The first edit on Wikipedia is dated 15 January 2001 1 M. After 25 years it has 68.6m articles in 348 active language editions 2 H. The record for human pageviews across all Wikimedia projects came in the pandemic year of 2020: 200.3bn. In the first nine months of 2026 there were 120bn: 9.6% fewer than a year earlier and 15.3% fewer than in 2024 3 H. In October 2026 the decline accelerated: −12.8% in the first 25 days against the same days of 2025 4 H.
Knowledge from Wikipedia has not become less needed: search engines and chatbots retell it. What has become rarer is visits to the site itself. Since May 2024, when Google switched on AI answers above search results for all users in the US, every month’s human pageviews have been lower than in the same month a year earlier 5 H. At the same time, so many machines now read Wikipedia that they account for almost half of all pageviews, and some of them disguise themselves as people.
Wikipedia has not lost its significance — it is losing visitors. Its knowledge increasingly reaches people through an AI retelling rather than through the page itself, even though it is visitors who bring Wikipedia donations and new editors.
In brief
- Scale. In 2025 all Wikimedia projects had 175.6bn human pageviews, 6.5% fewer than in 2024; English Wikipedia had 85.5bn, almost half 3 H. By our estimate, 2026 will end at about 159bn L.
- Bots. In the first nine months of 2026 bots generated 46.4% of all pageviews; on desktop the Foundation classes only 35% of requests as human. If desktop users are falling away at the same rate as on mobile web, the true fall in human pageviews is about −17% a year, not −9.6% L.
- The main suspect is search that answers by itself. According to Pew for the US (March 2025), when there is a Google AI answer people click a link in 8% of visits, and when there is none, in 15% 6 H; 78% of Wikipedia reading sessions begin in search engines 7 H. A study with a control group (preprint) puts the direct effect of AI Overviews for English Wikipedia at only −5.45% of referrals from search, so other factors explain the rest of the decline 8 M.
- Where the losses are. Pageviews from mobile web are down 17.6% on the year. Articles on mathematics and physics have kept a third of their 2016–2019 pageviews, biographies 87% 9 M. The Spanish and Portuguese Wikipedias have about half of their 2019 pageviews, English 87% 10 H.
- Editors. English Wikipedia has about 30,000 active editors a month — the same as in 2013; occasional contributors are down 21% since 2019 11 H. Since the March ban on AI text, English Wikipedia has been losing editors more slowly than most large language editions (−1% a year in April–September), so the hypothesis that the ban drives editors away is so far not confirmed L.
- Money. The Foundation’s revenue in fiscal year 2024–25 rose 12.5% to $208.6m, but donors fell by 2.4% and new donors by 10.2% 12,13 H. Paid access to data for big technology companies (Wikimedia Enterprise) brings in only 4% of revenue 14 H.
- 2030 and 2035. There is no official forecast. In our base scenario (−8% a year to 2030, then −4%) there are 114bn human pageviews in 2030 and 93bn in 2035, 54% fewer than at the 2020 peak; the range of scenarios for 2035 is from 49bn to 134bn L.
01 · ScaleWhat the Wikimedia counters show: 175.6bn human pageviews in 2025 and a decline since May 2024
The Wikimedia Foundation counts pageviews, not people. A pageview is a request for a page that received a response from the server. Requests fall into three classes: human (user), honest bots (spider) and bots that pass themselves off as humans (automated) 15 H. In this briefing “human pageviews” means the first class. “All projects” means all the language editions of Wikipedia together with Wikimedia Commons, Wikidata, Wiktionary and the other projects; they account for only a few per cent of human pageviews 3 H. A monthly series in the current definition exists from July 2015 16 H.
Until 2024 the picture was stable: 183–200bn human pageviews a year (2021–22 are understated because of a logging failure), with a sharp jump in April 2020, when people were locked down at home 3,17 H. English Wikipedia accounted for 46–50% of all pageviews throughout 18 H. By the Pew Research Center’s count, from December 2015 to December 2025 Wikipedia had about 1.9trn human pageviews — an average of 508m a day; English received 49% of them, Japanese 7%, and Spanish, German and Russian 6% each 19 M. Since May 2024 every month has been lower than the same month a year earlier. In 2025 there were 6.5% fewer human pageviews than in 2024. In January–September 2026 there were another 9.6% fewer for all projects and 7.9% fewer for English Wikipedia. September 2026 (6.27bn) was English Wikipedia’s weakest month in the whole series since 2015 18 H.
October 2026 shows an acceleration. In the first 25 days there were 10.1bn human pageviews across all projects against 11.6bn in the same days of 2025 — 12.8% fewer; on English Wikipedia, 14.0% fewer 4 H. After the bot recalculation in October 2025, the Wikimedia Foundation described the fall as “about 8%” 20 H; the new data show that it has not slowed since.
The overall flow of requests to pages is not falling, though. Adding humans and bots together, in 2016 the Wikimedia projects received 239bn pageviews and in 2025 302bn, and 2026 is running at roughly the same level 3,21,22 H. The legacy counter of 2008–2015 also counted all requests, bots included: 121bn in 2008 and 235bn in 2015 23 H. The gap between the lines on the chart is machines.
02 · History2001–2014: from a wiki on a few servers to the first result in Google
There is no monthly series of pageviews for 2001 to mid-2007: the Foundation did not then compile server logs 24 H. Growth can be seen from indirect signs. In 2004 the load on the servers grew by an average of 90% each quarter, and the site was kept running by three cache servers, seven web servers and two database servers 25 H. In October 2006 about 250 servers in Florida, Amsterdam and Seoul handled up to 17,500 requests a second 26 H. By Alexa ranking the site rose from about 10,000th place in 2002 to the world’s top ten: on the three-month ranking, in March 2007 27 M. According to comScore, in January 2007 Wikipedia entered the ten most visited sites in the US for the first time: 42.9m unique visitors a month 28 H.
The first direct measurement appeared in June 2007. The developer Brion Vibber estimated from a sample of logs 233.6m article pageviews a day across all projects, or about 7bn a month. Only 23.8% of pageviews came from Google, and another 43.8% through links inside Wikipedia itself 29 H. The Foundation now estimates that 78% of reading sessions begin in search engines. That is a different metric from Google’s share of pageviews in 2007, but it seems that the dependence on search which makes Wikipedia vulnerable today grew later L.
The Foundation’s old counters counted page requests together with bots, so they are not comparable with today’s “human” pageviews. By them, in May 2009 all the Wikipedias had 11.7bn pageviews (English 6.3bn) 30 H, in fiscal year 2010–11 an average of 13.6bn a month 31 H, and a Foundation post of 19 April 2013 cited 21.3bn Wikipedia page requests a month (without saying for which month) 32 M.
