Wikipedia Statistics 2026
By Axis Intelligence Research
Co-author: Sarah Mitchell | Last updated: September 22, 2026 | License: CC BY 4.0
Wikipedia carries 66 million articles across 342 language editions, answers roughly 508 million pageviews a day, and does it on an operating budget of $189.7 million. Human pageviews fell about 8% year over year. Machines now account for 35% of pageviews, 65% of the most expensive traffic, and nearly 40% of all edits. The encyclopedia that trains the models is being read by fewer people every year.
Quick Answer
Wikipedia in 2026 is simultaneously the most-cited source in generative AI and a platform losing human readers. The Wikimedia Foundation reported human pageviews down roughly 8% year over year for March–August 2025 after it rewrote its bot-detection logic. Wikipedia accounts for 47.9% of ChatGPT’s top-ten citation share, per Profound’s analysis of 680 million citations. Commercial data licensing — Wikimedia Enterprise — returned $8.3 million in FY2025 against $189.7 million in Foundation expenses. Axis Intelligence Research calculates that licensing covers 4.37% of what it costs to run the corpus the models depend on, and places the inaugural Knowledge Commons Drain Ratio (KCDR™) at 69.2 of 100.
Key Findings
- Axis Intelligence Research calculates a Knowledge Commons Drain Ratio (KCDR™) baseline reading of 69.2 out of 100 for Wikipedia as of September 2026, placing it in the index’s “drain” band, where machine consumption outpaces human participation.
- Axis Intelligence Research estimates Wikipedia operates at $1.57 per 1,000 human pageviews, dividing the Wikimedia Foundation’s $189,748,442 in FY2025 expenses (IRS Form 990) by an estimated 120.5 billion annual human pageviews.
- Axis Intelligence Research finds that commercial AI and search licensing covers 4.37% of the Wikimedia Foundation’s operating costs — $8.3 million of Wikimedia Enterprise revenue against $189.7 million in expenses for the fiscal year ending June 30, 2025.
- Axis Intelligence Research calculates a machine load multiplier of 1.86× for Wikimedia infrastructure: bots generate 35% of pageviews but 65% of requests reaching core datacenters, per the Wikimedia Foundation’s own April 2025 infrastructure analysis.
- Axis Intelligence Research calculates that roughly 196,992 of Wikipedia’s approximately 492,480 daily edits are made by bots, derived from the 2024 rate of 5.7 edits per second and the near-40% bot share of edits documented by Pew Research Center.
How Big Is Wikipedia in 2026?
The scale numbers are the ones every compilation repeats, so start with them and move on quickly — they are the denominator, not the story.
| Metric | Value | As of | Source |
|---|---|---|---|
| Articles, all language editions | 66,000,000+ | Dec 2025 | Pew Research Center |
| Articles, English Wikipedia | ~7,000,000 | Dec 2025 | Pew Research Center |
| Language editions | 342 | Late 2025 | Pew Research Center |
| Words, English Wikipedia | 5,000,000,000+ | Dec 2025 | Pew Research Center |
| Storage footprint | ~775 TB | Dec 2025 | Pew Research Center |
| Media files on Wikimedia Commons | 144,000,000 | Apr 2025 | Wikimedia Foundation |
| Cumulative pageviews since 2015 | 1.9 trillion | Dec 2025 | Pew Research Center |
| English share of all pageviews | 49% | 2015–2025 | Pew Research Center |
Pew Research Center’s 25th-anniversary data analysis, published January 13, 2026, is the cleanest independent accounting of Wikipedia’s scale currently available, and it is the source Axis Intelligence Research uses for the structural figures throughout this report.
One number in that table deserves a flag. The 508-million-pageviews-per-day figure is a decade average, not a 2025 reading. Sources that quote it as “Wikipedia’s current daily traffic” are quietly converting a ten-year mean into a present-tense claim. It is the correct input for long-run cost math; it is the wrong input for a year-over-year trend.
Sarah Mitchell: Scale figures are the part of Wikipedia’s story that everyone gets right and nobody learns anything from. 66 million articles is a training-corpus statistic. What matters for anyone building retrieval systems in 2026 is not how much text exists — it’s the rate at which humans are still checking it. A corpus can grow while its verification layer thins. Those are two different curves, and only one of them gets cited in keynote slides.
