, exposed to the same global news shocks, and Google rolled AI Overviews out by geography rather than by topic. The paper's Table 1 lists the milestones it works from: testing began on a subset of queries and a subset of search traffic in the United States on 22 March 2024, the feature reached all US users on 15 May 2024, it expanded beyond the US to India, Mexico, Japan, Indonesia, Brazil and the UK on 15 August 2024, and it reached more than 100 countries and territories on 28 October 2024.

That ordering is the entire design. English Wikipedia is the treated group, the Hindi, Indonesian, Japanese and Portuguese editions are the controls, and the panel closes on 14 August 2024. It is worth noticing what the following day holds: the international expansion listed in that same table covers India, Indonesia, Japan and Brazil, which are the readerships behind all four control editions. The window ends one day before the controls stop being controls. The panel itself opens on 28 October 2023 and holds 46,534,093 daily observations across 161,382 article and language pairs, built from 52,262 English articles matched to their versions in those four languages, all of it drawn from Wikimedia's public pageview API.

One substitution in that design deserves stating rather than glossing. Wikimedia's pageview data is not broken down by the reader's country, so the paper never observes who was exposed. It proxies exposure with the language edition, on the stated reasoning that English Wikipedia readership is most closely aligned with the first wave of the rollout in the United States. That is an approximation and the authors treat it as one. It is also the kind of design that the critical survey of 45 GEO studies found to be scarce: most of the generative engine optimization literature observes correlations across sites, and the papers collected on our research page are the exceptions that go looking for a counterfactual.

What the models estimate, and how far apart they are

The headline is one number, and the paper is unusually open about how much that number depends on the scale you estimate it on. Two specifications carry the main result. Without fixed effects, the difference in differences coefficient is a fall of 242.283 daily views with a standard error of 11.4507, which the authors convert to a 16.6 percent decline against the pre period English baseline. Adding article by language and date fixed effects moves the coefficient to a fall of 220.545 daily views with a standard error of 15.1456, and lifts the model's R squared from 0.0019 to 0.9166. That second figure is the one the abstract rounds to approximately 15 percent. Standard errors throughout are two way clustered by article and by date, across 52,262 article clusters and 292 date clusters.

The robustness section is where the honest range sits. Redefining the post period to begin at the full US launch on 15 May 2024, which drops the partial test window, yields a decline the authors put at about 16 to 17 percent. Aggregating to weekly data gives roughly 1,504 fewer weekly views per article, a 14.8 percent decline. A Poisson pseudo maximum likelihood specification, which models proportional rather than absolute change, gives 3.5 percent. A weighted log specification that weights articles by their pre treatment traffic share gives 8.1 percent.

Three point five and sixteen point six are a wide spread, and the paper explains rather than hides it: the authors note that the Poisson model down weights high traffic outliers and captures average proportional change rather than absolute losses, which makes it more conservative when the treatment disproportionately affects high volume articles. In other words the specifications disagree about whether you are counting articles or counting visits, and the answer differs because the loss is not spread evenly across the catalogue. The paper also reports linear pre trend tests and event study diagnostics and states it finds no significant and systematic deviations from the parallel trends assumption.

Reporting the range rather than the best number is the part worth copying. We hold our own scoring rules to the same standard, and the reason is that a single figure quoted without its specification is the easiest thing in this field to move by choosing a different one.

Wikipedia is a cited source, and the visits still fell

The reason to read this paper rather than one of the many traffic laments is that the authors set up both hypotheses before estimating either. They write that AI Overviews can substitute for Wikipedia by providing similar factual summaries directly at the top of the results page, and that on the other hand Wikipedia is a highly reputable and frequently cited source in those answers, so if being cited increases exposure or if the feature expands overall search engagement, Wikipedia could benefit. Those are genuinely opposed predictions and the design can tell them apart. The estimate came back negative.

Scaling the fixed effects coefficient across the whole sample of 52,262 English articles implies roughly 11.53 million fewer page views per day, which the paper puts at about 4.21 billion fewer per year. The authors state that both quantities should be read as lower bounds on the reallocation that would occur under full exposure, and they are explicit that Wikipedia carries no advertising, so the revenue discussion in the paper is an extrapolation to ad supported publishers at comparable scale rather than a measurement of one.

For anyone working on AI visibility, the finding lands on a specific assumption rather than on the whole practice. Citation is still the thing worth pursuing, because a source that is not cited gets neither the visit nor the credit. What the estimate undercuts is the conversion step: the belief that a citation in an answer is a click in waiting. Two earlier posts here point the same way from different directions. Citations in AI Overviews do not track the organic top ten, so the pages being cited are frequently not the pages that were already winning, and a separate framework found that more citations did not mean more of a page reached the answer text. This paper adds the third leg: even the source that is cited most may see fewer people arrive.

None of that is an argument for ignoring how to get cited in Google AI Overviews. It is an argument for being precise about what the citation buys, which on this evidence is presence in the answer rather than a guaranteed session.

