Research · October 2026

At least 43% of “AI bot” traffic is fake

We went through 15 days of web server logs from a group of small websites. Requests claiming to be AI crawlers outnumbered Googlebot about 7 to 1. But a large part of that “AI traffic” wasn't AI at all.

What we found

• 25,000 of 58,500 AI-labelled requests were probes for things like .env files, /proc/self/environ, ../ path traversal and random 20-character paths. Real crawlers fetch pages; they don't hunt for secrets.
• One scanner cycled through about 30 AI user agents within the same second: ClaudeBot, Claude-User, OAI-SearchBot, DeepSeekBot, KimiBot, Qwenbot, ChatGLM-Spider, Hunyuan and more.
• 1,300 requests said “Google-Extended”. Google says Google-Extended has no HTTP user agent; it's only a robots.txt token. Every one of those requests was fake.
• AI referrals were tiny: 71 visits from chatgpt.com in 15 days. A large “you.com” referrer count looked spoofed, so we left it out.

Share of requests that were attack probes, by claimed bot

User-triggered agents (the ones that fetch a page when someone asks an assistant) were impersonated the most, probably because sites are reluctant to block them.

Methodology and limits

• Data: nginx access logs, 2026-10-04 to 2026-10-18, one multi-site log plus five small WordPress sites.
• A request counted as fake when it claimed an AI user agent but asked for an attack path (secrets, shells, traversal, random probe paths). That only catches scanners that probe; scanners that fetch normal pages with a fake user agent slip through. So 43% is a lower bound.
• We could not verify by IP address on these servers: they sit behind a tunnel that hides the client IP. That is common, and it's why our log checker tells you when your log has the same problem and how to fix it.
• We store only aggregate counts. No IP addresses or site names are published.