Is Wikipedia Losing Traffic to AI? The 2025–2026 Pageview Data
Yes, and the way the number was produced is more interesting than the number.
In May 2025 the Wikimedia Foundation noticed an implausible traffic surge, much of it from Brazil, that passed as human. Investigation found bots engineered to defeat detection. The Foundation rewrote its classification logic and reclassified six months of data. What survived the reclassification was a decline of roughly 8% in human pageviews for March–August 2025 versus the same months in 2024, published in the Foundation’s New User Trends post on October 17, 2025.
The Foundation attributes the decline to generative AI and social platforms answering questions without sending anyone onward. It also cautions — correctly — that year-over-year comparison is complicated when the counting rules changed mid-stream.
By its April 2026 annual planning cycle, published on Diff as Global Trends 2026, the Foundation reported that traffic decreased year over year across the first half of 2025, that awareness and affinity toward Wikipedia declined, and that total active editors continued to decline year over year, with the steepest drops in account registrations and on emerging-language Wikipedias. The same document places AI-overview-driven traffic declines across publishers generally in the 20–60% range.
Against that 20–60% publisher range, Wikipedia’s 8% looks like resilience. It is worth asking why. Wikipedia’s traffic is unusually event-driven — the Charlie Kirk article took nearly 15 million views on September 10, 2025 alone — and event spikes are exactly the queries a static model cannot answer from memory. The articles that bleed are the stable, definitional ones. Which is to say: the decline is not uniform, it is selective, and the selection principle is “can a model answer this without me.”
What the Research Actually Establishes — and What It Doesn’t
Two peer-reviewed studies bear directly on this, and they disagree.
Reeves, Yin and Simperl examined twelve Wikipedia language instances before and after ChatGPT’s November 2022 release and found no evidence of a fall in engagement across pageviews, unique visitors, edits, or editor counts. Growth was lower in languages where ChatGPT was available than where it wasn’t — a constraint on growth, not a decline (arXiv:2405.10205).
Lyu, Siderius, Li, Acemoglu, Huttenlocher and Ozdaglar reached a sharper conclusion at WWW 2025. Using differences-in-differences with dissimilar articles as control, they found that newly created, popular articles whose content overlaps with ChatGPT 3.5 saw a greater decline in both editing and viewership than dissimilar articles (arXiv:2503.00757).
Axis Intelligence Research reads these as compatible rather than contradictory, and the reconciliation is the finding: aggregate engagement held while substitutable engagement fell. Averages hid a composition change. Anyone citing the 2024 study as evidence that “AI didn’t hurt Wikipedia” is citing a true result about the wrong unit of analysis.
Sarah Mitchell: The 2024 paper measured the platform; the 2025 paper measured the article. That distinction is the whole argument. A model with a 200K context window doesn’t substitute for Wikipedia — it substitutes for the eleven paragraphs of Wikipedia that answer a definitional question, which happen to be the paragraphs that historically recruited new editors. Nobody’s first edit was to a breaking-news article. It was a typo fix on something boring.
How Much of AI’s Answers Come From Wikipedia?
Enough that the dependency runs one direction and the traffic runs the other.
Profound analyzed 680 million citations across AI platforms for the August 2024–June 2025 window. Within ChatGPT’s ten most-cited sources, Wikipedia holds 47.9% — nearly as much as every other top-ten source combined. Across all ChatGPT citations, Wikipedia takes 7.8%.
| Platform | Wikipedia share of top-10 citations | Wikipedia share of all citations |
|---|---|---|
| ChatGPT | 47.9% | 7.8% |
| Google AI Overviews | 5.7% | 0.6% |
| Perplexity | Not in top 10 | Not reported |
Source: Profound, AI Platform Citation Patterns, 680 million citations analyzed, August 2024–June 2025 · tryprofound.com
The spread between platforms is the part worth noticing. ChatGPT leans on Wikipedia harder than any other major assistant by an order of magnitude; Google AI Overviews, which draws on its own index, treats Wikipedia as one source among many. Wikipedia’s exposure to the AI transition is therefore not evenly distributed — it is concentrated in the single platform with the largest consumer reach.
