ChatGPT's 404 rate is 6.7x Google's - but the traffic number may surprise you

New evidence from the Ahrefs study quantifies how often AI assistants link users to pages that return 404 errors compared with Google Search. Based on 16 million URLs, the research also estimates how much traffic assistants currently send to websites so marketers can size the issue relative to organic search.

AI search 404 rate - executive snapshot

Clicked 404 rates: ChatGPT 1.00% vs Google Search 0.15% (about 6.7x higher). Other assistants ranged 0.12-0.58% for clicked links.
Mentioned links: 2.38% of ChatGPT-mentioned URLs returned 404 vs 0.84% among Google’s top results.
Traffic share: AI assistants drove roughly 0.25% of visits to analyzed sites vs 39.35% from Google Search.
Drivers of bad links: a mix of stale pages that once existed and fabricated, pattern-based URLs that never existed.

Implication for marketers: The broken-link risk from assistant referrals is real but low-volume. Prioritize helpful 404 pages and add selective redirects where there is meaningful traffic.

Method and source notes

Ahrefs analyzed link quality and traffic from AI assistants and Google using a corpus of 16 million URLs. The study measured the share of URLs returning HTTP 404s among (a) links that users clicked from assistant outputs and Google results, and (b) links merely mentioned by assistants or shown in Google’s top results. It also estimated the share of total site traffic attributable to each channel across the observed sites.

Who, what, when: Ahrefs, “How often do AI assistants hallucinate links?” (2025) - methodology summarized on the Ahrefs blog. See the Ahrefs study.
Sample: 16,000,000 URLs; multiple assistants measured (ChatGPT, Claude, Copilot, Perplexity, Gemini, Mistral) plus Google comparisons.
Metrics: Percent of URLs returning HTTP 404 (clicked vs mentioned), and channel traffic share.
Caveats: “404 at time-of-check” can fluctuate as pages are fixed or redirected. Assistant models and search indices change quickly. Differences in how clicked links are detected and how assistants expose links may affect comparability. Site sample composition can bias traffic-share estimates.

Findings: AI assistants hallucinate links more than Google Search

Ahrefs study: AI assistants link to 404 pages more often than Google Search — Across assistants, clicked links were more likely to return 404 than Google Search referrals.

Clicked broken-link rates

ChatGPT: 1.00% of clicked URLs led to 404.
Claude: 0.58%.
Copilot: 0.34%.
Perplexity: 0.31%.
Gemini: 0.21%.
Mistral: 0.12% (lowest among assistants measured).
Google Search: 0.15%.

Mentioned vs clicked discrepancy

Among all URLs mentioned by assistants, a higher share were 404s than among clicked links. For ChatGPT, 2.38% of mentioned URLs returned 404 vs 1.00% of clicked URLs. For Google’s top results, 0.84% were 404s vs 0.15% among clicked links. This suggests users self-filter some bad links by not clicking them, but a meaningful portion still leads to dead ends.

Drivers of 404s

Stale URLs: Pages that existed historically but returned 404 at time of analysis, often cited when assistants rely on older snapshots.
Fabricated patterns: Assistants generated plausible-looking slugs (for example, “/blog/internal-links/”) that never existed, producing 404s when clicked.

Traffic magnitude and model differences

Across the analyzed sites, AI assistants accounted for roughly 0.25% of visits. Google Search contributed 39.35% - roughly two orders of magnitude more traffic, indicating the absolute volume of assistant-driven 404 visits is currently small relative to organic search. While Mistral showed the lowest clicked 404 rate (0.12%), Ahrefs observed it also sent the least traffic among assistants, limiting aggregate impact.

External context

In March, Google’s John Mueller anticipated a slight uptick in hallucinated links being clicked over a 6 to 12 month window and recommended focusing on helpful 404s rather than chasing accidental traffic spikes while the ecosystem stabilizes, per Search Engine Journal.

Interpretation and implications for marketers

AI assistants produce markedly higher broken-link rates than Google on both mentioned and clicked URLs, with ChatGPT outlier-high on clicks (1.00% vs Google’s 0.15%). Ensure a robust 404 UX.
Assistant traffic share is small today (~0.25% vs 39.35% for Google), so the operational impact of assistant-driven 404s is limited for most sites.
Two mechanisms - stale references and fabricated URL patterns - explain most assistant-driven 404s. Sites with predictable slug structures may be more exposed to fabricated patterns.
Model improvements may reduce 404 rates, but churn and index latency can keep stale links in circulation. Monitoring referral sources and 404 logs remains prudent.

Practical actions

Design 404 pages to help users complete tasks: add on-site search, top categories, and clear paths to active equivalents.
Add redirects only for fabricated or stale URLs that attract non-trivial traffic to avoid redirect sprawl and misrouting.
Track assistant referrals separately (UTMs or referrer patterns) to quantify impact and identify recurring hallucinated slugs.

Contradictions and gaps

Coverage and representativeness: The site mix and verticals are not fully detailed. Assistant usage varies by audience and geography, which could shift rates for specific industries.
Time sensitivity: “404 at time-of-check” can overstate persistent errors if publishers quickly fix content or add redirects, and the reverse can also occur. Longitudinal stability is unknown.
Assistant interface differences: Some assistants link sparingly or cite sources differently, which can suppress or inflate click-level 404 rates. Normalizing across interfaces is challenging.
AI-generated content prevalence: Detection methods and false positives can materially affect any percentages cited. Use caution when extrapolating.