Study Puts a Number on What AI Overviews Cost Wikipedia: 5.45%

A peer-reviewed study from the University of Washington provides the first causally-verified estimate of how much traffic Google's AI Overviews take from source websites. Using Wikipedia as a test case, the researchers found that default AI Overview availability cut English Wikipedia's search traffic by 5.45% against a German-language benchmark and 4.82% against a French one. For anyone building a brand's visibility strategy around AI search, this is the clearest evidence yet that being cited inside an AI answer, not just ranking well, is now the metric that matters.

What changed?

The paper, by Mehrzad Khosravi and Hema Yoganarasimhan of the University of Washington, reached its sixth version on 2 September 2026 after a long and contested revision process, according to the arXiv preprint. The researchers estimate the causal effect of AI Overviews on Wikipedia traffic by using the feature's staggered geographic rollout alongside Wikipedia's multilingual structure, comparing English-language articles exposed to AI Overviews with near-identical articles in language editions that weren't shown AI Overviews at the same time.

That design matters because it lets the researchers isolate cause from correlation, rather than simply noting that traffic fell after AI Overviews launched and assuming a link, according to Search Engine Journal.

The paper has had a bumpy publication history. Earlier versions used daily pageviews and reported a decline of roughly 15%. The methodology then switched to monthly search referrals and added German and French control groups in version five on 26 August, with further updates in version six on 2 September. An earlier version was withdrawn entirely before the current approach was settled on, per Search Engine Journal.

The latest revision also includes a further data point suggesting the impact could be larger in some markets. An English-Japanese comparison in the current version shows a 16.53% decline, though this uses a shorter time frame and a different comparison group, and the authors describe it as directional support rather than a confirmed figure.

Google disputes the finding. Its objection centres on methodology: Wikimedia's public data lumps all external search engines together, a limitation the paper itself acknowledges, and Google has argued this prevents the analysis from isolating its own product specifically. The paper's authors counter that other search engines carry only a small share of Wikipedia's overall referral traffic, according to Search Engine Journal.

Why does this matter for your business?

This is the first time an academic, peer-reviewed methodology has attached a specific, defensible number to the traffic cost of AI Overviews, rather than a platform-reported statistic or a marketing-vendor survey. That distinction matters because founders and marketers have spent much of 2026 hearing traffic-loss claims from SEO tool vendors with a commercial interest in the story. A university study using a control group that genuinely wasn't exposed to AI Overviews is a different order of evidence.

For Wikipedia specifically, the stakes were already high. Almost 90% of Wikipedia's visitors have traditionally come from Google search, according to SQ Magazine, which is why even a mid-single-digit percentage decline translates into a meaningful loss of readers and, for the Wikimedia Foundation, donors.

If a site as authoritative, well-linked and heavily cited as Wikipedia loses roughly 5% of its search referrals once AI Overviews go live by default, smaller commercial sites without Wikipedia's brand recognition or backlink profile are unlikely to fare better. The lesson for founders is blunt: ranking well in classic search results is no longer a guarantee of a visit. If Google's AI answer already contains everything the searcher needs, the click may never happen, however good your ranking is.

This reframes the whole SEO conversation. Position on the results page was always a proxy for something else: whether a human being actually reached your site. AI Overviews break that proxy. Being ranked number one and being invisible in practice can now be the same thing.

What should you do now?

The practical response isn't to abandon SEO, it's to treat citation inside AI-generated answers as a separate, additional goal that sits alongside ranking. That means checking whether your brand's key pages are the kind of source Google, ChatGPT, Perplexity and Gemini actually quote when they answer questions in your category, not just whether those pages rank on page one.

Concretely, that involves checking three things. First, whether your most important pages contain a clear, extractable answer near the top: the kind of two or three sentence definition an AI system can lift directly. Second, whether your content carries first-party proof, data, named authorship and evidence that make it a safe source to cite, rather than one of many similar-sounding options. Third, whether you actually know how often your brand shows up when real customers ask AI tools the questions that matter to your business, rather than assuming it does.

That last point is where most businesses are currently blind. Traffic dashboards show what happened on your own site, but they say nothing about whether AI systems are naming your brand in the answers they give to people who never land on your site at all. If you haven't checked, it's worth running your own brand through a free AI visibility audit at Sited to see where you stand before you plan next quarter's content.

Frequently asked questions

Does this study prove Google's AI Overviews are hurting all websites, not just Wikipedia?

The paper's causal design is specific to Wikipedia's unusual multilingual structure, which gave researchers a genuine control group. The same rigorous methodology hasn't yet been applied at this scale to commercial sites, so the exact percentage for other categories of website remains an open question.

Why does Google dispute the findings?

Google's objection is methodological rather than a flat denial. Wikimedia's public data lumps all external search engines together, which the paper acknowledges as a limitation, and Google argues this prevents the analysis from isolating its own product specifically, according to Search Engine Journal.

How is this different from earlier Wikipedia traffic-decline reports?

Earlier statements about Wikipedia's traffic decline described an overall drop in human pageviews without isolating AI Overviews as the specific cause. This paper instead uses a controlled comparison against language editions that weren't exposed to AI Overviews, designed specifically to separate the AI Overview effect from other causes such as social media.

Should smaller businesses worry more or less than Wikipedia does?

Likely more. Wikipedia benefits from being cited constantly across AI Overviews and chatbot answers precisely because of its scale and authority, which offsets some of the referral loss. A smaller commercial site without that citation volume has less to fall back on if its click-through traffic declines in the same way.

What is the difference between AI Overviews and a chatbot like ChatGPT for this issue?

AI Overviews appear directly inside Google's search results, so they intercept clicks that would otherwise go to a website from a search page a user already opened. Standalone chatbots like ChatGPT or Perplexity generate answers in a separate interface, but the underlying risk is the same: if the AI names your brand or quotes your content without sending a click, you need to know that's happening.