Articles grew in number, but editors did not. English Wikipedia had the most active editors in March 2007; after that there were fewer of them, even though readers kept increasing for several more years 11 H. A 2013 study by Halfaker et al. offered an explanation: between the first halves of 2006 and 2007, the share of good-faith newcomers whose first edits were reverted rose from 6.1% to 18.2%, and by mid-2008, according to data the authors cite from a 2012 paper by Geiger et al., 75% of newcomers received their first message from an automated tool rather than from a human 33,34 H. According to a 2011 Foundation study, the share of newcomers who were still editing a year after their first edit fell sharply between mid-2005 and early 2007 and never recovered 35 M. The goal of 200,000 active editors across all the Wikipedias by 2015 was not met even by half 36 H.
The first wave of fear that “Google is taking the readers” came in 2012–2014. On 16 May 2012 Google launched Knowledge Graph — panels of facts directly in the results, partly taken from Wikipedia 37 H. Desktop (full-version) pageviews of all the Wikipedias by the legacy counter, which also counted bots, peaked in January 2013 at 18.3bn a month and in two years fell by 30% 38 H. The press repeated the theory that Knowledge Graph was to blame, while the Foundation suspected counting errors 39 M. A significant part of the fall was probably explained by the move to smartphones L: at the start of 2014 mobile accounted for only about 20% of reading on English Wikipedia, and in June 2017 mobile pageviews outnumbered desktop ones for the first time 40,41 M. A Foundation analyst in 2015 found no fall in referrals from Google 42 M. Back then Wikipedia lived through a change in the way people read. Now the very need to visit the site is changing.
03 · Events2015–2026: the mobile era, bots, AI answers and the first rules against AI text
A new stage began in 2015 with a change in counting: the Foundation stopped counting search-engine crawlers as readers — almost 20% of the old “pageviews” 43 H. After that, events run along several parallel lines: mobile devices displace desktop, bots disguise themselves ever better, Google and chatbots start to answer for themselves, and the community of editors tightens its rules against AI text. The filter above the timeline shows each line separately.
04 · BotsWho reads Wikipedia in 2026: almost half of requests come from software
The overall flow of requests to Wikipedia pages is not falling. In 2025 all Wikimedia projects received about 302bn pageviews including bots; in 2016 the figure was 239bn. Only the human part is falling. In 2016 bots accounted for 18.9% of all pageviews, in 2024 for 36.4%, and in January–September 2026 for 46.4% 3,21,22 H. Automated pageviews in the first nine months of 2026 (52bn) already exceed those in the whole of 2025 (50.7bn) 22 H.
Wikimedia divides bots into two classes. “Spiders” are programs that honestly identify themselves, such as Google’s search crawler. “Automated” traffic comes from programs that pass themselves off as people. The Foundation began to separate this class only in May 2020; before that such requests were counted as human. The bot share for 2016–2019 on the chart is therefore understated, and human pageviews in those years were in fact somewhat lower 44 H.
For the Foundation this is not only a statistical problem but a cost one. In April 2025 it reported that bots generate about 35% of pageviews (under the classification used before the October 2025 recalculation) but at least 65% of the most expensive requests: those that are not served from the cache and reach the core data centres. The explanation is simple. People read popular articles, which already sit in the cache, while scrapers collecting data to train AI work through the whole “long tail” of pages. Traffic for downloading images and video has risen by 50% since January 2024, mainly because of scrapers 45 H.
In March 2026 the Foundation began to introduce global rate limits on requests to its API and said it was blocking or slowing about 30% of requests from scrapers that break the rules; newer scrapers disguise themselves as browsers and hide behind networks of residential proxies 46 H. On 5 October 2026 the Foundation said that agents which, in its view, belong to OpenAI had made millions of automated requests and passed through millions of pages, mostly on Wikidata and Wikimedia Commons 47 H.
The ratio of how much a machine reads to how many people it sends back is visible in Cloudflare’s data for the internet as a whole. In July 2025, for every referral a site received, there were about 38,000 page loads by Anthropic’s scrapers, about 1,100 by OpenAI’s, 195 by Perplexity’s and 5.4 by Google’s 48 H. Cloudflare does not publish separate data for Wikipedia, but the logic is the same: a model that has already read an article rarely sends a person back to read it again.
How many bots are hidden in “human” pageviews
A “human” pageview in the Foundation’s statistics is everything that could not be identified as a bot. The best-disguised bots are programs that arrive from desktop browsers. In 2016 the Foundation counted 72% of desktop requests to all projects as human (then still without a separate automated class); in January–September 2026 it counted only 35%, and 48% on English Wikipedia. On mobile web the human share is much higher: 78% 3,21,22 H. Two-thirds of desktop traffic, then, is already bots, and that is where those not yet caught are most likely to be hiding.
That bots still leak into the “human” series is clear from the Foundation’s own corrections. In spring 2025 programs, mostly from Brazil, passed themselves off as people; after the recalculation, human pageviews of all projects for May 2025 fell from 17.2bn to 14.5bn, a drop of 15.5% 49 H. In November 2025 a new wave of bots with spoofed browsers, from Vietnam, the US, the Philippines and India, which looked like desktop users in the statistics, accounted at its peak for about 10% of global pageviews; November could not be fixed, and December and the following months were recalculated only at the end of April, in May 2026 50 H. A surge of bots on Wikidata in July 2026 was fixed only on 1 October, without recalculating past data 51 H. The Foundation plans to move bot detection, including for public statistics, to a new system, KAPOW; as of 15 October 2026 the relevant task sat in the “ready to launch” column on the team’s board, although it formally remained open 52,53 M. Human pageviews will therefore probably be recalculated again soon, most likely downwards L. One more possible sign (the data do not distinguish residual bots from real growth): in March 2025 Brazil rose from 12th–15th to 7th place among countries by pageviews of English Wikipedia, and it has stayed in the top ten ever since 54 H.
The most convincing indirect sign is the divergence between devices. In January–September 2026 “human” pageviews from mobile web were 17.6% lower than a year earlier, from the app 8.5% lower, and from desktop 6.6% higher 3 H. Nothing confirms a mass return of people to desktop computers. If we assume that people on desktop are dropping off as fast as on mobile web, about 10.4bn “human” desktop pageviews in the nine months of 2026 (23% of them) are undetected bots. The true fall in human pageviews of all projects is then about −17% a year, not −9.6%; for English Wikipedia it is about −13.5% instead of −7.9% L. The Foundation itself, in its monthly report for July 2026, writes that the overall figure (−3.6% y/y) substantially understates the decline, and cites mobile web (−16%) as less distorted by bots. Unique devices from May 2025 it left out of that report altogether, because bots distort them too much 55,56 H.
