Skip to main content
Anyone can send a request with User-Agent: GPTBot. Detection tells you who a request claims to be; verification checks that claim against network evidence the provider publishes. Quotable never trusts the user agent alone.

The three outcomes

Unverified is not spoofed. It means the evidence wasn’t there: the log didn’t carry the client IP, the provider publishes no ranges, or a DNS lookup timed out. Only a direct contradiction with published ranges marks a request as spoofed.

Methods

Published IP ranges

OpenAI, Perplexity, Google (its crawlers and Google-Agent), Microsoft (Bingbot), DuckDuckGo (DuckDuckBot), Mistral (MistralAI-Index), Common Crawl (CCBot) and Yahoo (Slurp) publish machine-readable lists of the IP ranges their agents use. Quotable checks the request’s source IP against the list for the agent it claims to be:
  • Inside a range: verified, with the method ip range and the name of the range set that matched.
  • Outside every range the provider publishes for that agent: spoofed.
  • No range set loaded for that agent: unverified.
Range lists are refreshed from the providers’ published documents regularly. Changes are reviewed before they affect verdicts, so a provider-side glitch can’t suddenly turn real crawlers into spoofs. For Anthropic (ClaudeBot, Claude-SearchBot, Claude-User) and DuckDuckGo (DuckAssistBot), which don’t publish a machine-readable list, Quotable maintains the range sets.

Forward-confirmed reverse DNS

Apple, Amazon, You.com, Yandex, Baidu and Huawei document hostnames instead of stable ranges, and Google, Microsoft, Common Crawl and Yahoo document both. For these agents, Quotable:
  1. Looks up the reverse DNS name of the source IP.
  2. Keeps only names under the provider’s documented domains: googlebot.com (and google.com for Googlebot), search.msn.com, applebot.apple.com, crawl.amazonbot.amazon, crawl.commoncrawl.org, crawl.yahoo.net, search.you.com, yandex.ru, yandex.net or yandex.com, baidu.com or baidu.jp, petalsearch.com.
  3. Resolves that name forward and checks that it returns the same IP.
If both directions agree, the request is verified with the method fcrdns. If they don’t, or DNS fails or times out, the request stays unverified. A DNS problem can never turn a request into a spoof.

No published method

Meta (meta-externalagent, Meta-ExternalFetcher, Meta-WebIndexer), Mistral (MistralAI-User, MistralAI-Training), Microsoft (AdIdxBot, MicrosoftPreview), Diffbot (Diffbot, Diffbot-User), Allen Institute for AI (AI2Bot), ByteDance (Bytespider) and Cohere (cohere-ai) have no ranges or hostnames in the registry. Their requests are always unverified, and the user agent is the only evidence.

Published, but not supported yet

Some providers publish a check that Quotable doesn’t read yet. Their requests are unverified for now, never spoofed:
  • Claude-Agent: Anthropic publishes its addresses in a format Quotable doesn’t read yet.
  • Amzn-SearchBot and Amzn-User: Amazon publishes the addresses inside a web page.
  • Kagibot: Kagi lists a handful of addresses on its crawler page.
  • ExaSearchBot and ExaUser: Exa signs each request instead of publishing addresses.
  • Google-GeminiNotebook: Google publishes shared ranges for all its user-triggered fetchers, including ones run by its cloud customers, so a match would not prove which fetcher sent the request.

What verification needs from your logs

Verification needs the client IP of each request. Every built-in connector sends it, except where noted on the connector page. If you build your own sender with Custom HTTP, include client_ip; without it, every event stays unverified. The IP is used in memory for the check and then discarded. The stored event keeps only the status, the method and the name of the matched range set. Make sure the IP in your logs is the visitor’s IP, not the IP of a proxy or load balancer in front of your origin. If your logs record your CDN’s address, every request will be unverified (or spoofed, for providers with published ranges). Connect the CDN’s logs instead of your origin’s.

When verdicts change

Each event records which version of the agent registry classified it. When the registry or a range set changes, Quotable can relabel existing events without the IP or user agent, which were never stored:
  • A verified event whose range set was withdrawn becomes unverified.
  • A spoofed event for an agent that no longer has any published range becomes unverified.
  • Intent follows the current registry entry for the agent.
Nothing is ever upgraded to verified after the fact, because the evidence to do so is gone. An update that adds an agent counts it from that update on. Requests the agent made earlier were not stored, so there is nothing to relabel. The same holds when an update gives part of an agent’s traffic its own entry, as the 11 November 2026 update did for Googlebot-Image, Googlebot-Video and AdIdxBot: older events keep the Googlebot or Bingbot label, because the user agent that would tell them apart was never stored.

Intent from behaviour

Intent normally comes from the purpose the provider documents. One behaviour overrides it: when an agent that fetches for users (for example ChatGPT-User) sends POST, PUT, PATCH or DELETE, the request is labelled Agent / action, because it’s acting, not reading. Claude-Agent and Google-Agent are browsers that act for a user by design, so every request they make is labelled Agent / action, whatever the method.

Next steps

Recognized agents

The full registry with each agent’s method.

FAQ

Common questions about verification and counts.
Last modified on November 18, 2026