In October 2025, the Wikimedia Foundation confirmed what reputation teams had been whispering about for months: human pageviews on Wikipedia fell roughly 8% year over year. The story ran everywhere within a day, usually under some version of "even Wikipedia is losing to AI."
We build and maintain Wikipedia pages for companies, so you might expect this article to bury that number in paragraph eleven. Most coverage written by people who sell Wikipedia work does. We are going to do the opposite: state it first, decode it honestly, and then walk through the 2026 data the doom headlines left out — because the same twelve months that took 8% of Wikipedia's human readers also made it the most-cited domain in ChatGPT answers, the corroboration layer that survived a 3-billion-entity purge of Google's Knowledge Graph, and the most heavily machine-read reference work in history.
The short version: your Wikipedia page's job changed. It used to be read by people. It is now read by machines that talk to people. This article ends with a buy-or-wait framework and a re-run of our five-year ROI math under that new reality.
The October 2025 statement, decoded
The 8% arrived with a backstory most headlines skipped, and the backstory matters more than the number.
Around May and June 2025, Wikimedia's analysts noticed unusually high traffic that looked human, much of it originating in Brazil. They investigated, rebuilt parts of their bot-detection systems, and reclassified months of data. The conclusion: a meaningful slice of what had been counted as human readers was actually bots engineered to evade detection. After the correction, human pageviews from March through August 2025 came out roughly 8% below the same months in 2024.
Marshall Miller, senior director of product at the Wikimedia Foundation, said it plainly: "We believe that these declines reflect the impact of generative AI and social media on how people seek information, especially with search engines providing answers directly to searchers, often based on Wikipedia content."
Three things in that sequence deserve decoding.
The decline is real, but it is a correction, not a cliff. Nobody watched readers walk out over a single quarter. The Foundation re-measured and found the human baseline had been lower than reported for a while. Direction: down. Magnitude: single digits.
Measuring "human" traffic is now genuinely hard. The restatement happened because bots got good enough to pass as people; every precise-sounding traffic figure carries that asterisk in 2026.
Read the second half of Miller's sentence again. The answers that intercept the click are "often based on Wikipedia content." The Foundation's own diagnosis is not that people stopped consuming Wikipedia. It is that they consume it somewhere else.
Where the readers went
The Foundation names two destinations, and both are worth taking at face value.
First, AI-mediated search. Google's AI Overviews, ChatGPT, and their peers answer factual questions directly, and those answers lean heavily on encyclopedic sources. The user gets the Wikipedia fact without the Wikipedia visit — the zero-click pattern that has eroded publisher traffic for years, now applied to the world's largest reference work.
Second, social video. Younger users increasingly pick up information on short-video platforms, which generate discussion rather than outbound clicks.
Notice what is absent from that list: any claim that the content lost its audience. The supply chain still starts at the same encyclopedia; the retail layer moved. The headlines collapse two different things — audience for Wikipedia's content, which moved rather than vanished, and traffic to Wikipedia's domain, which fell. The second is the Foundation's problem, because visits feed its donation funnel. It is mostly not yours, and conflating the two is how brands mis-decide.
The paradox, in one table
Here is 2025–2026 with humans and machines side by side.
| Signal | Direction | What the data shows | Source |
|---|---|---|---|
| Human pageviews on Wikipedia | Down | ~8% year-over-year decline after bot reclassification | Wikimedia Foundation, Oct 2025 |
| Share of US ChatGPT citations | #1 domain | 13.15% of citations in Q1 2026 — the largest single domain; Wikipedia plus Reddit exceed a quarter of all citations | 5W Citation Source Audit |
| Same metric, different method | #1 domain | ~14% peak in March 2025, settling near 7% — still the top-cited domain | Azoma |
| Presence across AI engines | High | Top-tier citation source across ChatGPT, Perplexity, Gemini, and Google AI Overviews | 5W |
| Bot share of Wikimedia's costliest traffic | Up | 65% of the most resource-intensive traffic comes from bots, vs. ~35% of pageviews | Wikimedia Foundation, Apr 2025 |
| OpenAI crawl-to-referral ratio (web-wide) | Extreme | ~1,217 pages crawled per referred visit in January 2025; ~1,091 by July | Cloudflare |
| Google Knowledge Graph entities | Purged | ~3 billion entities removed — roughly 6% of the graph | SOCi |
One honesty note, because two of those rows disagree. 5W's Q1 2026 audit puts Wikipedia at 13.15% of US ChatGPT citations; Azoma's tracking has it settling near 7% after a ~14% peak. Different samples, query sets, months, and definitions of "citation." We show both on purpose: AI citation measurement in 2026 is directional, not precise, and any vendor quoting a one-decimal number as settled fact is overselling. What every study agrees on is the ranking — Wikipedia is the most-cited domain in ChatGPT.