For the corresponding view from the other side of that relationship, see our AI search statistics report, which tracks referral share across assistants, and our Google AI Overviews statistics report on citation behavior inside Google’s answer layer.
Bot Traffic on Wikipedia: The Infrastructure Numbers
This is where Wikipedia’s situation stops resembling a publisher’s and starts resembling a utility’s.
The Foundation’s April 2025 infrastructure analysis, How Crawlers Impact the Operations of the Wikimedia Projects, reports that bandwidth for multimedia downloads grew 50% since January 2024, that bots represent about 35% of total pageviews, and that 65% of the most expensive traffic — requests that miss edge caches and hit core datacenters — comes from bots.
That asymmetry has a number.
Axis machine load multiplier = bot share of core-datacenter requests ÷ bot share of pageviews = 65% ÷ 35% = 1.86×
Crawlers cost Wikimedia 1.86 times what their pageview share implies. The mechanism is caching: human readers cluster on popular articles that sit in regional caches cheaply, while scrapers walk the long tail and bulk-pull media, forcing origin fetches. A crawler that reads everything is, from an infrastructure ledger’s view, the most expensive possible reader.
Pew puts machine reads at more than 88 billion pageviews in 2025. Note that this figure and the Foundation’s 35% share are measured under different definitions and windows; Axis Intelligence Research reports them side by side rather than reconciling them into a single bot-traffic number, because dividing one by the other would produce a ratio with no defensible meaning.
For sector context outside Wikimedia, Cloudflare’s network-wide telemetry showed AI and search crawler traffic up 18% year over year to May 2025, with Googlebot at 50% of crawl share, GPTBot at 7.7% and ClaudeBot at 5.4% (Cloudflare). Those are Cloudflare’s customers, not Wikimedia’s servers — useful as a directional read on crawler economics, not as a Wikipedia measurement.
Bot Edits: Machines Inside the Editorial Layer
Bots don’t only read Wikipedia. They write it.
Pew reports 5.7 edits per second across Wikipedia in 2024 and a near-40% bot share of all edits. Axis Intelligence Research converts that into daily terms:
5.7 edits/sec × 86,400 sec = 492,480 edits per day 492,480 × 0.40 = ~196,992 bot edits per day
Most of those are maintenance — category sorting, link repair, vandalism reversion — not authorship. The distinction matters, and the aggregate edit count is a poor proxy for human contribution precisely because of it. But the direction is the point: the Foundation reports active human editors declining while automated edit volume holds. The encyclopedia is becoming more machine-maintained and less human-attended at the same time.
Sarah Mitchell: There’s a failure mode people building RAG pipelines on Wikipedia dumps should sit with. Retrieval grounding assumes the corpus has a human verification loop attached. Bot edits keep the corpus tidy — they don’t keep it true. If the editor base thins while the bot layer holds, the file keeps looking healthy by every automated metric right up until the point where nobody caught the error. That’s not a 2026 problem. It’s a 2029 problem being priced at zero today.
Who Pays for the Data AI Takes? Wikimedia’s Revenue Model in 2026
The Wikimedia Foundation’s IRS Form 990 for the fiscal year ending June 2025, filed May 13, 2026, reports $205,389,744 in total revenue, $189,748,442 in total expenses, and $296,550,030 in net assets (ProPublica Nonprofit Explorer).
Against that, the commercial answer to AI extraction is Wikimedia Enterprise, the paid, SLA-backed data service launched in 2021. Its FY2024–2025 financial report shows $8.3 million in revenue, up 148% from $3.4 million, against $4.4 million in expenses, with 13 paid commercial customers at period end — Ecosia, Pleias, ProRata.ai and Nomic AI among the new ones. Cumulative net profit since January 2022: $646,000.
