What the Wikimedia Foundation told the UN
During the 81st session of the UN General Assembly, the Wikimedia Foundation's chief executive, Bernadette Meehan, spent a week in New York arguing that the AI industry depends on a source it is weakening. The Foundation's account of the week, published on 30 September, makes two claims that matter to anyone who cares what Wikipedia says about them.
The first is about cost. “Wikipedia's content is free. We don't run ads. We don't monetize. But our infrastructure is not free,” Meehan said, adding that scraping by “some of the big technology and frontier AI companies” takes a toll on it. The second is about visibility. An AI-dominated internet, she argued, has made Wikipedia “more vital, but less visible”: when AI platforms and search engines give machine-generated answers without attribution, they “block pathways for people to access the original sources of information to fact-check, research further, and contribute to Wikipedia.” The Foundation asked large commercial reusers for attribution and financial support.
The statement contains no traffic figures. Wikimedia publishes its own, however, and they are public. So we pulled them.
What happened to human pageviews
Wikimedia classifies every pageview by agent: “user” for traffic it believes is human, “spider” for crawlers that identify themselves, and “automated” for bots that do not. We took the monthly totals for English Wikipedia from January 2019, the earliest month we requested, to September 2026, and added them into twelve-month periods ending each September.
Billions of pageviews classed as “user”, all access methods
Source: Media Morpheus Research, from the Wikimedia Pageviews API, English Wikipedia, monthly aggregates, retrieved 1 October 2026.
Wikimedia reclassified March to August 2025 with improved bot detection; earlier months were classified under older rules, so some bot traffic that evaded detection before 2025 is likely still counted as human in the earlier bars.
Human pageviews in the twelve months to September 2026 were 80.4 billion. That is 7.5 per cent fewer than in the year to September 2025 and 13.1 per cent fewer than in the year to September 2024, which, at 92.5 billion, was the highest since the year that included the first months of the pandemic. Human pageviews have dropped before, in the year to September 2022, and recovered. This time they have fallen two years running.
Single months show the same thing. September 2026 recorded 6.27 billion human pageviews, the lowest month anywhere in the series. The first nine months of 2026 had 59.5 billion, against 64.6 billion in the same months of 2025 and 69.3 billion in 2024.
Billions of pageviews classed as “user” in the month
Source: Media Morpheus Research, from the Wikimedia Pageviews API, English Wikipedia, monthly aggregates, retrieved 1 October 2026.
September 2025 and September 2026 both fall after Wikimedia's 2025 change to bot detection, which makes them the most like-for-like pair in the series. September 2026 is complete: the daily series has all 30 days.
The September comparison is the one we would put most weight on. In October 2025 the Foundation explained that, around May 2025, it saw unusually high amounts of apparently human traffic, mostly from Brazil. It updated its bot detection and reclassified March to August 2025, finding that much of the traffic in May and June came from bots “built to evade detection”. It warned that the data must be interpreted “with care, as our bot detection systems apply different rules at different points in time.” September 2025 and September 2026 were both counted after that change. Between them, human pageviews fell 9.3 per cent.
The Foundation reached a similar conclusion from its own data at the time: after the revision, it saw human pageviews across all language versions of Wikipedia down “roughly 8%” on the same months of 2024. It said it believes the 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.” Our figures cover English Wikipedia only and run a year further.
How much of the traffic is bots
While human readers have fallen, non-human pageviews have not. In the twelve months to September 2026, English Wikipedia recorded 37.9 billion pageviews from bots and crawlers, 32 per cent of the total. In the year to September 2021, the first full year in which Wikimedia separated out automated traffic, the bot share was 21.5 per cent.
Share of all pageviews in each twelve-month period, by Wikimedia's agent class
Source: Media Morpheus Research, from the Wikimedia Pageviews API, English Wikipedia, monthly aggregates, retrieved 1 October 2026.
The “automated” class, for bots that do not identify themselves, begins in April 2020, so the first full twelve months that include it run from October 2020. Shares are of pageviews only; Wikimedia's infrastructure figures, such as API calls and media downloads, are separate and not shown.
