Executive Summary

First, we have to pin down exactly what the number is. In October 2025, Marshall Miller of the Wikimedia Foundation reported that human page views on Wikipedia had fallen 8% between March and August 2025 compared with the same period a year earlier. What makes the figure matter is the methodology behind it. That spring, the Foundation updated its bot-detection logic, and it confirmed this number only after filtering out a wave of human-disguised traffic flooding in from Brazil. In other words, the 8% is the decline in "real people left after the bots are removed."

The −23% often quoted in the press is a different number. That is a third-party estimate from Similarweb, capturing a decline spread over three years — from roughly 165 million daily visits in 2022 to about 128 million in 2025. Because the period and the definition both differ, the two should not be blurred into a single sentence. This report treats the Foundation's official 8% as its baseline and uses the Similarweb figure only as corroboration of a longer-term trend.

Figure 1. Comparison of the two measurement baselines. An original diagram by Pebblous, based on Wikimedia Foundation (Oct 2025) and Similarweb (2025).

The Foundation names generative-AI summaries and social media as the causes of the decline. But this is circumstantial judgment, not a controlled experiment. No study has yet cleanly separated out the contribution of other factors — search-algorithm changes, pandemic base effects, shifts in the SEO landscape. Distinguishing correlation from causation is the first discipline in handling this subject honestly.

That said, the circumstantial case is thick. The share of "zero-click" searches — those that end without any click at all — has already passed half. SparkToro counted 58.5% of Google searches in 2024 as leading to no click of any kind, and Pew Research reported that on searches where an AI Overview appeared, only 1% of users clicked through to an actual source link. The more firmly the habit of receiving answers directly on screen sets in, the fewer footsteps return to Wikipedia, the origin of those answers. And this is not something Wikipedia faces alone. Monthly visits to news sites, too, have shed roughly a quarter over the year since mid-2024 (Similarweb). The structure in which summaries stand in for originals is leaving the same mark across the knowledge commons as a whole.

Just how much AI leans on Wikipedia shows up most sharply in citation data. When the analytics firm Profound examined roughly 680 million AI citations between August 2024 and June 2025, Wikipedia's share among the top 10 source domains that ChatGPT pulls from most in its answers was 47.9%. On Google's AI Overviews, Wikipedia's share by the same measure was just 5.7%. Looked at within the top sources alone, ChatGPT's dependence on Wikipedia is overwhelming.

Here the definition of the number has to be attached. The 47.9% is a "share within the top 10 sources," not "half of all answers." Measured against total citations, Wikipedia's share is 7.8%, and the proportion of responses that cite Wikipedia is 2.49%. Even so, the direction of the story doesn't change. The fact that AI reaches for Wikipedia first when choosing a source stays the same. The diagram below places the two platforms' within-top-sources Wikipedia shares side by side.

Figure 2. Wikipedia's share of top cited sources by AI platform. Source: Profound (2025), recomposed by Pebblous. Figures use the "share within the top 10 domains" definition.

The paradoxical part is its weight in training corpora. Across the major pretraining datasets, Wikipedia typically accounts for 2–4.5% of tokens: about 3.8% in SlimPajama, 4.5% in LLaMA-1, roughly 3.0% in GPT-3. By volume alone, it is a small sliver. And yet the reason Wikipedia is called the "anchor" of factuality and neutrality lies in quality density. Sentences that have passed through edit histories, source citations, and the neutral-point-of-view rule are far more trustworthy than the average text scraped off the open web. Small share, large role.

When that anchor shakes, the ripple runs well past the token share. A model's ability to verify facts and stay current depends disproportionately on the minority of high-quality sources inside the corpus. So a drop in Wikipedia's freshness is tied directly to a blurring of the model's factual accuracy.

The principle that keeps Wikipedia alive is a simple virtuous loop. People visit; some of them become editors; edits keep articles current; some become donors and hold up operations. The visit is the mouth of this loop. Narrow the mouth, and the water flowing down through it thins too.

Yet look at the Foundation's finances and it seems, at first, like the opposite. Total revenue for FY24-25 was $208.6 million, an all-time high. Pull that figure out on its own and it's easy to misread as "there's no crisis." Which is precisely why there's a number that has to be read alongside it. Over the same period new donors fell 10.2% (from about 2.6 million to 2.3 million), and total donors dropped 2.4% as well. Revenue rose not because more people came, but because existing supporters gave more. It is a signal that the top of the funnel — new inflow — is drying up first.

Figure 3. The virtuous-loop funnel. Visits, new editors, and new donors all fall, yet total revenue holds at a record — the paradox. Source: Wikimedia FY24-25 audit and fundraising reports, recomposed by Pebblous.

The editor side shows a similar grain. Active editors on English Wikipedia still number around 290,000, but new account registrations have slid steadily downward since 2020–21. Fewer new editors won't crash the article count overnight, but the capacity to keep knowledge current and catch errors quietly thins over time.

If falling traffic is the "revenue" side of the commons, bot crawlers present an entirely different invoice on the "cost" side. According to the infrastructure analysis Wikimedia published in April 2025, bots make up just 35% of all page views yet generate 65% of expensive core-datacenter traffic. People tend to view popular, frequently accessed articles, which are served cheaply from cache; bots, by contrast, sweep through rarely visited articles too and bypass the cache. Multimedia bandwidth has grown 50% since January 2024, and a large share of that increase has been attributed to AI crawlers. Behind this bandwidth surge is scraping — pulling images from Wikimedia Commons en masse to train computer-vision models. It is not just text: crawlers are sweeping through the more than 140 million media files piled up on Wikimedia Commons.

