wandered
Home/Resources/AI agents/Maintained reference

AI crawler user-agent directory: purpose, policy, and verification

AI-related user agents do not all perform the same job. This directory separates model-development crawlers, search crawlers, user-triggered fetchers, and special policy tokens.

15 minute referenceUpdated Evidence verified
Working definition

An AI crawler user-agent directory records documented machine identities, their stated purpose, robots token, available verification method, and analytics class. A match is evidence of a declared identity, not proof by itself.

  • A user-agent token states an identity; stronger policy decisions may require provider-supported verification.
  • Classify search, model-development, user-retrieval, and control tokens separately.
  • Keep automated requests out of human visitor, engagement, and conversion metrics.
Download the crawler directory CSV
01

Documented machine identities and control tokens

The same provider can publish several tokens because the requests serve different purposes.

OpenAI, Perplexity, Anthropic, and Google document search crawlers, model-development crawlers, user-requested retrieval, generic fetchers, or robots product controls. The declared purpose should be retained because it affects analytics, policy, and the business decision.

Google-Extended is included as a policy control rather than an HTTP user-agent identity. Always follow the linked current provider documentation before writing a WAF, robots, or legal policy.

TokenProviderPurpose classVerification / classification note
GPTBotOpenAIModel developmentUse current OpenAI documentation
OAI-SearchBotOpenAISearchUse current OpenAI documentation
ChatGPT-UserOpenAIUser-requested retrievalAutomated request, not a human visit
PerplexityBotPerplexitySearchProvider publishes live IP data
Perplexity-UserPerplexityUser-requested retrievalProvider publishes live IP data
ClaudeBotAnthropicModel developmentToken documented by Anthropic
Claude-SearchBotAnthropicSearchToken documented by Anthropic
Claude-UserAnthropicUser-requested retrievalAutomated request, not a human visit
Google-ExtendedGoogleRobots product controlNo separate HTTP user agent
GoogleOtherGoogleGeneric product crawlerFollow Google crawler verification
02

Classify purpose before volume

Ten thousand search-crawler requests and ten thousand attempted agent tasks do not describe the same activity.

Use purpose classes such as search or answer indexing, model development, user-triggered retrieval, monitoring, delegated action, and unknown. Store the provider’s stated purpose and the site’s observed behavior separately.

Unknown must remain a valid class. A browser-like user agent, spoofed token, or undocumented system should not inherit a trusted provider classification because its requests look similar.

PurposeUseful question
Search / answer indexingWhich public pages are being discovered?
Model developmentWhich content is being collected under current policy?
User retrievalWhich resources are fetched in response to a user?
Delegated actionWhich authorized tasks are attempted and completed?
UnknownWhat requires investigation or restriction?
03

Use layered identity verification

A user-agent match is useful evidence but can be spoofed.

Begin with the full raw user-agent string and exact token. Add provider-supported IP, reverse DNS, signature, or published-range verification where available and appropriate. Record the verification method, result, timestamp, and rule version.

Do not invent a network verification method when the provider does not publish one. In that case, retain lower confidence and combine the declared token with behavior, rate, authentication, and policy evidence for analysis.

  • Exact token match
  • Full raw user agent
  • Provider-supported network evidence
  • Behavior and request sequence
  • Authentication / authorization
  • Confidence and timestamp
04

Store a policy-ready machine record

A bot-name column cannot explain what happened or whether the request was allowed.

Record provider, token, purpose, raw user agent, source network evidence, verification result, requested host and path, method, response status, bytes, latency, robots decision, WAF decision, authentication state, and time under an approved retention policy.

Aggregate for routine reporting and restrict access to detailed logs. Redact query strings, credentials, payloads, and personal data that are not necessary for the stated operational purpose.

provider · token · purpose · confidence · verification_method · verification_result
host · path_class · method · status · bytes · latency · policy_decision · timestamp
05

Report machine activity with machine denominators

Machine requests belong beside human analytics, not inside it.

Show verified and declared actors, requests, machine sessions, pages requested, purpose, status, denial reason, latency, retries, and rate limits. For agents, add discovery, authorization, attempt, completion, and failure stages where the site can observe them.

Do not calculate human engagement, bounce, conversion, or audience share from machine requests. Compare page-level machine access with citations or referrals only as contextual evidence.

Machine metricDenominator
Request shareAutomated requests
Successful fetch rateEligible machine requests
Denied request ratePolicy-evaluated requests
Agent completion rateAuthorized attempted tasks
06

Match the control to the risk

robots.txt communicates crawl preferences; it is not authentication or a security boundary.

Use robots rules for compliant public crawl preferences, WAF and rate limits for capacity and abuse control, authentication for identity, authorization for permitted actions, and application checks for consequential tasks. Keep the business rationale and owner beside every rule.

Monitor allowed, denied, challenged, and failed requests after changes. A policy that cannot be observed and reviewed will drift as provider identities and product behavior change.

  • robots.txt preference
  • WAF and rate limits
  • Authentication
  • Action authorization
  • Sensitive-path protection
  • Post-change monitoring
07

Change log and correction policy

A maintained directory should show what changed and how the correction was verified.

Version 2.0 adds layered verification, a policy-ready log schema, purpose-specific analytics, and control guidance. The visible table and downloadable CSV share the same last-verified date.

Send corrections to hello@wandered.io with the provider documentation, exact token, stated purpose, verification method, and date. Do not submit credentials, private IP logs, personal data, or sensitive request payloads.

Evidence note

This directory supports classification and policy review; it must not be the sole control protecting sensitive resources.

Methodology and verification.

Last verified August 17, 2026. The page is updated when the underlying analytics or provider documentation changes materially.

  1. 01

    Used provider-owned crawler documentation and the Robots Exclusion Protocol as primary references.

  2. 02

    Separated declared identity, stated purpose, observed behavior, and independent verification instead of treating a token as proof.

  3. 03

    Mapped each purpose to an analytics denominator and appropriate control so the directory supports real operational decisions.

Change log
2.0

Added purpose taxonomy, layered verification, machine log schema, analytics denominators, control guidance, and correction requirements.

1.0

Initial major-provider token directory and downloadable CSV.

Verify the evidence.

Provider behavior and analytics definitions change. These are the primary references reviewed for this page.

Questions teams ask.

Is a user-agent match enough to verify an AI bot?+

No. It identifies a declared token. Use provider-supported network or signature evidence when available and retain confidence.

Why is Google-Extended listed without a user agent?+

It is a robots.txt product control token, not a separate HTTP request identity.

Are human AI referrals included?+

No. This directory covers machine identities. Human referral sources are maintained separately.

Should ChatGPT-User or Claude-User count as a visitor?+

No. They are documented automated fetchers used in response to a user request, not human browser sessions.

How should unknown automation be classified?+

Keep it Unknown with the available request, behavior, and confidence evidence. Do not assign a trusted provider solely from similarity.

Know which machines are using your site.

One launch email. No account, no weekly drip, no noise.