Read the table top to bottom and the paradox states itself. The only declining line is the human one, and it declines by single digits. Every machine-side line — citation share, crawl volume, infrastructure load — points up, by multiples rather than percentages.
How AI consumes Wikipedia: three different pipes
"AI reads Wikipedia" is true in three different ways, and they behave differently. We cover the mechanics in depth in Why Wikipedia Is ChatGPT's #1 Source; here is the structural summary.
Pipe one: the training corpus. Before a model answers anyone, it is trained on a frozen snapshot of text in which Wikipedia is one of the most heavily represented sources — partly because it is large and reliable, mostly because it is freely licensed and therefore duplicated across thousands of mirrors and derivative datasets that get crawled too. Facts ingested this way become weights; no pageview is ever recorded. The model does not look up your founding date; it remembers it — as of whatever your article said at training cutoff.
Pipe two: live retrieval. When a question needs current information, systems like ChatGPT search the web at answer time and ground the response in fetched documents. Wikipedia surfaces constantly here because it is authoritative, cleanly structured, and easy to extract facts from — and this is the pipe the citation studies actually measure. A retrieval fetch is a bot request, not a human visit.
Pipe three: knowledge-graph grounding. Entity systems — Google's Knowledge Graph above all — resolve which "Mercury" or "Apple" a query means by checking structured records: Wikidata identifiers, Wikipedia articles, and the cross-references between them. This pipe runs on structured statements rather than prose, and it is what builds Knowledge Panels and disambiguates brands inside AI answers.
The machine readership is formal enough to have a price list. Wikimedia Enterprise, the Foundation's commercial API, sells SLA-backed real-time feeds of Wikipedia to large reusers — Google became the first announced paying customer in 2022. The load shows at the infrastructure level too: by April 2025, Wikimedia's engineers reported that 65% of their most expensive traffic came from bots that generate only about 35% of pageviews. Companies now pay to read Wikipedia without visiting it.
None of the three pipes registers as a human pageview. A page can shed a tenth of its visitors while multiplying its effective reach through answers, and standard analytics will only ever show you the first number.
The 3-billion-entity purge: scarcity raises the value of what survived
While attention was on the traffic story, a quieter development moved the economics more. Through late 2025, Google removed roughly 3 billion entities from its Knowledge Graph — about 6% of the graph, a cleanup tracked by entity-SEO researchers and reported by SOCi.
What got purged, broadly: thin entities — auto-generated, weakly corroborated, single-source records the graph had accumulated over years. What survived: entities anchored by strong corroboration, with the Wikipedia–Wikidata pair the strongest known anchor in the public web's entity layer.
The strategic read is straightforward scarcity economics. When a graph prunes billions of nodes, each surviving node carries more resolving weight: it is what search and AI systems disambiguate against when they decide which company a name refers to and whether it merits a Knowledge Panel at all. Pre-purge, a thin entity presence was cheap to acquire and easy for machines to ignore. Post-purge, the bar is higher — and clearing it is worth more, precisely because fewer entities clear it. An anchor that survived the cut is a scarcer asset in 2026 than the same anchor was in 2023.
Is Wikipedia institutionally at risk?
An asset case built on Wikipedia should answer the uncomfortable question directly rather than wave it off.
The near-term answer is no. The Foundation runs on donations, not pageview advertising, so there is no ad-revenue collapse mechanism. Revenue was roughly $185 million in fiscal 2023–2024, the endowment crossed $100 million back in 2021, and Wikimedia Enterprise adds a direct line to monetize the machine readership described above.
The honest long-term concern is the one the Foundation itself named in the October post: "With fewer visits to Wikipedia, fewer volunteers may grow and enrich the content, and fewer individual donors may support this work." Readers are the top of the funnel for both editors and donors. An 8% dip does not threaten that funnel today; a decade of compounding dips would. That is a real institutional question for the 2030s, with mitigations already in motion — the Enterprise revenue line and the Foundation's stated strategy of using AI to support volunteers rather than replace them.
