AI Summaries Are Draining Wikipedia's Traffic

Fewer and fewer internet users are landing directly on Wikipedia, since search engines and chatbots increasingly hand them a ready-made answer first. The Wikimedia Foundation has calculated that real human visits to the encyclopedia fell 8 percent over the past year, while researchers in a separate study on Google's AI Overviews found losses reaching as much as 15 percent of traffic to English-language articles.

What Wikimedia's numbers show

The Wikimedia Foundation had been tracking the traffic decline for months, but only presented hard numbers in the fall of 2025. Marshall Miller, senior director of product at the Wikimedia Foundation, wrote on the Diff blog that after stripping out traffic generated by sophisticated bots posing as human users, the number of real pageviews from March through August 2025 was 8 percent lower than a year earlier.

The foundation had previously acknowledged that some of the unusually high traffic in May and June came from exactly these disguised bots, mostly originating in Brazil. After updating its detection systems and recalculating the data, the picture became clearer: people really are visiting Wikipedia less often, while traffic generated by AI systems is climbing fast.

We believe that these declines reflect the impact of generative AI and social media on how people seek information, especially with search engines providing answers directly to searchers, often based on Wikipedia content - Marshall Miller, Senior Director of Product, Wikimedia Foundation

Bots read more, people read less

Wikimedia Switzerland added a further contrast in March 2026: while pageviews generated by humans fell 8 percent, traffic from automated systems, including crawlers training language models and AI assistants pulling data in real time, rose 50 percent. Artificial intelligence is becoming an increasingly important intermediary for accessing the knowledge stored on Wikipedia, even as it stops sending human traffic there.

The mechanism is simple. Search engines like Google and Bing increasingly display a ready-made answer directly in the search results instead of directing users to the source page. Chatbots such as ChatGPT, Gemini and Claude draw enormous amounts of training data from Wikipedia's resources, but don't send users back to the articles they drew on.

What AI summaries really cost

The scale of the phenomenon is shown by a study published in May 2026 that analyzed the impact of Google's AI Overviews feature on traffic to the English Wikipedia. The authors calculated that mere exposure to these summaries cuts daily traffic to articles by about 15 percent. Across a sample of more than 52,000 articles, that translated into 11.5 million fewer visits per day, or roughly 4.21 billion fewer visits a year.

The losses aren't distributed evenly. Culture articles were hit hardest, with traffic down 19.6 percent, followed by geography articles at 16.6 percent. Historical and social topics lost 9.9 percent of traffic, while STEM articles fared best, down just 7.4 percent. The authors also estimated that for publishers relying on advertising, this would amount to a loss of between $35.71 and $102.12 million a year.

The source-data paradox

Prof. Dariusz Jemielniak of Kozminski University (Akademia Leona Koźmińskiego), vice president of the Polish Academy of Sciences (Polska Akademia Nauk), described the situation as paradoxical in January 2026. He pointed out that as much as 15 percent of the input data used to train most large language models came from Wikipedia itself. People are increasingly choosing the convenient, fast answer from a language model instead of looking up information in the encyclopedia themselves, even when that answer turns out to be fabricated and unsourced.

That creates a feedback-loop risk. The less often people land directly on Wikipedia, the fewer volunteers get involved in editing and expanding articles. A decline in the number of editors could, over time, lower the quality and freshness of the very content that feeds those same AI models, deepening the problem at its source.

What's next for Wikipedia

The Wikimedia Foundation isn't just sounding the alarm. It has already signed licensing agreements granting access to its content to Amazon, Meta, Microsoft, Perplexity and Mistral AI, though OpenAI is still missing from that list. Miller also argued that AI platforms should more clearly credit Wikipedia as a source and make it easier to click through to the original articles, rather than replacing a visit to the site altogether.

For readers in Poland, the problem carries practical implications well beyond the encyclopedia itself. The same mechanism, in which search engines and chatbots serve a ready answer instead of routing traffic to the source, also affects newsrooms, how-to guide creators, and expert websites that have spent years building visibility through search engine links. Wikipedia, as one of the resources most frequently cited by AI models on the internet, is today the most visible example of what traffic loss looks like in an era of answers generated directly within search results.