The change is almost entirely in the automated class: bots that do not declare themselves, which Wikimedia identifies by behaviour. Self-identified crawlers, which include search engines, were a similar share at both ends of the period, though their volume swings widely from month to month and peaked in the year to September 2025. These are pageviews only. The Foundation's infrastructure figures are a separate measure and a starker one: in April 2025 it said bots made up about 35 per cent of pageviews but at least 65 per cent of its “most expensive” traffic, because crawlers bulk-read unpopular pages that are not cached, and that bandwidth for downloading media had grown 50 per cent since January 2024. In March 2026 it said it was blocking or throttling about 30 per cent of all automated requests, “up to billions of requests per day”.
Where the readers went missing
Splitting human pageviews by how they arrived shows where the latest fall happened.
Twelve months to September 2026 against the twelve months to September 2025
Source: Media Morpheus Research, from the Wikimedia Pageviews API, English Wikipedia, monthly aggregates, retrieved 1 October 2026.
Bar length shows the size of the change, not its direction; the value gives the direction. The desktop rise may partly reflect how traffic is classified rather than more people reading on computers.
Mobile web, which carries most of Wikipedia's human readership, lost 7.0 billion pageviews in a year, a fall of 12.3 per cent. Desktop human views rose slightly. That pattern fits the explanation the Foundation gives, since phone searches are where AI summaries and social video most often stand between a question and a click, but the data cannot confirm it. Classification may matter here too: if bots that once passed as mobile readers are now being caught, the mobile figure would fall for reasons unrelated to people. How AI summaries affect clicks on search results more generally is covered in do AI Overviews reduce clicks.
What it means for a company's Wikipedia article
It is tempting to read falling traffic as falling importance. For a company with a Wikipedia article, the opposite is closer to the truth.
The article's readers are increasingly machines and the people they answer. The Foundation says that “almost all large language models (LLMs) train on Wikipedia datasets” and that search engines and social platforms prioritise its information when answering questions. A company article read by fewer people on wikipedia.org may still shape what many more people are told about the company by an assistant, without a link back to the page or its sources. Why assistants lean on reference sources is discussed in which sources AI assistants cite.
Fewer readers can mean fewer eyes on errors. The Foundation's worry is that fewer visits mean fewer volunteers and donors. For a company the practical point is narrower: an inaccuracy on a lightly read article may sit longer before someone notices it, while being repeated elsewhere. Watching the article, rather than waiting to be told, matters more. Why a company should not fix it by paying an editor is in can you pay someone to edit your Wikipedia article.
The sources under the article carry the weight. An article is only as strong as the independent coverage it cites, and that coverage is what both editors and, indirectly, AI systems draw on. What company articles actually cite is in what sources Wikipedia company articles cite, and which outlets Wikipedia's editors trust in Wikipedia's perennial sources.
A company can measure its own article. The same API that produced the figures above returns daily human pageviews for any single article, and the Pageviews tool linked from every article's history page does the same without code. Setting an article's trend beside English Wikipedia's overall decline shows whether interest in the company is changing or the whole site is.
Methodology and limitations
The Wikimedia Foundation's UN General Assembly statement and its Diff posts of 1 April 2025, 17 October 2025 and 26 March 2026 were read in full on 1 October 2026. Pageview data come from the Wikimedia Pageviews REST API aggregate endpoint, wikimedia.org/api/rest_v1/metrics/pageviews/aggregate/en.wikipedia.org/{access}/{agent}/monthly/2019010100/2026100100, requested on 1 October 2026 with a descriptive User-Agent, for each combination of access (all-access, desktop, mobile-web, mobile-app) and agent (user, automated, spider). Every series returned months up to and including September 2026; the automated series begins in April 2020. We confirmed September 2026 was complete by requesting the daily all-access user series, which returned all 30 days. Twelve-month periods run October to September. Bot share is automated plus spider pageviews divided by the sum of all three agent classes; the five-year comparison uses October 2020 to September 2021 because it is the first full period with the automated class. Percentages were calculated from unrounded monthly totals. Five limitations apply. Wikimedia's classification of humans and bots has changed over time, most notably with the reclassification of March to August 2025, and earlier months were not revised, so comparisons across that change partly measure detection, not readers. Pageviews count page loads, not people. The data cover English Wikipedia only, not all language editions or Wikimedia's other projects. They cannot show why readership fell; the attribution to AI and social media is the Foundation's view, reported as such. And they say nothing about any individual company article; we did not sample company articles for this report.
