Wikipedia traffic drops 8% after bot cleanup - why AI isn't sending clicks

Wikipedia is reporting fewer human pageviews year over year after reclassifying stealth bots, while AI platforms draw on its content at large scale. The available data connects three dynamics: a measured decline in reported human visits, rising AI bot bandwidth scraping content, and high Wikipedia representation in AI citations. This report summarizes what changed in measurement, how AI answer surfaces may be affecting referral traffic, and practical implications for publishers and marketing leaders.

Wikipedia Traffic Down As AI Answers Rise — Wikipedia human pageviews fell about 8% YoY as AI answer surfaces and bot scraping increased.

Wikipedia traffic - Executive snapshot

The Wikimedia Foundation reported a decline in human pageviews after retroactively filtering previously undetected bots. In parallel, AI systems cite Wikipedia at scale. Research from Profound analyzing 680 million AI citations shows Wikipedia is a leading source for ChatGPT and materially present in Google AI Overviews. WMF telemetry, highlighted by IBM, also reported a 50% surge in bandwidth from AI bots since January 2024, largely from computer vision training scrapers. Separately, Wikimedia Enterprise formalizes SLA-backed feeds for high-volume reusers, reflecting a shift toward structured data delivery over pageview referrals.

Human pageviews down ~8% YoY for Mar-Aug after bot reclassification
680 million AI citations analyzed across platforms
Wikipedia share of ChatGPT top-10 citations: 47.9%
Wikipedia share of Google AI Overviews top-10 citations: 5.7% - Reddit 21.0%, YouTube 18.8%
Bandwidth from AI bots up ~50% since Jan 2024

Implication for marketers: high citation in AI answers does not guarantee referral traffic. Expect more zero-click information retrieval to compress organic visits even for authoritative sources.

Method and source notes - AI search, bot detection, and AI citations

The Wikimedia Foundation (WMF) investigated traffic anomalies beginning in May 2025 and determined some human-like visits were stealth bots designed to evade detection. Updated bot rules were applied retroactively to March-August 2025, producing an estimated 8% decline in human pageviews versus the same months in 2024. WMF cautions that cross-period comparisons should be treated carefully while detection logic evolves. Source

Research from Profound aggregates 680 million citations across AI platforms and reports top-10 citation share for services including ChatGPT and Google AI Overviews. The study reflects observed outputs rather than platform-internal logs, and top-10 share is a constrained view of overall attribution.

IBM highlights WMF telemetry showing a ~50% increase in bandwidth from AI bots since January 2024, with much of the growth attributed to scrapers training computer vision models. Bandwidth is not equivalent to visits but indicates higher automated consumption of Wikimedia content. Source

Wikimedia Enterprise (launched 2021) provides commercial, high-availability data feeds and APIs for reusers like search and AI companies, signaling an institutional pathway for data access beyond public crawling.

Key limitations

Evolving bot detection complicates clean year-over-year trend lines.
AI citation shares reflect visible references only and may miss uncredited use.
Bandwidth metrics capture load rather than discrete request counts or user impact.

Findings - generative AI answers, bot scraping, and citation patterns

WMF’s reassessment shows stealth bots inflated human-like sessions in 2025, especially in May. After reclassification, human pageviews for March-August were about 8% lower than the same months in 2024. WMF links the broader downward trend to shifts in information access - search engines and social platforms increasingly answer queries directly, often with content derived from Wikipedia, reducing click-through to the source site.

Evidence also shows Wikipedia is a leading source for AI-generated answers. Among ChatGPT’s top-10 cited domains, Wikipedia holds 47.9% of share. For Google AI Overviews, Wikipedia accounts for 5.7%, while Reddit and YouTube are cited more heavily in that surface. This indicates platform-specific sourcing patterns.

Concurrently, WMF reports a ~50% increase in bandwidth from AI bots since January 2024, especially from computer vision training scrapers, raising infrastructure demand without corresponding user engagement. WMF’s Wikimedia Enterprise product provides a commercial channel for structured access to Wikimedia data for large reusers, reflecting a strategic response to the shift toward machine consumption.

Taken together, these signals point to a structural decoupling: Wikipedia’s content remains central to AI answer generation while direct human visits decline or grow more volatile under stricter bot filtering.

Interpretation and implications - SEO, PPC, content strategy, and measurement

Likely

Zero-click answers and AI Overviews are shifting a share of informational search away from click-throughs, reducing organic referral traffic even to top authorities. Plan for lower SEO traffic per query as answer surfaces expand.
Automated consumption will continue to rise, increasing server load and cost without proportional user value unless mediated by licensable access paths such as SLA-backed feeds.
Being highly cited by AI models does not ensure brand-site sessions. Measure presence and influence within AI outputs separately from web traffic metrics.

Tentative

Differences in citation mixes across AI surfaces suggest that content formats and communities (encyclopedic vs. discussion or video) may gain or lose exposure depending on the interface. Channel-specific packaging may influence inclusion in AI answers, but traffic uplift is uncertain.
Over-time comparisons of site traffic will require baselines that account for evolving bot detection. Treat 2025 YoY comparisons as provisional until classification stabilizes.

Speculative

If AI answer surfaces keep absorbing demand without referral, publishers may need alternative value capture - licensing, partnerships, direct subscriptions, or community retention - to sustain high-cost content categories. Wikimedia Enterprise illustrates one institutional approach; applicability varies by scale and data utility.

Operational measurement implications

Expand KPIs to include AI answer presence, citation share, and brand mention quality alongside sessions, CTR, and rankings. Use controlled experiments to isolate the impact of AI answer units on click-through where possible.
Budget and forecast models should assume reduced incremental SEO traffic from net-new informational queries and reallocate to channels that produce attributable engagement when warranted by ROI.

Contradictions and gaps - data quality, comparability, and AI Overviews

The 8% decline depends on revised bot rules. Year-over-year comparisons may be distorted and absolute baselines remain uncertain.
Profound reports top-10 citation share, not total citation volume by domain, and AI platforms sometimes omit citations. Window and coverage affect generalizability.
The 50% bandwidth increase reflects data transfer volume, not the number of distinct bots or requests, and does not directly translate to pageview displacement.
The contribution of generative AI versus broader search UX changes to Wikipedia’s traffic decline is not isolated in WMF reporting.

Data appendix - key figures referenced

Human pageviews: ~8% YoY decline for Mar-Aug vs. 2024 after bot reclassification; stealth bots identified inflating May traffic; caution on cross-period comparisons.
AI citations dataset: 680,000,000 citations analyzed; Wikipedia share of ChatGPT top-10 citations 47.9%; Wikipedia share of Google AI Overviews top-10 citations 5.7%; Reddit 21.0%, YouTube 18.8% in Google AI Overviews’ top-10.
AI bot bandwidth: ~50% increase since Jan 2024; primary driver cited as computer vision training scrapers.
Data access model: Wikimedia Enterprise launched 2021 for commercial, high-volume reusers such as search and AI firms.