Wikipedia still draws about 508 million human page views a day. In a January 13, 2026 short read marking the project at 25, Pew Research Center put the previous decade at 1.9 trillion human views, about 49% of them on the English edition. These are page views (a page opened), not unique people, and not an all-time total. The count leaves the machines out on purpose.
The machines are a second audience, not a larger one. Pew says web crawlers, AI bots, and other nonhuman agents produced more than 88 billion views in 2025. Near that figure, the rate is roughly 240 million views a day, about half the human daily average Pew reports. The windows differ: one number is a decade of human views, the other a single year of automated ones. Together they are enough to retire the idea that bots already out-read people by a wide margin. The strain is the split. Human visits are slipping at the front door while automated readers keep taking the text.
The visit thinned after the logs were cleaned up
In an October 17, 2025 post, the Wikimedia Foundation said that after it revised how it detects bots, human page views were down by roughly 8% compared with the same months in 2024. The drop was not obvious in the raw logs. Around May 2025 the Foundation saw a spike that looked human, mostly from Brazil. Updated detection, applied back across March through August 2025, found that much of the May-June surge came from bots built to evade the old filters. Only after that cleanup did the 8% decline appear. Rules for what counts as a person have changed over time, so the series has to be read carefully. It is a real loss of visits. It is not an emptying building.
Wikimedia’s explanation is specific. Generative AI and social platforms answer or entertain before anyone clicks, and they often do it with Wikipedia’s own sentences. The Foundation says almost all large language models train on Wikipedia datasets, and that search and social products prioritize its information. Pew ties the same pattern to AI search summaries. Fewer direct visits, the Foundation warns, can mean fewer new volunteers and fewer individual donors, even while the models keep using the text.
Richard Cooke, whose book The Last Best Place on the Internet: A Human History of Wikipedia is a human history rather than a traffic study, noticed the exposure while he was finishing. He also says he was seeing editors add large amounts of weak machine-written text. That second claim is his observation. Neither Pew’s readership study nor the Foundation’s October post measures a flood of bad AI prose inside articles. What those sources measure is the outside problem: reuse without the visit.
Trust cuts the other way from the old sneer. The familiar line, the “free encyclopedia that anyone can edit, ” froze a lot of casual talk at “don’t trust it, ” which is how Cooke characterizes the reputation, even as the page became the first stop for a dog breed or a head of state. Citing its own Brand Health Tracker, the Foundation says Wikipedia remains highly trusted and valued as a neutral, accurate source. That is a survey run by the institution, not an outside audit. Both can be true. People mock the encyclopedia in conversation and still treat it as the source their chatbot quietly used.
Who pays once the summary box keeps the visit
For anyone shipping automated answers, this is not a media story about a beloved website. It is a bill that moved.
Picture a support or sales agent that quotes a help-center page, or a Wikipedia-grade explanation, and hands the customer a finished answer. The click never comes back. Nobody on the customer’s side sees who wrote the citation, who is wrong, or who should be paid to fix it. Accuracy can keep flowing out of a knowledge base long after the people who maintain it stop showing up.
Search used to return some of that attention as a visit, a correction, or a donation. A summary box does not. Wikimedia Enterprise is the Foundation’s attempt to charge the heaviest reusers, alongside an attribution framework and stricter reuse rules. Most companies do not have a donation banner and twenty-five years of volunteers behind the API.
The decision is narrower than a new policy deck. Name an owner for the source the agent quotes (correction, credit, and the cost of keeping it true) and fund that owner from the same budget that ships the answer. If the summary keeps the visit and nobody is paid to maintain what it summarizes, the product is borrowing a maintenance model it is also wearing down.
Bigger than a reference desk, and not all handmade
Jimmy Wales and Larry Sanger launched Wikipedia in 2001. Cooke reports that the co-founders once thought 100, 000 articles would be “a reasonable size, ” because that was roughly Encyclopaedia Britannica’s scale. As of December 2025, Pew puts the project at more than 66 million articles across 342 languages. English Wikipedia holds around 7 million articles and over 5 billion words. Pew estimates one person would need about 38 years to read every English article, and that only holds as nonstop reading. The files (text, images, video, and the rest) come to about 775 terabytes.
Do not read “66 million articles” as 66 million volunteer essays. Pew notes that bots create pages in some languages. Lsjbot, built by linguist Sverker Johansson, creates Cebuano and Swedish articles. Cebuano ranks second in article count with relatively few active users. Size and human labor are no longer the same statistic.
What people open is narrower than the mythology. In Pew’s decade-long look at English Wikipedia, Donald Trump’s article drew nearly 300 million views, the most of any English page in that analysis. ChatGPT also sits near the top, which is a small irony: readers have been using the encyclopedia to understand the systems now answering in its place. “Deaths in [year]” pages for 2015 through 2025 totaled nearly half a billion views. Since 2015, Pew reports, 46% of the articles on each month’s most-viewed English list were about individuals, 31% about movies or TV, and 14% about events such as the FIFA World Cup. Outside the spikes, Trump’s page still averages just under 45, 000 views a day. The article on dogs is Pew’s reminder that a lot of the traffic is someone settling an ordinary argument.