Conclusion
The Wikimedia Foundation's case at the UN was that AI depends on Wikipedia while drawing people away from it. Wikimedia's own public data support the second half of that, with the necessary caveats. Human pageviews of English Wikipedia fell 7.5 per cent in the year to September 2026, after falling the year before; September 2026 was the lowest month since at least 2019; and the cleanest comparison, September against September after the Foundation's bot-detection update, shows a fall of 9.3 per cent. Meanwhile almost a third of counted pageviews now come from bots, most of them bots that do not say what they are. The data cannot prove that AI caused the fall, and some of it may be better bot detection. But for a company, the conclusion does not depend on the cause. Its Wikipedia article is being read less often by people and more often by the systems that answer people's questions, which makes the accuracy of the article, and the strength of the independent coverage beneath it, more important rather than less.
Frequently Asked Questions
Are fewer people reading Wikipedia?
Yes, on Wikimedia's own numbers. English Wikipedia recorded 80.4 billion human pageviews in the twelve months to September 2026, 7.5 per cent fewer than the year before and 13.1 per cent fewer than the year to September 2024. September 2026, at 6.27 billion, was the lowest month in the series we pulled, which starts in January 2019.
Is AI the cause?
The Wikimedia Foundation believes generative AI and social media are the main reasons, with search engines answering questions directly, often from Wikipedia content. The pageview data cannot prove cause; it shows the timing and size of the decline, and part of the fall from earlier years may reflect better bot detection rather than fewer readers.
How much of Wikipedia's traffic is bots?
In the twelve months to September 2026, 32 per cent of English Wikipedia's counted pageviews were classed as bots: 17.4 per cent as automated traffic that does not identify itself and 14.6 per cent as self-identified crawlers. Five years earlier the bot share was 21.5 per cent. The Foundation separately said in 2025 that bots account for at least 65 per cent of its most expensive traffic.
Did Wikimedia change how it counts human visitors?
Yes. In 2025 it found that much unusually high traffic in May and June, mostly from Brazil, came from bots built to evade detection, updated its detection and reclassified March to August 2025. It says the data must be interpreted with care because different rules apply at different times.
Does a falling pageview count make a company's Wikipedia article matter less?
Not necessarily. Fewer people read the article on wikipedia.org, but the Foundation says almost all large language models train on Wikipedia and search and AI products draw on it to answer questions. What the article says may reach more people through those products than through visits, and without a link back.
How can a company see its own article's traffic?
Wikimedia publishes daily and monthly human pageviews for every article through the same public API, and through the Pageviews tool linked from each article's history page. Comparing an article against all of English Wikipedia separates its own trend from the overall decline.
Sources
- official Wikimedia Foundation, Human knowledge at stake: Wikimedia Foundation makes its case at the United Nations General Assembly, 30 September 2026. Meehan's remarks on infrastructure, scraping and attribution. Read 1 October 2026.
- official Marshall Miller, Wikimedia Foundation, New User Trends on Wikipedia, Diff, 17 October 2025. Reclassification of March to August 2025; human pageviews down roughly 8 per cent; the Foundation's explanation; LLM training on Wikipedia. Read 1 October 2026.
- official Birgit Mueller, Chris Danis and Giuseppe Lavagetto, Wikimedia Foundation, How crawlers impact the operations of the Wikimedia projects, Diff, 1 April 2025, and Quo Vadis, Crawlers? Progress and what's next on safeguarding our infrastructure, Diff, 26 March 2026. Bots as 65 per cent of the most expensive traffic; 50 per cent growth in media bandwidth; about 30 per cent of automated requests blocked or throttled. Read 1 October 2026.
- live read Media Morpheus Research, from the Wikimedia Pageviews API, English Wikipedia, monthly aggregates by access method and agent type, January 2019 to September 2026. Retrieved 1 October 2026.