Axis Intelligence Research runs three ratios on those two filings.
| Axis metric | Value | Formula |
|---|---|---|
| Licensing cost coverage | 4.37% | $8,300,000 ÷ $189,748,442 |
| Revenue per active editor | $342.32 | $205,389,744 ÷ 600,000 |
| Expense per active editor | $316.25 | $189,748,442 ÷ 600,000 |
| Cost per 1,000 human pageviews (est.) | $1.57 | $189,748,442 ÷ 120,523,000,000 × 1,000 |
Sources: IRS Form 990 FY2025 via ProPublica · Wikimedia Enterprise FY2024–2025 report · Pew Research Center (editor and pageview inputs) · Axis Intelligence Research calculation, September 21, 2026
The human pageview denominator is an Axis estimate: 508 million daily pageviews × 365 × 65% human share = 120.5 billion. Because the daily figure is a decade average, the resulting cost-per-pageview is a long-run rate, not a FY2025 reading. Axis Intelligence Research labels it an estimate for that reason.
A cross-check worth stating: $8.3M ÷ $205.4M = 4.04%, which matches the Foundation’s own reported “4% of total revenue.” The two independently sourced filings agree.
One more ratio, offered as framing rather than as a price. Machine traffic pulled more than 88 billion pageviews in 2025; commercial licensing returned $8.3 million the same year. $8,300,000 ÷ 88,000,000,000 × 1,000 = $0.094 per thousand machine pageviews. Enterprise revenue is contractual and not metered to crawl volume, so this is a ratio of two independently measured quantities, not a rate card. It is, however, the clearest single expression of the gap between what the corpus is worth to model builders and what flows back.
Sarah Mitchell: $8.3 million growing 148% is a real product finding a real market — 13 paying customers in a segment that mostly prefers to scrape for free is not nothing. It’s also 4.37% of the bill. Every AI lab’s inference stack has a Wikipedia-shaped dependency in it, and the aggregate contractual acknowledgment of that dependency is smaller than the annual compute spend on a single mid-sized training run. The market has priced the corpus at roughly the cost of not being sued.
The Axis Knowledge Commons Drain Ratio (KCDR™)
Wikipedia’s condition can’t be read from any single metric. Traffic alone misses the contributor side; editor counts alone miss the machine load. So Axis Intelligence Research built a composite.
KCDR™ — the Knowledge Commons Drain Ratio — measures the degree to which machine consumption of an open knowledge base is outrunning the human participation that sustains it. Scale: 0–100, higher means more drain.
Components, Inputs and Weights
| Component | Weight | Input | Value | Score | Normalization |
|---|---|---|---|---|---|
| Human engagement trend | 30% | Human pageview change YoY | −8% | 50.0 | Decline in pp × 6.25; 16% annual decline = 100 |
| Machine load concentration | 25% | Core-DC request share ÷ pageview share | 1.857× | 57.1 | (multiplier − 1) ÷ 1.5 × 100; 2.5× = 100 |
| AI answer dependency | 25% | Wikipedia share of ChatGPT top-10 citations | 47.9% | 95.8 | share × 2; 50% = 100 |
| Contributor substitution | 20% | Bot share of all edits | 40% | 80.0 | share × 2; 50% = 100 |
KCDR™ = (0.30 × 50.0) + (0.25 × 57.1) + (0.25 × 95.8) + (0.20 × 80.0) = 69.2
Inputs: Wikimedia Foundation (Diff, Oct 2025 and Apr 2025) · Profound (Aug 2024–Jun 2025) · Pew Research Center (Jan 2026). Reading as of September 21, 2026. This is the baseline reading — the first computation of the metric, with no prior history to compare against.
Interpretation bands: 0–33 balanced commons · 34–66 strain · 67–100 drain.
At 69.2, Wikipedia sits just inside the drain band, and the composition explains why. The traffic component is the mildest of the four. What pushes the reading up is dependency (95.8) and contributor substitution (80.0) — machines consuming the corpus at scale while an increasing share of the maintenance work is itself automated. An index built only on pageviews would read 50 and miss the structural half of the problem.
Anchors are fixed at their stated values so future readings remain comparable. If a component’s source changes methodology, Axis Intelligence Research re-baselines transparently and publishes the change rather than restating history.
What KCDR™ does not capture: donation behavior, article quality, reader satisfaction, or non-Wikipedia knowledge commons. Its AI-dependency component reads ChatGPT only, because it is the platform with both the highest Wikipedia concentration and published measurement.
