Trust & security

How to tell if a visitor really is ChatGPT, and not something pretending

Listen · 9 min

Narrated by an AI voice. Code examples are on the page.

Short answer

A user-agent string proves nothing, because anyone can type it. A stronger check is to match the visitor's IP address against the lists OpenAI publishes for its bots. The strongest check, for ChatGPT agent, is a cryptographic signature that your site or CDN verifies as the request arrives.

Key takeaways

  • The user-agent string is a claim, not proof.
  • OpenAI publishes IP lists for GPTBot, OAI-SearchBot and ChatGPT-User.
  • Signed requests can be checked only live, not from a log line.

Can you trust the user-agent string?

No. The user-agent string is the weakest check, because the sender writes it. A user-agent is a line of text in each request that says what software is asking. Nothing stops any script from sending any text it likes.

OpenAI documents the strings its bots use. For example, ChatGPT-User identifies itself as compatible; ChatGPT-User/1.0; +https://openai.com/bot. Anyone can copy that line. A request that carries it has told you what it wants to be called. It has not told you who it is.

Use the user-agent for a first guess and for grouping your logs. Do not use it to grant access, skip a rule or trust a request.

Does the IP address prove it is ChatGPT?

It is much stronger, but it still has limits. Real traffic from OpenAI's bots comes from addresses OpenAI controls. A fake cannot easily send traffic from those addresses.

OpenAI publishes a JSON list of IP ranges for each of its main bots:

Bot What it does Published list
GPTBot Collects content for training OpenAI's models https://openai.com/gptbot.json
OAI-SearchBot Surfaces sites in ChatGPT search https://openai.com/searchbot.json
ChatGPT-User Visits a page when a person asks ChatGPT to https://openai.com/chatgpt-user.json

To use a list, compare the visitor's IP address with the ranges in the right file. If the address is inside a range and the user-agent matches, the request is very likely genuine. If the user-agent says ChatGPT-User but the address is not on the list, the request claims to be ChatGPT but fails verification.

There is an older method too. Google documents a "reverse DNS" check for its own crawlers. You look up the hostname behind the IP address, confirm it belongs to the company, then look that hostname up again and confirm it points back to the same address. Google also publishes IP range files for larger sites. OpenAI's bot documentation points to its published IP ranges, so that is the method to use for OpenAI.

Two cautions apply. First, the lists change, so fetch them regularly instead of copying them once. Second, an IP check says the request came from OpenAI's network. It does not say which product or person is behind it.

You can read more about the bot behind the third list in the ChatGPT-User directory entry.

What are signed requests and the Signature-Agent header?

A signed request carries a digital signature that only the real sender can produce. It is the strongest check, because a fake cannot copy it without the sender's private key.

The standard is called HTTP Message Signatures, published by the IETF as RFC 9421. The approach built on it for bots is often called Web Bot Auth. For the full background, see Web Bot Auth and signed agent requests.

OpenAI's documentation says ChatGPT uses Web Bot Auth to identify certain automated browser requests, and that this covers signed cloud browser requests, not web crawlers. A signed request carries these headers:

  • Signature and Signature-Input: the signature itself, and a description of what it covers.
  • Signature-Agent: set to "https://chatgpt.com", with the quotation marks.

The Signature-Agent header tells you where to find the sender's public keys. OpenAI lists them at https://chatgpt.com/.well-known/http-message-signatures-directory. Your site, or your CDN, fetches the key and checks the signature against it.

OpenAI's steps for sites that do not recognize this traffic automatically are:

  1. Confirm the Signature-Agent header matches exactly, including the quotation marks.
  2. Fetch the matching public key from the key directory.
  3. Check the Signature and Signature-Input headers against RFC 9421.
  4. Only after that, apply any allowlist exception.

The signature proves the request came from OpenAI's system. It does not tell you which person started the task. That is a separate question, and one your own rules have to answer.

Can you spot ChatGPT agent or Atlas by user-agent?