The official series, then, most likely understates the decline. But the comparison base is flawed as well: before May 2020 “human” pageviews contained undetected bots: in May 2020 the new automated class took 6.4% of all pageviews of English Wikipedia 57 H, while the share for 2016–2019 is unknown L; and in 2024 bots were probably also hiding in the “human” series. In this briefing we show the official numbers and, alongside them, a range.
05 · AIThe effect of AI: people have not stopped looking for knowledge – they just click through to the page less often
A popular explanation runs: ChatGPT appeared, and Wikipedia’s traffic fell. The dates do not match. ChatGPT was launched on 30 November 2022. In 2023 human pageviews of all Wikimedia projects totalled 190.2bn, more than in 2022 (183.4bn) 3 H. A comparison of 12 language editions found no fall in pageviews in the first months after ChatGPT’s launch, although growth slowed where it was available 58 M.
The sustained decline began in May 2024. That month Google switched on AI Overviews – short AI-generated answers above the search results – for all users in the US 59 H. Since then every month’s human pageviews of all projects have been lower than in the same month a year earlier, by between 1% and 18% 5 H. In time, then, the decline coincides not with the launch of chatbots but with the arrival of AI answers in search. This is the most likely main factor, but it is not proven: the only estimate with a control group (a preprint) puts the loss at −5.45% of referrals from search to English Wikipedia.
The mechanism: an answer instead of a link
In March 2025 Pew Research Center tracked 68,879 searches by 900 US adults. When Google showed an AI answer, people clicked a regular link in 8% of cases; when it did not, in 15%. Links inside the AI answer itself were clicked in about 1% of cases 6 H. Wikipedia, YouTube and Reddit were the most frequently cited sources both in AI answers and in standard results 60 H. Google, in other words, shows Wikipedia’s text, but the person does not get as far as Wikipedia.
For Wikipedia this is a weak spot. By the Foundation’s estimate, 78% of reading sessions begin in search engines, and about 90% of those in Google 7 H. According to Alphabet, in summer 2025 AI Overviews were seen by more than 2bn users a month in more than 200 countries 61 M, and AI Mode, Google’s conversational search, passed 1bn users a month in July 2026 62 M.
Calculator: how many pageviews AI answers take
The model is simple: the share of pageviews that come from search × the share of queries with an AI answer × how much less people click when an answer is present. To this is added the share of queries that people put to a chatbot instead of searching. The defaults are: 18% of searches with an AI answer and a 47% drop in clicks (8% instead of 15%), from Pew data for March 2025 6 H; 90%, the “historical” share of visitors from Google under the Foundation’s annual plan (by the Foundation’s estimate, measured by sessions, it is about 70%, so 90% is an upper bound) 63 H; and 10% of queries shifted to chatbots, which is our assumption L. The result is compared with the actual fall in human pageviews of all projects in January–September 2026 against the same months of 2024.
One of the first studies with a control group specifically on AI Overviews and Wikipedia is by Khosravi and Yoganarasimhan (University of Washington, preprint, not peer-reviewed). They compared English Wikipedia with German and French, where AI answers arrived later. Their estimate: AI Overviews reduced referrals from search to English Wikipedia by 5.45% relative to German and by 4.82% relative to French 8 M. The first version of the paper gave about 15%, so the result depends on the method. Another paper (Lyu et al., WWW 2025) found that new popular articles similar in content to ChatGPT’s answers lost more pageviews after its launch than dissimilar ones 64 H.
| Study | Method | Period | Result | Does it establish causation? |
|---|---|---|---|---|
| Reeves, Yin, Simperl, 2024 58 | Comparison of 12 language editions before and after ChatGPT’s launch | 2022–2023 | No fall in pageviews found; slower growth where ChatGPT is available | No |
| Lyu et al., WWW 2025 64 | Difference-in-differences: articles similar and dissimilar to ChatGPT’s answers | 2022–2023 | Similar articles lost more pageviews | Partly |
| Khosravi, Yoganarasimhan, 2026, preprint 8 | Difference-in-differences: English versus German and French Wikipedia | December 2023–December 2024 | AI Overviews: −5.45% of referrals from search | Yes, but not peer-reviewed |
| Luis Villa, 2026, volunteer analysis 65 | 5,000 articles, 8 languages, compared with 2016–2019 | April 2025–March 2026 | −26% overall; mathematics −67%, biographies −13% | No |
| Wikimedia Foundation, October 2025 20 | Recalculation of traffic after hidden bots were found | Several months of 2025 (not named) | About 8% fewer human pageviews than a year earlier | No, it is a description |
Where exactly the readers went
The cleanest measure is mobile web, where bots disguise themselves worst (see section 04). Human pageviews from mobile web in January–September 2026 were 17.6% lower than a year earlier and 23.8% lower than in 2024 3 H.
Language editions are losing pageviews unevenly. In January–September 2026 the Spanish and Portuguese Wikipedias had about half the human pageviews of the same period of 2019, Indonesian 43%, Arabic 59% and German 73%. English kept 87%, Ukrainian 98% and Chinese 98.5% 10 H. Among the large editions English holds up best, if Chinese and Ukrainian, where separate factors are at work, are set aside; in Villa’s volunteer analysis it lost 15%, against 26% on average 65 M. One possible explanation is freshness: among English articles, the most recently updated 10% lost 7% of pageviews, while the 10% that had gone longest without updates lost 34% 66 M.
The biggest losses are in articles on topics where a chatbot is a good substitute for the page; the author of the analysis himself cautions against categorical conclusions. In Villa’s sample of 5,000 vital articles in eight languages, mathematics retained only a third of its 2016–2019 pageviews, the physical sciences 34% and technology 41%. Biographies kept 87% 9 M.
Chatbots: a large audience, few referrals
ChatGPT had 900m weekly users in February 2026 67 M, and the Gemini app 950m monthly users in July 2026 62 M. By OpenAI’s own analysis, the share of “seeking information” messages rose from 14% to 24% between July 2024 and July 2025 68 H. But news surveys show that only a minority has changed its habits: in 2026, 10% of respondents in 48 countries used chatbots for news every week, against 7% a year earlier 69 H.
AI learns from Wikipedia: in GPT-3’s training mix English Wikipedia had a weight of 3% 70 H, and in LLaMA 4.5% (Wikipedia in 20 languages) 71 H. But ChatGPT cites it less and less; in Google AI Mode and Perplexity the share, according to Semrush, has held steady. According to Semrush, the share of ChatGPT search answers linking to Wikipedia fell over the study’s 13 weeks (14 July–12 October 2025) from about 55% to under 20%, mostly in mid-September 72 M. According to Ahrefs, in the US from March to September 2026 the share of ChatGPT answers linking to Wikipedia fell from 14.1% to 2.9% on the same queries 73 M. Referrals from chatbots are small in any case: Google sends news publishers about 500 times more visitors than ChatGPT 74 M.