Cite as: Axis Intelligence Research, Knowledge Commons Drain Ratio (KCDR™) — Wikipedia baseline reading 69.2, September 21, 2026.
What Happens Next: Three Things Worth Watching
Whether licensing scales past a rounding error. Enterprise tripled off a small base. The question is whether the next cohort of customers is AI labs or search startups — and whether any of the frontier labs sign. Thirteen customers is a product; it is not yet a funding model.
Whether editor decline reaches the quality layer. Active editors are falling and bot edits are not. Nothing in the current metrics distinguishes “tidy” from “accurate.” Whoever publishes a defensible article-quality trend line will own the most important Wikipedia statistic of the next three years, and it does not currently exist.
Whether attribution becomes measurable. The Foundation has said it is developing attribution frameworks. Any AI platform that publishes per-source referral data would convert this entire debate from inference to measurement. Currently none do.
Methodology
Collection. Nine primary sources were fetched and read on September 21, 2026: three Wikimedia Foundation Diff posts (October 17, 2025; April 1, 2025; July 16, 2026), the Wikimedia Enterprise FY2024–2025 financial report (November 24, 2025), Pew Research Center’s January 13, 2026 analysis, the Wikimedia Foundation’s IRS Form 990 for FY ending June 2025 via ProPublica Nonprofit Explorer, Profound’s citation-pattern analysis, Cloudflare’s crawler telemetry, and two peer-reviewed studies (arXiv:2405.10205; arXiv:2503.00757). Every figure in this report carries a row in the accompanying CSV with source organization, document, URL and retrieval date.
Formulas. Machine load multiplier = 65 ÷ 35. Daily edits = 5.7 × 86,400. Bot edits/day = daily edits × 0.40. Human pageview estimate = 508,000,000 × 365 × 0.65. Cost per 1,000 human pageviews = 189,748,442 ÷ 120,523,000,000 × 1,000. Licensing coverage = 8,300,000 ÷ 189,748,442. Revenue per editor = 205,389,744 ÷ 600,000. KCDR™ as shown in the component table above.
Estimates vs. observations. Figures labeled “Axis Intelligence Research finds” are direct ratios of two reported values. Figures labeled “estimates” — cost per 1,000 human pageviews, annual human pageviews — depend on the decade-average daily pageview figure and are long-run rates rather than single-year readings.
What we did not merge. Pew’s 88 billion machine pageviews (2025) and the Foundation’s 35% bot pageview share (April 2025) are measured under different definitions and windows. They are reported separately. Cloudflare’s crawler shares describe Cloudflare’s network, not Wikimedia’s, and are used only as sector context.
Currency of data. Wikipedia’s raw counters change continuously. Figures here are dated snapshots from the sources named, not live reads.
About This Dataset
File: wikipedia-statistics.csv — 59 rows, 13 columns, one observation per row, full provenance columns (source_org, source_document, source_url, retrieved_date, is_primary, axis_calculated, method_note). Every number in this article exists as a row.
Coverage: Wikipedia and Wikimedia project scale, traffic, bot traffic, editorial activity, financials, AI citation share, and Axis-calculated metrics including KCDR™ components. Temporal coverage 2015–2026; spatial coverage global.
License: CC BY 4.0. Free to use, redistribute and build on, including commercially, with attribution.
Citation: Axis Intelligence Research, Wikipedia Statistics 2026, 2026. https://axis-intelligence.com/wikipedia-statistics/
Mirrors: Hugging Face · Kaggle · GitHub.
Update policy: Relevance-driven. This dataset is revised when a component source publishes new data, not on a calendar.
Cite This Report
APA: Axis Intelligence Research. (2026, September 21). Wikipedia statistics 2026: Traffic, editors, revenue, and the cost of being AI’s most-cited source. Axis Intelligence. https://axis-intelligence.com/wikipedia-statistics/
MLA: Axis Intelligence Research. “Wikipedia Statistics 2026: Traffic, Editors, Revenue, and the Cost of Being AI’s Most-Cited Source.” Axis Intelligence, 21 Sept. 2026, axis-intelligence.com/wikipedia-statistics/.