Where the decline landed, by topic

The average conceals a spread that matters more than the average does. The paper estimates the effect separately by topic and reports a negative, statistically significant result in all four categories, with the magnitude varying by a factor of more than two and a half. Culture articles fall 19.6 percent, Geography 16.6 percent, History and Society 9.9 percent, and STEM 7.4 percent. A footnote states that these percentage changes are computed from the regression without fixed effects, which is worth carrying with the numbers rather than dropping.

The authors read this as substitution being strongest when a short synthesised answer can satisfy the informational intent, and the ordering is consistent with that. What a Culture article most often answers is a question with a short factual core: a date, a cast list, a plot summary. A STEM article is more likely to be consulted for a derivation, a table of values or a method, which is harder to reproduce usefully in a paragraph above the links.

That pattern generalises past Wikipedia in a way a site owner can act on, and it does not require accepting the exact figures. If your pages answer questions that compress well, the answer engines can compress them, and the visit is the thing at risk rather than the citation. If your pages carry the sort of specificity that does not survive compression, prices, configurations, comparisons against a reader's own situation, then the summary is a route to you rather than a replacement for you. That is the same asymmetry behind an earlier finding on this site, that ChatGPT citations reached only 6.8 percent of prompts and the resulting visit tended to land on the homepage: the traffic that does arrive from an answer surface often arrives without the specific intent the cited page was written to serve.

The practical version of this is unglamorous. Work out which of your pages exist to state a fact that fits in two sentences, and stop scoring those by sessions. Then work out which pages exist to do something a summary cannot do, and make sure those are the ones an assistant can actually read when it goes looking, whether that is Google's surfaces or ChatGPT.

What this design cannot tell you about your own site

Four limits sit between this paper and any conclusion about a particular website, and two of them are ones the authors raise themselves.

The first is the proxy. Exposure is inferred from language edition rather than observed, because Wikimedia does not publish pageviews by reader country. The authors name two consequences and argue both push the estimate downward: exposure within the United States was gradual across users and query classes, so not all English traffic was treated uniformly during the treatment period, and English Wikipedia has a global readership, so some readers counted as treated were in countries the rollout had not reached. Both dilute the treated group with untreated traffic.

The second is the window. The panel ends on 14 August 2024, which is close to two years before this post. It describes AI Overviews in its first months in one country. It says nothing about how the feature behaves in 2026, nothing about AI Mode, and nothing about any other answer engine. A separate SIGIR study on this site found that AI Overviews retrieved less from sites blocking Google-Extended using queries collected in December 2025, which is a different feature in a different year and should not be stacked on top of these figures as though the two measured the same thing.

The third is that Wikipedia is one publisher, and an unusual one. It carries no advertising, it is a general reference rather than a commercial site, its subject range is enormous, and it enjoys a level of trust that most domains do not. A 15 percent estimate for that catalogue is not a forecast for a storefront, a SaaS documentation set or a local service business, and nothing in the paper claims it is.

The fourth is not the paper's fault but is the one that decides what you do next. This is a measurement of an outcome, and outcomes are the hardest thing for an individual site to attribute. The publisher-side variables that do sit in your control are more mundane, and this site has covered several: who blocks AI crawlers, and how that splits by editorial credibility, and the fact that publisher tolls charged to AI crawlers are invisible in robots.txt. Those you can inspect. A counterfactual for your own traffic, you generally cannot.

What a site owner can actually measure

Google's own guidance describes the same feature in the opposite register. The AI features documentation on Search Central, carrying Last updated 2025-12-10 UTC and read on 7 August 2026, states that AI features surface relevant links to help people find information quickly, that there are no additional requirements to appear in AI Overviews or AI Mode and no special optimisations necessary, and that sites appearing in AI features are included in the overall search traffic reported in Search Console under the Web search type.

That last clause is the one to sit with, because it is a measurement statement rather than a marketing one. If AI feature traffic is folded into the same totals as ordinary search traffic, then your own Search Console property cannot separate the two by subtraction, which is precisely why a study of this kind has to go and find a control group in another language. A site owner who wants to know what an answer surface did to their traffic is not being denied a report by oversight. The report is aggregated by design.

What is left is the layer underneath, and it is fully inspectable. Whether a named AI crawler is permitted to fetch your page, and whether the text it receives contains what you think the page says, are both answerable in a single request. That is what what GPTBot sees returns, and the two questions in that order are what two layers decide if AI can read your site is about. Prose parity, the share of the rendered page that survives into the served HTML, is a number you own outright: it does not need a counterfactual, it does not move when Google ships a feature, and if it is low then nothing further up the chain is available to you at all.

Above that sits the question of what assistants actually say about you, which what AI says exists to sample, and which is a different measurement again from a session count. It is worth pairing this paper with Google's own Mythbusting section on generative AI features, which names five tactics the company says do nothing. Between the two you get the useful shape of the field on 7 August 2026: a short list of things that demonstrably do not work, a readability layer you can verify yourself in seconds, and one causal estimate saying that the payoff at the far end is smaller than the pitch assumes even for the site that gets cited most.