Do not count on it. OpenAI's page on signed requests does not give a user-agent string for them. It identifies them by the signature headers instead.

We could not find OpenAI documentation that says how ChatGPT Atlas, or agent mode inside it, can be recognized by user-agent. Third-party write-ups disagree on the details, so we leave the question open. If your site needs to know, rely on the signature, not the string.

Why can't you check a signature from a log line?

Because a signature belongs to one request, and a log line is only a summary of it. Most access logs keep the time, the address, the path and the user-agent. They do not keep the signature headers, and they rarely keep every detail the signature covered.

A signature is also tied to when it was made. RFC 9421 lets a signer include a creation time, an expiry time and a unique value, so a copied signature can be rejected later. Replaying a stored one proves little.

So the check has to happen live, as the request arrives. Your CDN, web server or a plugin verifies the signature at that moment and records the result next to the request. Later you read the recorded result: verified, failed or unsigned. You cannot recreate it afterwards.

What does a fake look like?

Here is an invented example of a request that claims to be ChatGPT but fails verification. It is for illustration only.

GET /products/blue-widget HTTP/1.1
Host: shop.example.com
User-Agent: Mozilla/5.0 (compatible; ChatGPT-User/1.0; +https://openai.com/bot)
Accept: text/html

Here is what is wrong with it:

  • The user-agent claims to be ChatGPT-User, which proves nothing alone.
  • The request comes from an IP address that is not in the ChatGPT-User list.
  • There is no Signature-Agent or Signature header. That alone is not suspicious for a plain ChatGPT-User visit, because OpenAI's signing guidance covers its cloud browser. It means this request offers nothing stronger than a claim.

The word to use is "unverified." You do not need to call anyone malicious. A developer testing a script, a monitoring tool and a scraper can all send this exact request.

A checklist you can use

  • Treat the user-agent as a claim, never as proof.
  • Compare the visitor's IP with OpenAI's published list for the bot it claims to be.
  • Refresh those lists on a schedule.
  • For ChatGPT agent, expect the Signature-Agent header "https://chatgpt.com" and verify the signature against OpenAI's key directory.
  • Verify signatures live. Never try to verify them from stored logs.
  • Record the result of each check with the request: verified, failed or unsigned.
  • Keep rules for verified bots separate from rules for unverified traffic that only claims a name.

How Sill helps

Sill shows you which AI agents and bots visit your site and the pages they read. It verifies signed agent requests (Web Bot Auth) on Cloudflare and WordPress installs, so the check happens live as the request arrives. You can start with the free readiness scan at /scan, no account needed.

Questions people ask

Can someone fake the ChatGPT-User user agent?

Yes. A user-agent is just a text header the sender chooses. Any script can send the exact string OpenAI documents, so it should never be your only check.

Where does OpenAI publish its bot IP addresses?

OpenAI publishes JSON files at openai.com/gptbot.json, openai.com/searchbot.json and openai.com/chatgpt-user.json, one per bot.

What is the Signature-Agent header?

It is a request header that names who signed the request. OpenAI's documentation says signed ChatGPT agent requests carry a Signature-Agent header set to "https://chatgpt.com", including the quotation marks.

Can I check a signature later in my server logs?

Not reliably. A signature covers parts of one specific request and is meant to be checked as that request arrives. A stored log line usually lacks the signature headers and the details needed to re-check it.

Sources

  1. OpenAI: Overview of OpenAI crawlers · checked 15 Nov 2026
  2. OpenAI: Help ensure cloud browser access to your site · checked 15 Nov 2026
  3. IETF RFC 9421: HTTP Message Signatures · checked 15 Nov 2026
  4. Google: Verifying requests from Google's crawlers and fetchers · checked 15 Nov 2026

Written by The Sill Signals Team. Last reviewed 15 Nov 2026.

Share this articleLinkedInXBlueskyEmail

Keep reading

See which AI agents visit your site, and decide what they may do.

Scan your site free