What full replacement looks like: Stack Overflow
There is a site where AI replacement has been almost complete. On Stack Overflow, a question-and-answer site for programmers, 2.18m questions were asked in 2016, 1.34m in 2022 and 109,000 in 2025. In July 2026 there were 1,442, against 207,204 in March 2014 75 M. A peer-reviewed study with a control group estimated that by the end of the sixth month after ChatGPT’s launch, activity on Stack Overflow was about 25% lower than on comparable sites where ChatGPT was unavailable or less useful (averaged over the six months, about 15%) 76 H. The homework service Chegg had revenue of $51.8m in the second quarter of 2026, against $105.1m a year earlier, a fall of 51%; the company attributes this to Google’s AI answers 77 H.
Wikipedia is not Stack Overflow. There, AI replaced the question itself: a person gets an answer from a chatbot and does not post a question on the site. On Wikipedia, quick look-ups are combined with reading that an AI summary does not replace: biographies, history, geography. So in two years Wikipedia lost about 15% of human pageviews, while Stack Overflow in three years lost 92% of new questions. These are different measures (reading and contribution), so the comparison only illustrates scale.
Separate “AI encyclopaedias” are not yet taking Wikipedia’s traffic. Grokipedia, from Elon Musk’s xAI, launched on 27 October 2025 with more than 800,000 articles, a large share of them adapted from Wikipedia 78 M. In March 2026 Ahrefs estimated its organic traffic at ~1.3m pageviews a month, against ~2.1bn for Wikipedia 79 M.
06 · EditorsThe AI-text ban and editors: the core holds, newcomers are fewer, and the decline began before the ban
English Wikipedia’s rules on AI text became strict only recently. The first draft of rules on large language models appeared on 12 December 2022. A discussion that ended on 14 October 2023 declined to approve it as a guideline or policy: 30 participants were against, 8 in favour. Some found the rules too lenient, others too strict, and many held that the existing sourcing rules were enough 80 H. The rules then tightened step by step.
| Date | What changed on English Wikipedia |
|---|---|
| 4 December 2023 | WikiProject AI Cleanup created: a group of volunteers who look for and remove AI text; by October 2026 it had 319 members 81 H |
| 12 January 2025 | Administrators allowed to disregard, strike or collapse comments in discussions that are obviously AI-generated 82 H |
| 4 August 2025 | Criterion G15: an administrator can speedily delete a page without discussion if it is obviously AI-generated without human review (addressing the user, invented sources, technical traces of the model) 83 H |
| 24 November 2025 | NEWLLM guideline: do not create new articles from scratch with language models 84 H |
| 20 March 2026 | Guideline (not a policy) “Writing articles using LLMs”: generating or rewriting article text with language models is “strictly prohibited”; the closer of the discussion counted 44 votes in favour and 2 against 85 H. The exceptions as of October 2026 are grammar fixes to one’s own text and translation from other Wikipedias with human verification; in July 2026 it was clarified that other people’s articles may not be edited this way. Using AI to find sources is allowed 86 H |
| 2 May 2026 | Persistent use of language models in breach of the rules added to the grounds for blocking (131 participants in the discussion) 87 H |
| 24 September 2026 | Paste check switched on: an editor who pastes text from ChatGPT, Gemini, Claude, Copilot or DeepSeek sees a warning 88 H |
Other language editions followed the same path. German Wikipedia banned AI text in articles after a vote on 1–15 February 2026; by a preliminary count, about a third of the votes were against 89 M. Ukrainian Wikipedia translated the English guideline on 30 March 2026 and approved it on 6 May 90 H. According to a Meta-Wiki table, at least 25 language editions of Wikipedia had rules against AI text with a known adoption date as of October 2026 91 M.
The main reason for the rules is the load on volunteers. An article can be generated in a minute; checking every source takes hours. On 26 October 2026, 8,114 English Wikipedia articles carried the “suspected AI text” tag, 1,221 of them tagged in October alone; 30,252 articles were waiting in the new-page review queue 92,93 H. In a Princeton University study, detectors flagged 4.4–5.4% of new English articles from August 2024 as likely AI-generated; this is the detectors’ estimate, not a confirmed fact 94 H.
What the editor data show
English Wikipedia had its largest number of active editors (5+ article edits a month) in March 2007: 53,095. By 2013 the figure was about 30,000, and it has barely changed since: 30,740 a month on average in 2013 and 30,453 in January–September 2026 11 H. In September 2026 it was 28,773, 3.6% lower than a year earlier and the lowest for a September since 2018. The number of very active editors (100+ edits a month) has even grown: 3,069 in 2014, 4,180 in 2025 11 H.
A different layer is thinning: occasional contributors with 1–4 edits a month, from whom regular editors usually grow. In 2019 there were 77,455 a month; in January–September 2026 there were 61,509, 21% fewer 11 H. The decline began in 2021, long before any rules on AI. Registered users of English Wikipedia created 515,000 new articles in 2007, 246,000 in 2021 and 202,000 in 2025 95 H.
Testing the hypothesis: does the ban drive editors away
If the ban drove editors out, English Wikipedia would have lost them faster after March 2026 than before, and faster than other editions. The data do not show this. In April–September 2026 English Wikipedia had only 1.0% fewer active editors than a year earlier; in January–March it had 1.9% fewer. On German Wikipedia, which adopted its ban in February, the figures are 9.5% and 10.4% fewer. On Ukrainian Wikipedia they are 12.0% and 17.2% fewer, and the decline began before the rules were approved in May 11 H. English, with some of the strictest rules, is losing editors more slowly than ten of the twelve other editions; only the Chinese and Swedish editions are losing them more slowly (−0.7%). This is a coincidence in time, not proof that there is no effect.
This does not mean the rules are safe for newcomers. WikiProject AI Cleanup’s own guide acknowledges that AI text is often used by bewildered newcomers or by people for whom English is not a native language, and that after being accused they often quit editing 96 H. The signs of AI text are unreliable: the page “Signs of AI writing” warns that they are not proof 97 H. There is a historical precedent: a 2013 study by Halfaker et al. showed that after 2007 good-faith newcomers were increasingly reverted by automated tools and strict rules, which, in the authors’ view, explains the start of the editor decline 33 H. How common AI text is among newcomers can be partly seen from Wiki Education data: among 6,357 new editors in its education programme in autumn 2025, automated alerts about AI text in articles were triggered for 5% of participants (25% had been expected) 98 H. This figure measures how widespread AI text is, not the effect of the rules.