What the data does not support is a shutdown narrative. Anyone selling you urgency on the theory that Wikipedia is dying is selling fiction; anyone telling you Wikipedia stopped mattering because its traffic dipped has not looked at the citation tables. Both errors cost money.
The decision framework: when a page is worth more in 2026 — and when to wait
If the page's job changed, the buying logic changes with it. Here is the matrix we use in scoping conversations.
| Your situation | 2026 verdict | Why |
|---|---|---|
| Buyers research you through ChatGPT, AI search, or due diligence (B2B, finance, health, enterprise) | Stronger case than 2023 | The answer layer is assembled from encyclopedic and entity sources; one maintained page feeds all three pipes |
| Your branded search shows AI Overviews or a Knowledge Panel | Stronger case | The description of you now happens off your site; accuracy at the source layer is the only lever you have |
| Substantial independent press coverage exists | Buy | Notability is the gate and coverage is the raw material; eligibility is likely and durable |
| Coverage is thin or borderline | Wait — audit first | A deleted page costs more than no page; a notability audit prices the survival question before you fund the asset |
| You need traffic or pipeline this quarter | Wait — wrong instrument | A Wikipedia page is a trust asset; under machine readership it refers fewer clicks than ever, by design |
| Consumer-impulse brand discovered through social video | Usually wait | Your buyer's journey rarely touches the encyclopedic layer at all |
Note the column that is missing: traffic to you. It was never the point of a Wikipedia page, and after 2025 it is less the point than ever. A provider pitching a page as a click source is describing a product that does not exist.
Re-running the five-year math under machine readership
Our five-year ROI analysis priced the entry scenario at about €4,030: a professionally created English company page from €1,930, plus baseline annual support from €420 per year. Those are WikiBusines published prices; over five years they work out to roughly €2.20 a day. The question is what the traffic news does to that math.
Under the old mental model — value per human visit — the damage is easy to compute and easy to absorb. Take a clearly labeled hypothetical: a mid-sized company article drawing 500 human reads a month collects about 30,000 reads over five years, around €0.13 per read at the entry price. Apply the 8% decline and it becomes roughly €0.15. If the case for the asset rested on per-visit economics, it would have been a weak case in 2023 too.
But the denominator changed. The page's consumption now happens upstream of any visit: in training snapshots that memorize it, retrieval fetches that ground tonight's answers in it, and entity lookups that resolve your company against it. The unit of value is no longer the human read; it is the AI answer that describes you correctly because the record it drew on was accurate. That unit is uncountable from your analytics — Cloudflare's measurements suggest the scale, with OpenAI crawling on the order of 1,100–1,200 pages for every visit it refers — and every measurable proxy for it grew while human traffic shrank. Fixed numerator, larger denominator: the cost per unit of presence fell.
Two honest costs come with the shift. Attribution gets worse — you will never see an AI answer in your referral logs, so the audit question becomes "do the major models describe us accurately," not "how many sessions did Wikipedia send." And maintenance matters more, not less: a stale fact on a machine-read page no longer just misleads the occasional visitor — it gets memorized at the next training cutoff and repeated confidently until corrected. Drift compounds now. The support line in the budget is not optional under machine readership; it is the part that protects everything else.
FAQ
Is Wikipedia shutting down?
No. The October 2025 announcement was a traffic restatement, not a financial event. The Foundation runs on donations, reserves, a nine-figure endowment, and now Wikimedia Enterprise revenue from the same AI economy that reads it. The real long-term question — volunteer and donor recruitment over the next decade — is worth watching and nowhere near a shutdown.
Will AI replace Wikipedia?
AI consumes Wikipedia — the most-cited domain in every ChatGPT study, at 13.15% of US citations in the latest audit. Replacing it would starve the replacement: models trained on model output instead of a maintained, human-curated corpus degrade, which is why AI companies pay the Foundation for structured feeds rather than building a substitute. What AI is displacing is Wikipedia's traffic, not its function — and the function became more load-bearing, not less.
Do fewer human visits make my page less valuable?
Only if you think the visits were the value. They were a proxy for it. The value was always being the record that search engines, journalists, due-diligence teams, and now AI systems resolve your company against — and that consumption grew while the proxy shrank. Fewer people read the page; far more people hear from it.
If the framework says buy, the work starts where it always did: a page sourced well enough to survive and watched well enough to stay accurate. Wikipedia page creation explains how we scope that — including the 93% success rate and the refund terms we publish, because in this market priced risk is the only honest offer. If the framework says wait, wait properly: build independent coverage first, and let the audit tell you when the math flips.