Reach is uneven in another way. Mandarin and Hindi are the world’s second- and third-most-spoken languages, yet those Wikipedias rank 8th and 25th in all-time page views, Pew reports. Wikipedia has been blocked in mainland China since April 2019. Baidu Baike is the local analogue people there can actually open.
Why this is hard to hire
The visit numbers are the urgent part. A replacement is hard to hire for an older reason, and Cooke’s history is the short turn worth taking: this was not a content strategy that learned to scale.
Ward Cunningham invented the wiki in the 1990s, a webpage anyone could change, with a tolerance for incompleteness that Cooke traces to constructivism and to wabi-sabi, the Zen sense that beauty includes the unfinished. Wales came from tech and finance and had spent years writing to philosophers. Sanger was an academic philosopher who treated online argument as serious work. Cooke’s phrase for the culture they drew on is a “republic of letters”: people fighting over facts in public. There are still no paid fact-checkers. Donors, he writes, sometimes assume the money is “for the factcheckers.” It is not. The checking is the volunteers, the citations, and the arguments on the talk pages.
The early editors he recovers are specific and unglamorous. Gaytha Langlois, whom he identifies as the first woman to edit the site, wrote about Rhode Island and standard poodles. Ruth Ifcher, in his account, had the first biographical page for a woman (what would now be called a user page). She had a PhD, a fierce sense of grammar, and had left the workforce because of severe agoraphobia. Obsessive, difficult, often homebound volunteers built a public utility. The institution has not always been gracious about them.
The standing fight is what belongs in the record. Cooke describes inclusionists and deletionists (keep the minor page, or cut it) as the project’s deepest rift. Enthusiasts he calls “road geeks” left after work on minor roads was deleted and started their own encyclopedia at wiki.aaroads.com. Too much writing on a favorite subject even has a name in his telling: fancruft. Gender imbalance, he says, is longstanding and stubborn. Sandi Toksvig is, in his account, the latest prominent critic, involved in a more balanced alternative called Mappa Mundi, which he thinks could also feed material back into Wikipedia. He does not claim it has. The site became usable by arguing about what counts, not by being polite. That argument is the maintenance model. It does not transfer cleanly into a vendor contract.
People ask Cooke whether Wikipedia is under threat. He thinks yes, and he also thinks it will survive. The line he wrote the book to earn:
We saw this made, and we won’t see anything like it again.
That is an opinion, not a forecast. An 8% drop in human page views does not prove it. Survival and irreplaceability are different claims.
The Foundation, in that October post, is spending on reuse enforcement, attribution, and enterprise access, with more attention on mobile editing and a newcomer’s first edit. It is also testing video, games, and chatbots on platforms where many readers never clicked through in the first place. The post ends confident that Wikipedia will keep supplying free, accurate, human knowledge. Confidence is not a reversed trend.
Questions worth settling before the next strategy deck
- Is Wikipedia losing readers to AI? Human page views were roughly 8% lower than the same months in 2024, after Wikimedia reclassified evasive bot traffic, as the Foundation reported on October 17, 2025. It blames generative AI and social platforms that answer before the click, often with Wikipedia’s text. That is a material loss of visits, not an emptying site.
- Are the machines already the bigger audience? Not on the figures Pew published. Humans account for about 508 million page views a day across the last decade. Crawlers, AI bots, and other nonhuman agents produced more than 88 billion views in 2025, roughly 240 million a day near that total, and those are different windows. Machine demand is enormous. It has not, on that count, overtaken people.
- How big is it, if you refuse the round numbers? Pew’s December 2025 snapshot: more than 66 million articles in 342 languages, about 7 million of them in English, and over 5 billion English words. Some of the article count is bot-generated. The verified readership figure is 1.9 trillion human views in a decade, not an all-time total, and not a count of people.
- Do people trust it or not? Casual dismissal is what Cooke describes, stuck to the phrase “anyone can edit.” Wikimedia’s own Brand Health Tracker says the site stays highly trusted as a neutral, accurate source. One is street reputation. The other is the Foundation surveying its own standing. Neither cancels the other.
- Are editors flooding articles with bad AI text? Cooke says he has seen large amounts of poorly written machine text added by editors. Pew’s traffic research and the Foundation’s October post do not measure that. What they measure is models and answer engines consuming the encyclopedia while sending fewer humans back.
- Will something like this get built again? Cooke doubts it. Wikimedia expects the project to endure. Those are judgments. The quantified part is the visit, the scale, and the machine readership, not the forecast.
- What should a company decide? If an automated answer sits between the customer and the source, assume the click is gone. Name who pays to correct, credit, and maintain that source, and fund them from the same budget that ships the answer. Otherwise you are keeping the answer and discarding the upkeep.
The site can still look, as Cooke puts it, much as it did twenty years ago. That plainness hides a fight about what belongs in the record, a donor base that sometimes thinks it is paying fact-checkers, and a layer of automated readers happy to recite the record without visiting. The last best place on the internet was never a product launch. It was a habit of argument. Habits do not survive on nostalgia, and they do not survive on being summarized for free.