The Foundation’s monthly report for July 2026 gives a similar picture. Active editors across all projects combined are 0.3% more numerous than a year earlier, but only thanks to Wikidata (+31%), where the Foundation links the growth to promotional editing aimed at visibility in AI answers; on Wikipedias with a developed community there are 4.3% fewer active editors 99 H. Newcomers who stay on after their first edits were 11% fewer on those Wikipedias in July (6.6% fewer on a three-month average): no change on English, −8% on Italian, −27% on Russian 100 H. So the problem is the inflow and retention of newcomers, and it is not concentrated on English Wikipedia with its strict rules.
A bigger threat to editors is the fall in the number of readers. Editors come from readers. The Wikimedia Foundation itself writes that fewer visits may mean fewer new volunteers 20 H, and its plan for fiscal year 2026–27 sets a goal of raising the number of retained editors by 20% 101 H.
Verdict on the hypothesis
- Confirmed: the rules really are strict and are tightening, from speedy deletion (August 2025) to a ban on AI text (March 2026) and blocks (May 2026).
- Partly confirmed: newcomers and occasional editors are fewer, but the decline began in 2021, long before the rules.
- Not confirmed: a mass exodus caused by the ban. English Wikipedia is losing editors more slowly than most large editions (−1.0% in April–September 2026), although in September the figure was −3.6%. The rules are seven months old, so there is no final answer yet L.
07 · MoneyThe Foundation’s money: revenue is rising even as fewer people visit the site
The Wikimedia Foundation is not getting poorer yet. In fiscal year 2024–25 (July–June) its audited revenue was $208.6m, 12.5% more than a year earlier; expenses were $190.9m 12 H. According to tax filings, the Foundation’s revenue for the fiscal year that ended in 2011 was $27.2m, and for 2024–25 it was $205.4m 102 M. Net assets at 30 June 2025 were $296.6m; by the Foundation’s calculation, that is 17.1 months of expenses (it divides by the planned expenses of fiscal year 2025–26) 103 H. The separate Wikimedia Endowment held $169.4m 104 H.
But the same filings already show what is happening to readers. The number of donors fell by 2.4% over the year, to 7.7m, and the number of new donors by 10.2%, to 2.3m. Revenue grew because the average gift rose to $10.15 and regular monthly donations grew by 19% 13 H. Donation banners brought in $62.4m, or 30% of total revenue 105 H. A banner is seen only by a person on the page; an AI answer in search does not show it.
The second source is selling data to large companies through Wikimedia Enterprise. The first paying customer was Google, from June 2022 106 H. On 15 January 2026, Wikipedia’s 25th anniversary, the Foundation named Amazon, Meta, Microsoft, Mistral AI and Perplexity as partners 107 H. The terms of the deals have not been disclosed. Enterprise revenue for 2024–25 was $8.3m, 2.5 times more than a year earlier, but that is only 4% of the Foundation’s revenue 14 H.
For fiscal year 2026–27 the Foundation plans a budget of $240m, 15% more than a year earlier, and a donations target of $230m 108 H. In the same plan it states plainly that traffic, new editors and donors fell last year and are expected to keep falling 109 H, and that the fall in pageviews and referrals from Google is “not a temporary phenomenon but a structural shift” 110 H. Money lags attention: today each donor gives more, tomorrow fewer people will see the banner.
08 · UkraineUkrainian Wikipedia: pageviews doubled by 2023 and are falling faster in 2026
Ukrainian Wikipedia is an exception to the general picture of 2016–2023. It had 531m human pageviews in 2016 and 1.06bn in 2022, twice as many 111 H. The monthly peak, 115.9m, came in March 2023. Some readers moved over from Russian Wikipedia: according to Wikimedia data, in September 2021 readers from Ukraine made up to 113m pageviews a month on Russian Wikipedia, and in September 2026 up to 39m (the API gives only a range and an upper bound) 112 M.
In 2025 Ukrainian Wikipedia had 824.5m human pageviews, and in January–September 2026 it had 24.3% fewer than a year earlier. That is three times faster than English (−7.9%). The 12-month average pageviews fell from 101m a month (February 2024) to 56m (September 2026), by 45% 111 H. Part of the 2023–24 peak may have come from bots that were not yet detected at the time; this cannot be checked with open data. Compared with 2019, Ukrainian Wikipedia still stands at 98% (January–September against January–September; 95% on the 12-month average), because by 2023 it had grown strongly. We found no dedicated research on why it is losing readers faster.
Editors are also becoming fewer. Ukrainian Wikipedia had on average 1,134 active editors (5+ article edits a month) a month in 2023 and 821 in January–September 2026; occasional contributors (1–4 edits) numbered 2,121 and 1,457 113 H. The guideline against AI text, translated from English, was approved on 6 May 2026 90 H. On 30 July 2026 criterion C5 was added to the speedy deletion criteria: articles with obvious signs of unverified AI text (addressing the user, invented sources) can be deleted without discussion 114 M. The decline began earlier: in January–March 2026 there were 17.2% fewer active editors than a year earlier 113 H.
09 · ForecastWikipedia traffic in 2030 and 2035: three scenarios
We found no official forecast of Wikipedia pageviews for 2030 or 2035. The Wikimedia Foundation’s plan for 2026–27 says only that traffic, new editors and donors will keep shrinking, without numbers 109 H. Christophe Henner, a former chair of the Foundation’s board of trustees, put the trend in published pageviews at −8.4% a year over the past 12 months and −9.0% a year on a fitted curve in his own analysis on Meta-Wiki 115 L. News publishers surveyed by the Reuters Institute in late 2025 (the report came out in January 2026) expect their search traffic to fall by 43% over three years 116 M.
Forecasts about search are often wrong. In February 2024 Gartner predicted that the volume of queries to traditional search engines would fall by 25% by 2026 117 M. As of October 2026 there is no sign of this: in the second quarter of 2026 Alphabet reported growth in the number of queries (without absolute figures), and Google’s search revenue rose by 17% 118 M. People have not stopped searching; they have stopped clicking through to websites. Meanwhile the internet’s audience is growing: in 2025 about 6bn people were online, 240m more than a year earlier 119 H. So Wikipedia is losing not people on the internet but visits per person.
We built three scenarios. The starting point is our estimate for 2026: January–September multiplied by the ratio of the full year to the first nine months in 2025. This gives 158.8bn human pageviews for all Wikimedia projects and 78.8bn for English Wikipedia, 9.6% and 7.9% lower than in 2025. The estimate may be slightly too high: the Foundation has not cleaned November 2025 of bots, and the acceleration in October 2026 is not included L. We then set annual rates of change for 2027–2030 and 2031–2035.
- Plateau (−3% a year to 2030, then −1%). AI answers have already taken the simplest reference queries, and biographies hold up best (−13% against −66% in the physical sciences). This fits the hypothesis that the losses are concentrated in the “reference” layer L.