A more dramatic asymmetry shows up in the traffic crawlers send back. According to figures Cloudflare published in July 2025, traditional search — Google — returns one visitor for every 14 pages it crawls. OpenAI, by contrast, sends one visitor per 1,700 crawls, and Anthropic one per a staggering 73,000. The balance between how much data is taken and how much traffic is returned is qualitatively different.

Figure 4. Crawl-to-referral ratio by crawler type (log scale). An original diagram by Pebblous, based on Cloudflare (Jul 2025).

Table 1. The structure of the bots' free ride. Source: diff.wikimedia.org (Apr 2025), Cloudflare (Jul 2025). Referral ratios are inbound visitors per crawl request.

This cost paradox is a separate axis from the traffic paradox. Even if the earlier story — "summaries steal visits" — turned out to be false, the fact that bots drive up costs still holds. Wikipedia is caught in a double bind: fewer people come, yet spending rises because of the bots.

If everything so far is "what is happening now," this section is the theory of "why that is dangerous to AI itself." Two lines of research support the self-cannibalization thesis.

The first is model collapse. In a paper published in Nature in 2024, Shumailov and colleagues showed that recursively training successive generations of models on AI-generated synthetic data causes an irreversible degradation in performance. Early on, the rare information that sat in the tails of the distribution disappears; with each generation, outputs converge toward the average and lose diversity. And what that experiment retrained, generation after generation, was none other than a Wikipedia-based language model.

The second is data exhaustion. Epoch AI estimates that the stock of publicly available, high-quality human-generated text will run out somewhere between 2026 and 2032 — around 2028 at the median. Overlay the two studies and a single feedback loop emerges. The human commons dries up → training leans more on synthetic data → the risk of model collapse rises → factuality and diversity fall. The diagram below is that cycle.

Figure 5. The self-cannibalization feedback loop. An original diagram composed by Pebblous, based on Shumailov et al. (Nature 2024) and Epoch AI (2024).

Of course, real-world training pipelines mix synthetic and human data, reuse data, and use filtering to dodge the worst of it. This is no prophecy that collapse arrives tomorrow. But the direction is clear. The thinner the fresh human commons grows, the higher the ratio of synthetic data trying to replace it — and the gains once won through scaling are easily offset. "When Wikipedia runs dry, so does AI" is not a metaphor; it is a summary of this loop.

So who should pay to maintain the commons? Answers to this question are already being tried along several tracks. They share one thing: each is an attempt to dismantle the assumption that "data is free."

Table 2. A map of alternatives for funding the commons. Source: Wikimedia audit report (2025), Cloudflare (Jul 2025), industry press aggregated.

These alternatives operate at different layers. Wikimedia Enterprise has the data provider set its own price; Cloudflare collects a toll at the network gate; licensing deals and lawsuits fix the price through contracts and law. Rather than any one being the answer, together they point to the same thing: the era of "take it for free, without limit" is drawing to a close.

For any organization that uses data, this is not someone else's problem. If you run RAG, retrieval augmentation, or fine-tuning, you already lean implicitly on a free commons. So there are three things to ask now. Which public commons do our models and pipelines depend on, and how heavily? Are that commons' finances and contributions healthy? And is there room to hedge the risk of that source disappearing — through licensing or a referral arrangement? The problem of securing the provenance of data now spills over into the problem of managing that source's sustainability.

Policy, Statistics & Primary Sources

• 1.Miller, M. / Wikimedia Foundation (2025-10-17). "New User Trends on Wikipedia." diff.wikimedia.org. — Primary source for the −8% human-visit figure (Mar–Aug 2025)
• 2.Wikimedia Foundation (2025-04-01). "How crawlers impact the operations of the Wikimedia projects." diff.wikimedia.org. — Bots 65% of cost / bandwidth +50%
• 3.Wikimedia Foundation (2025). "Highlights from the Wikimedia Foundation's fiscal year 2024–2025 audit report." diff.wikimedia.org. — Total revenue $208.6M, Enterprise $8.3M
• 4.Wikimedia Meta-Wiki. "Fundraising/2024-25 Report." meta.wikimedia.org. — Total donors −2.4%, new donors −10.2%
• 5.Cloudflare (2025-07-01). "Introducing pay per crawl." blog.cloudflare.com. — Crawl-to-referral 14:1 / 1,700:1 / 73,000:1, default bot blocking

Academic Papers

• 6.Shumailov, I., Shumaylov, Z., Zhao, Y., Papernot, N., Anderson, R., & Gal, Y. (2024). "AI models collapse when trained on recursively generated data." Nature, 631(8022), 755–759. nature.com. — The canonical model-collapse paper
• 7.Villalobos, P., et al. / Epoch AI (2022, upd. 2024). "Will we run out of data? Limits of LLM scaling based on human-generated data." arXiv:2211.04325. — Public-text exhaustion 2026–2032 (median 2028)
• 8.Soboleva, D., et al. (2023). "SlimPajama-DC: Understanding Data Combinations for LLM Training." — Reference for Wikipedia's share of pretraining sets (2–4.5%)

Industry Analysis & Press

• 9.Profound (2025). "AI Platform Citation Patterns." tryprofound.com. — ChatGPT top-citation 47.9%, 680M citations analyzed
• 10.Engadget (2025). "Wikimedia says AI bots and summaries are hurting Wikipedia's traffic." engadget.com. — Hook article
• 11.Search Engine Journal (2025). "Wikipedia Traffic Down As AI Answers Rise." searchenginejournal.com. — Hook article
• 12.Similarweb / DataReportal (2025). "Digital 2025: exploring trends in Wikipedia traffic." — −23% over 3 years (third-party estimate, differing period/definition)
• 13.Fishkin, R. / SparkToro (2024). "In 2024, Less than One Third of Google Searches Still Send a Click." — Zero-click 58.5%
• 14.Pew Research Center (2026). "AI Overviews and news referral clicks." — 1% source-link click when an Overview is shown