Chicago: Axis Intelligence Research. “Wikipedia Statistics 2026: Traffic, Editors, Revenue, and the Cost of Being AI’s Most-Cited Source.” Axis Intelligence, September 21, 2026. https://axis-intelligence.com/wikipedia-statistics/.
Frequently Asked Questions
Is Wikipedia losing traffic because of AI?
Partly. The Wikimedia Foundation reported human pageviews down roughly 8% year over year for March–August 2025 and attributed the decline to generative AI and social platforms. Peer-reviewed work adds precision: articles whose content overlaps with what ChatGPT answers well declined more than dissimilar articles, meaning the loss is concentrated in definitional content rather than spread evenly.
Do AI bots count in Wikipedia’s official pageview statistics?
They are separated, but the separation is imperfect and changes over time. The Wikimedia Foundation classifies traffic as human or bot, and in May 2025 it found bots engineered to pass as human. It rewrote its detection logic and reclassified six months of data. Any Wikipedia traffic comparison spanning that revision is comparing two different counting rules.
Does Wikipedia get paid when AI companies use its content?
Only a fraction of it is contractual. Wikimedia Enterprise, the Foundation’s paid data service, earned $8.3 million in the fiscal year ending June 2025 from 13 commercial customers. Axis Intelligence Research calculates that this covers 4.37% of the Foundation’s $189.7 million in operating expenses. Wikipedia’s content is openly licensed, so use without payment is generally permitted.
Can I legally use Wikipedia data to train a model?
Wikipedia text is published under a Creative Commons license requiring attribution and share-alike terms; images on Commons carry varied individual licenses. Legal permission to use is not the same as permission to crawl at will — the Foundation offers Wikimedia Enterprise specifically so high-volume reusers get SLA-backed access without placing load on infrastructure funded by reader donations. Check the specific license on the specific content.
How many Wikipedia edits are made by bots rather than humans?
Bots accounted for nearly 40% of all Wikipedia edits in 2024, per Pew Research Center. At the 2024 rate of 5.7 edits per second, Axis Intelligence Research calculates roughly 196,992 bot edits per day out of approximately 492,480 total. Most are maintenance tasks — link repair, categorization, vandalism reversion — rather than authorship.
Why do Wikipedia article counts differ between sources?
Because “article” is defined differently depending on the counter. Some totals include only content pages in main namespace; others include redirects, disambiguation pages, stubs, or non-Wikipedia Wikimedia projects. A figure without a namespace definition and a date is not comparable to another such figure. Every count in this report states its source and as-of date.
Is the Wikimedia Foundation running out of money?
No. The Foundation reported $205.4 million in revenue against $189.7 million in expenses and $296.6 million in net assets for the fiscal year ending June 2025. The structural concern is not the current balance sheet — it is that the donation model depends on human visitors, and human visits are the metric declining.
What happens to AI answer quality if Wikipedia’s editor base keeps shrinking?
This is the open question with no published measurement. Wikipedia functions as a verification layer for models that retrieve from it, and that layer is maintained by humans whose numbers the Foundation reports as declining while bot edit volume holds. Automated metrics of corpus health would not detect a decline in accuracy caused by fewer human reviewers. Axis Intelligence Research tracks this as the most consequential unmeasured variable in the AI-Wikipedia relationship.
Which AI platform relies on Wikipedia most?
ChatGPT, by a wide margin. Wikipedia holds 47.9% of ChatGPT’s top-ten citation share versus 5.7% for Google AI Overviews, per Profound’s analysis of 680 million citations from August 2024 to June 2025. Google’s answer layer draws on its own index; ChatGPT’s dependency on a single open corpus is structurally different.
What is the Knowledge Commons Drain Ratio?
KCDR™ is an Axis Intelligence Research composite metric measuring how far machine consumption of an open knowledge base has outrun the human participation sustaining it. It combines four sourced inputs — human pageview trend (30%), machine load concentration (25%), AI citation dependency (25%) and bot share of edits (20%) — on a 0–100 scale. Wikipedia’s baseline reading is 69.2 as of September 21, 2026.