- Base (−8% a year to 2030, then −4%). The current pace continues (−6.5% in 2025, about −10% in 2026, Henner’s trend of −8–9% a year), slowing gradually as the substitution runs out L.
- Collapse (−15% a year to 2030, then −10%). Conversational mode becomes the default in Google, and AI agents read websites instead of people. This is close to the expectations of news publishers (−43% over three years), but far from Stack Overflow (−92% over three years) L.
| Scenario | All Wikimedia projects, 2030 | All Wikimedia projects, 2035 | English, 2030 | English, 2035 | 2035 vs the 2020 peak |
|---|---|---|---|---|---|
| Plateau | 140.6bn | 133.7bn | 69.8bn | 66.3bn | −33% |
| Base | 113.8bn | 92.8bn | 56.5bn | 46.0bn | −54% |
| Collapse | 82.9bn | 48.9bn | 41.1bn | 24.3bn | −76% |
Calculator: your scenario to 2035
Choose all Wikimedia projects or English Wikipedia and set the annual change in human pageviews for 2027–2030 and 2031–2035. The buttons fill in our three scenarios. The starting point is our estimate for 2026 L. The peak is 2020, when the pandemic produced the most human pageviews in the whole data series (since 2015): 200.3bn for all Wikimedia projects and 94.3bn for English Wikipedia 3 H.
| Indicator | What is likely by 2030–2035, by our estimate | Confidence | What could change the picture |
|---|---|---|---|
| Human pageviews | Keep falling; in the base scenario about 114bn in 2030 and 93bn in 2035 for all Wikimedia projects | Medium on direction, low on numbers | A change to Google’s interface (more or fewer links in AI answers); Foundation deals with AI companies on links |
| Bot share | Will exceed half of all pageviews (46.4% in the first nine months of 2026, 19% in 2016) | Medium | Strict rate limits; AI companies moving to paid access through Wikimedia Enterprise |
| Active editors (English) | Slow decline from the current ~30,000 a month, because the inflow of newcomers is thinning | Low | AI tools for editors that make checking sources easier; or, conversely, a flood of AI-generated text |
| Foundation money | Banner donations come under pressure because fewer people see them; the share of revenue from selling data rises from the current 4% | Medium | Large contracts with AI companies, or a law requiring compensation for the use of content |
| Wikipedia’s role | Remains an important source for training AI, but ChatGPT cites it less and less often, and people read it more often in retelling than on the page | Medium | Competition from AI encyclopaedias; so far their traffic is negligible |
10 · LimitsHonesty and limits
The strongest counterargument
One can object that Wikipedia is not declining but changing shape. By the Foundation’s estimate, about 5bn people use Wikipedia’s content in retold form, while about 1.5bn devices visit the site itself every month 120 H. The Foundation’s revenue is growing, the core of editors has not shrunk since 2013, and AI companies are starting to pay for data. Part of the fall is explained by measurement: until May 2020 “human” pageviews included undetected bots (the new class automated then took 6.4% of English Wikipedia’s pageviews) 57 H, and in 2024 bots were probably also detected less well than now L. After recalculating for bots, the Foundation put the decline at “approximately 8%”; that is several months of 2025 against the same months of 2024 20 H. By our data, the whole of 2025 was 6.5% below 2024. If the 2024 base contained more hidden bots than the current data, the true fall over two years is smaller than our −15.3%.
We think this argument is right about Wikipedia’s importance and wrong about its traffic. First, the error runs both ways: on desktop, the 2026 “human” traffic seems itself to contain undetected bots, so the mobile figure (−17.6% over the year) is closer to the truth than the overall one (−9.6%). Second, October 2026 (−12.8% over the first 25 days) shows acceleration, not stabilisation 4 H. Third, the Foundation itself calls this a structural shift in its plan for 2026–27 110 H.
Limits of this briefing
- No data for 2001–2007. There is no public series of pageviews for those years. For that period we cite separate estimates from different sources, which are not comparable with one another.
- Methodology breaks. The legacy counter (2008–2016) counted all requests, bots included, and was never recalculated; the new pageview definition has applied since May 2015 (the API series run from 1 July 2015); the class of “automated” bots dates from May 2020; in 2025–2026 the Foundation recalculated its data retroactively several times 121,49,50 H. From 4 June 2021 to 27 January 2022 a fault caused the Foundation to undercount global pageviews by 2–8%, and it has not corrected this gap, so 2021 is understated in the series 122 H. The numbers in this briefing are taken from the API as of 26 October 2026 and may change.
- Unique devices are not used by us as a measure of audience: in 2024–2026 bots distorted this series several times, and since May 2025 the Foundation has removed it from its monthly report 123,56 H.
- Causality. Only two studies with a control group estimate the effect of AI on Wikipedia specifically: one on AI Overviews (a preprint, not peer-reviewed) and one on ChatGPT (Lyu et al., WWW 2025; it compares articles that are similar and dissimilar to ChatGPT’s answers and does not give an overall effect). We show the coincidence in time and the mechanism, not a proven contribution from each factor. Among other causes the Foundation names social video services, where young people are going 20 H.
- Editors. Wikimedia’s series count registered editors separately in each language edition; the sum across 30 editions counts one person several times. The “before and after the ban” comparison is descriptive: there is no edition without the rules to serve as a control. We do not trust new temporary accounts (since November 2025 on English Wikipedia) as a measure of registrations 124 H.
- Scenarios. The rates of change in the scenarios are our choice, based on the current trend and analogues, not a model of proven accuracy L.
- Company data. The audiences of ChatGPT, Gemini and AI Mode are company claims without independent audit; the data from Semrush, Ahrefs and Similarweb are estimates from commercial panels.
- How the data were gathered. We downloaded the Wikimedia series directly from the Foundation’s API. The rest of the data were gathered by AI-based research agents with search, and separate reviewer agents checked the key figures against primary sources; we replaced corrected numbers, and unverified ones are marked with the label L.
11 · QuestionsFrequently asked questions
Is Wikipedia’s traffic really falling?
Yes. Human pageviews of all Wikimedia projects in 2025 were 175.6bn, 6.5% fewer than in 2024. In January–September 2026 they were a further 9.6% lower than a year earlier and 15.3% lower than in 2024. Every month since May 2024 has been lower than the same month a year earlier, and in October 2026 the decline accelerated to −12.8%.
Is ChatGPT to blame?
Mostly not directly. In 2023, ChatGPT’s first year, human pageviews even rose. The sustained decline began in May 2024, at the same time as the launch of Google’s AI answers above search results. According to Pew data for the US, with them people click on links in 8% of visits instead of 15%, and 78% of Wikipedia reading starts with search. The causal link is only partly established: the only estimate with a control group (a preprint) is about −5% of search referrals to English Wikipedia. Chatbots send roughly 500 times fewer people to news sites than Google does; for Wikipedia there is no such figure.
How much of Wikipedia’s traffic comes from bots?
In January–September 2026 bots accounted for 46.4% of all pageviews of Wikimedia projects, against 19% in 2016. Of desktop requests, the Foundation recognises only 35% as human. Some bots disguise themselves as people, so the official “human” series probably understates the decline. By our estimate, which assumes that desktop users drop off at the same rate as on mobile web, the true fall over a year may be closer to −17% than to −9.6%.
Does the ban on AI text drive editors away?
The data do not show this so far. English Wikipedia banned generating article text with language models on 20 March 2026, and since then it has lost active editors more slowly than most large language editions: −1% y/y in April–September. The core of editors (about 30,000 a month) has not changed since 2013. The number of occasional contributors and newcomers is falling, but that decline began in 2021, long before the ban.
What will happen to Wikipedia’s traffic in 2030 and 2035?
We found no official forecast. In our Base scenario (−8% a year to 2030, −4% after that), human pageviews of all projects would be about 114bn in 2030 and 93bn in 2035, which is 54% below the 2020 peak. Our scenarios for 2035 range from 49bn to 134bn. Wikipedia’s value as a source of knowledge for AI may persist, but ChatGPT cites it less and less often: according to Ahrefs, from 14.1% to 2.9% of US answers between March and September 2026.
12 · MoreRead also
13 · SourcesSources and methodology
Each number in the text leads to a source. Labels: H — primary source (Wikimedia Foundation data, a scientific paper, financial statements), M — secondary source or one checked against several, L — estimate, company claim, a single unconfirmed source or our calculation. We downloaded the pageview and editor series from the Wikimedia Analytics API on 26 October 2026; the rest of the data were gathered and cross-checked by AI-based research agents with search. Human pageviews are the agent type user (requests that the Foundation has not identified as bots). All sums are in US dollars.
- History of Wikipedia — first edit, 15 January 2001. en.wikipedia.org M
- Wikimedia — language edition statistics, 26 October 2026. meta.wikimedia.org H
- Wikimedia Analytics API — human pageviews, all projects (user), monthly 2015–2026. wikimedia.org H
- Wikimedia Analytics API — daily human pageviews, October 2026. wikimedia.org H
- Wikimedia Analytics API — year-on-year change in human pageviews (calculation). wikimedia.org H
- Pew Research Center — Google clicks when an AI summary appears, July 2025. pewresearch.org H
- Diff (Wikimedia Foundation) — where reading sessions start, March 2026. diff.wikimedia.org H
- Khosravi, Yoganarasimhan — AI Overviews and Wikipedia traffic, arXiv preprint 2026. arxiv.org M
- Luis Villa — pageview losses by topic. lu.is M
- Wikimedia Analytics API — human pageviews by language edition (example: Spanish Wikipedia). wikimedia.org H
- Wikimedia Analytics API — English Wikipedia editors by activity level, 2001–2026. wikimedia.org H
- Wikimedia Foundation — audited financial statements FY2024–25. upload.wikimedia.org H
- Wikimedia Foundation — Fundraising Report 2024–25. meta.wikimedia.org H
- Wikimedia Foundation — Wikimedia Enterprise revenue in audited statements. upload.wikimedia.org H
- Wikimedia — pageview definition and agent types. doc.wikimedia.org H
- Wikimedia — Pageviews API reference: data from 1 July 2015. meta.wikimedia.org H
- Wikimedia Analytics API — April 2020 pandemic records. wikimedia.org H
- Wikimedia Analytics API — human pageviews, English Wikipedia. wikimedia.org H
- Pew Research Center — Wikipedia at 25: What the data tells us, January 2026. pewresearch.org M
- Wikimedia Foundation — New User Trends on Wikipedia, 17 October 2025. diff.wikimedia.org H
- Wikimedia Analytics API — spider pageviews. wikimedia.org H
- Wikimedia Analytics API — automated pageviews. wikimedia.org H
- Wikimedia Analytics API — legacy pagecounts, 2008–2016. wikimedia.org H
- Erik Zachte — Wikimedia traffic fact sheet, 2007. meta.wikimedia.org H
- Wikimedia — server load worksheet, 2004. meta.wikimedia.org H
- Wikimedia Foundation — press kit, October 2006. meta.wikimedia.org H
- Wikipedia — Awareness statistics: Alexa rank. en.wikipedia.org M
- comScore — Wikipedia enters the US top 10, January 2007. web.archive.org H
- Brion Vibber — sampled pageview logs, June 2007. web.archive.org H
- Erik Zachte — Wikimedia in figures, August 2009. ftp3.us.freebsd.org H
- Wikimedia Foundation — Annual Report 2010–2011. meta.wikimedia.org H
- Diff (Wikimedia Foundation) — Wikimedia projects: 500 million, April 2013. diff.wikimedia.org M
- Halfaker, Geiger, Morgan, Riedl — The Rise and Decline of an Open Collaboration System, 2013. meta.wikimedia.org H
- Halfaker et al., 2013 — automated tools and newcomers. stuartgeiger.com H
- Wikimedia Foundation — Editor Trends Study, 2011. meta.wikimedia.org M
- Wikimedia Foundation — Strategic Plan 2011–2015. upload.wikimedia.org H
- Google — Knowledge Graph launch, 16 May 2012. blog.google H
- Wikistats-1 — desktop Wikipedia pageviews. web.archive.org H
- The Economic Times — Wikipedia pageview decline, January 2014. digit.in M
- The New York Times — mobile Wikipedia readership, February 2014. wikimedia.org H
- Wikimedia Analytics API — mobile overtakes desktop, June 2017. wikimedia.org H
- Oliver Keyes (Wikimedia Foundation) — Google referrals, 2015. hillwebcreations.com M
- Wikistats — new pageview definition from May 2015. web.archive.org H
- Wikimedia Phabricator T238357 — labelling bots as automated. wikitech.wikimedia.org H
- Diff (Wikimedia Foundation) — How crawlers impact the operations of the Wikimedia projects, 1 April 2025. diff.wikimedia.org H
- Diff (Wikimedia Foundation) — Quo Vadis, Crawlers?, March 2026. diff.wikimedia.org H
- Wikimedia Foundation — statement on automated agents, 5 October 2026. wikimediafoundation.org H
- Cloudflare — crawl-to-refer ratios of AI bots, August 2025. blog.cloudflare.com H
- Wikitech — data incident: May 2025 bot spike. wikitech.wikimedia.org H
- Wikitech — data incident: November 2025 bot wave. wikitech.wikimedia.org H
- Wikitech — Wikidata bot spike, July 2026. wikitech.wikimedia.org H
- Wikitech — KAPOW bot detection. wikitech.wikimedia.org H
- Wikimedia Phabricator T425668 — KAPOW scores for pageviews and unique devices. phabricator.wikimedia.org M
- Wikimedia Analytics API — readers of English Wikipedia by country. wikimedia.org H
- Wikimedia Foundation — monthly Movement Metrics report, July 2026. analytics.wikimedia.org H
- Wikimedia Foundation — Movement Metrics: unique devices excluded since May 2025. analytics.wikimedia.org H
- Wikimedia — the automated class from May 2020 (Phabricator T238357, API data). wikimedia.org H
- Reeves, Yin, Simperl — ChatGPT and Wikipedia engagement, arXiv 2024. arxiv.org M
- Google — AI Overviews launch, 14 May 2024. blog.google H
- Pew Research Center — most-cited sources in Google AI summaries. pewresearch.org H
- Alphabet — Q2 2025 results. blog.google M
- Alphabet — Q2 2026 results. blog.google M
- Wikimedia Foundation — Annual Plan 2026–2027: nearly 90% of visitors historically via Google. meta.wikimedia.org H
- Lyu et al. — ChatGPT and Wikipedia article views, WWW 2025. arxiv.org H
- Luis Villa — Vital Articles pageviews in eight languages, The Signpost, 2026. lu.is M
- Luis Villa — article freshness and pageview losses. lu.is M
- TechCrunch — ChatGPT at 900 million weekly users, February 2026. techcrunch.com M
- OpenAI / NBER — How People Use ChatGPT, September 2025. nber.org H
- Reuters Institute — Digital News Report 2026. reutersinstitute.politics.ox.ac.uk H
- Brown et al. — Language Models are Few-Shot Learners (GPT-3), 2020. arxiv.org H
- Touvron et al. — LLaMA, 2023. arxiv.org H
- Semrush — most-cited domains in AI, November 2025. semrush.com M
- Ahrefs — Wikipedia in ChatGPT answers, October 2026. ahrefs.com M
- Reuters Institute / Chartbeat — referrals from Google and ChatGPT. reutersinstitute.politics.ox.ac.uk M
- PPC Land — Stack Overflow questions from Stack Exchange Data Explorer. ppc.land M
- del Rio-Chanona, Laurentsyeva, Wachs — PNAS Nexus, 2024. academic.oup.com H
- Chegg — Form 10-Q, Q2 2026. sec.gov H
- Grokipedia — launch and scale (media summary). en.wikipedia.org M
- Ahrefs — Wikipedia vs Grokipedia, March 2026. ahrefs.com M
- Wikipedia — RfC «Is this proposal ready to be promoted?», August–October 2023. en.wikipedia.org H
- Wikipedia — WikiProject AI Cleanup. en.wikipedia.org H
- Wikipedia — RfC on LLM-generated comments, December 2024 – January 2025. en.wikipedia.org H
- Wikipedia — speedy deletion criterion G15. en.wikipedia.org H
- Wikipedia — NEWLLM guideline RfC, November 2025. en.wikipedia.org H
- Wikipedia — RfC banning LLM-generated article text, March 2026. en.wikipedia.org H
- Wikipedia — guideline Writing articles with large language models. en.wikipedia.org H
- Wikipedia — RfC on blocks for LLM use, May 2026. en.wikipedia.org H
- MediaWiki — LLM Paste Check. mediawiki.org H
- German Wikipedia — Meinungsbild on banning AI text, February 2026. de.wikipedia.org M
- Ukrainian Wikipedia — guideline on writing articles with LLMs. uk.wikipedia.org H
- Meta-Wiki — AI policies by project. meta.wikimedia.org M
- Wikipedia — category of articles with suspected AI text (API). en.wikipedia.org H
- Wikipedia — New Pages Patrol backlog. en.wikipedia.org H
- Brooks, Eggert, Peskoff — The Rise of AI-Generated Content in Wikipedia, 2024. arxiv.org H
- Wikimedia Analytics API — new articles, English Wikipedia. wikimedia.org H
- Wikipedia — WikiProject AI Cleanup guide. en.wikipedia.org H
- Wikipedia — Signs of AI writing. en.wikipedia.org H
- Wiki Education — generative AI and Wikipedia editing: what we learned in 2025. wikiedu.org H
- Wikimedia Foundation — Movement Metrics, July 2026: active editors. analytics.wikimedia.org H
- Wikimedia Foundation — Movement Metrics, July 2026: editors and newcomers. analytics.wikimedia.org H
- Wikimedia Foundation — 2026–27 OKRs: editor retention. meta.wikimedia.org H
- ProPublica Nonprofit Explorer — Wikimedia Foundation Form 990. projects.propublica.org M
- Wikimedia Foundation — FY2024–25 audit FAQ. upload.wikimedia.org H
- Wikimedia Endowment — FY2024–25 financials. meta.wikimedia.org H
- Wikimedia Foundation — audit FAQ: donation mix. meta.wikimedia.org H
- Wikimedia Foundation — Google becomes the first Wikimedia Enterprise customer, June 2022. enterprise.wikimedia.com H
- Wikimedia Enterprise — partners at Wikipedia’s 25th anniversary, January 2026. enterprise.wikimedia.com H
- Wikimedia Foundation — budget overview 2026–2027. meta.wikimedia.org H
- Wikimedia Foundation — Product & Technology OKRs 2026–2027. meta.wikimedia.org H
- Wikimedia Foundation — Annual Plan 2026–2027. meta.wikimedia.org H
- Wikimedia Analytics API — human pageviews, Ukrainian Wikipedia. wikimedia.org H
- Wikimedia Analytics API — readers by country (rank and order of magnitude only). wikimedia.org M
- Wikimedia Analytics API — Ukrainian Wikipedia editors. wikimedia.org H
- Ukrainian Wikipedia — speedy deletion criteria (C5). uk.wikipedia.org M
- Christophe Henner — personal traffic analysis on Meta-Wiki, August 2026. meta.wikimedia.org L
- Reuters Institute — Journalism, media and technology trends and predictions 2026. reutersinstitute.politics.ox.ac.uk M
- Gartner — search volume prediction for 2026, February 2024. mediapost.com M
- Alphabet — Search revenue, Q2 2026 (secondary source). digitalapplied.com M
- ITU — Facts and Figures 2025. itu.int H
- Wikimedia Foundation — Annual Plan 2026–2027: audience funnel. meta.wikimedia.org H
- Wikimedia Phabricator — legacy counts never reprocessed. phabricator.wikimedia.org H
- Wikitech — traffic data loss, June 2021 – January 2022. wikitech.wikimedia.org H
- Wikimedia — unique devices changelog. wikitech.wikimedia.org H
- Wikimedia Phabricator — temporary accounts on English Wikipedia. phabricator.wikimedia.org H