Short answer
An AI agent's identity is only as strong as the check behind it. A user-agent string is a claim anyone can send, IP lists and reverse DNS prove which network a request came from, and a cryptographic signature (Web Bot Auth) proves which operator sent it. None of these tells you who the person behind the agent is, so your own rules still decide what each agent may do.
Key takeaways
- A user-agent name is a claim; IP and DNS checks confirm the network.
- Signed requests are the strongest proof of which operator sent a request.
- Some agents look like ordinary Chrome and carry no bot name at all.
What does AI agent identity mean for a website owner?
It means answering one question with evidence: which company or service really sent this request? Agent identity is not the name a visitor gives itself. It is the name you can confirm.
There are three separate questions hiding inside "who is this?":
- What does the visitor say it is? This is the user-agent string, and it is only a claim.
- Which operator actually sent it? This is what IP lists, reverse DNS and signatures can confirm.
- Who is the person behind it, and what did they approve? No network check answers this. It is a matter for the agent's operator, the payment system and your own rules.
Most mistakes come from treating the first answer as if it were the second. A log line that says "GPTBot" or "Googlebot" tells you what the sender typed, not who the sender is.
The rest of this guide goes through each signal, how far you can trust it, and what to do with the result.
What kinds of AI visitors reach your site?
AI visitors fall into three broad groups, and each one identifies itself differently.
Crawlers. These visit on a schedule, without a person asking at that moment. They collect pages for search indexes or for model training. Examples include GPTBot, ClaudeBot and Googlebot. They send a fixed bot name in the user-agent string, and their operators usually publish ways to verify them.
User-triggered fetchers. These fetch a page because a person asked a question right now. Examples include ChatGPT-User, Claude-User and Perplexity-User. They also carry a bot name. Because a person started the request, they often do not follow robots.txt the way crawlers do. Google says its user-triggered fetchers "generally ignore robots.txt rules", for example.
Google now documents an agent in this group too. Google-Agent is used by "agents hosted on Google infrastructure to navigate the web" when a user asks them to act. Its user-agent includes the token compatible; Google-Agent, and Google publishes its IP ranges in a file called user-triggered-agents.json. Our Google-Agent explainer covers it in detail.
Agentic browsers. These drive a full web browser, clicking, typing and filling in forms. Some run in the cloud, like ChatGPT agent. Others run on the person's own computer. Anthropic's Claude in Chrome, for example, is installed from the Chrome Web Store and works "using your existing logins".
Agentic browsers are the hardest to identify. Cloudflare wrote in its Web Bot Auth announcement that "agents often use the Chrome user agent", a practice it calls discouraged, because there is no clear way to declare a semi-automated browser in that header. So an agentic browser can look exactly like a normal shopper using Chrome.
Which identity signals can you trust, and how much?
Each signal answers a narrower question than it first appears to. Here is how they compare.
| Method | What it proves | How hard to fake | Main limits |
|---|---|---|---|
| User-agent string | Only what the sender claims to be | Trivial | Anyone can type any name; agentic browsers may send plain Chrome |
| Published IP list | The request came from the operator's network | Hard | Lists change; shared addresses; says nothing about the end user |
| Reverse DNS (with forward check) | The IP belongs to a hostname the operator controls | Hard | Only some operators document it; needs two lookups |
| Web Bot Auth signature | The operator's key signed this exact request | Very hard without the private key | Must be checked live; not every agent signs yet |
| Cloudflare verified bot | Cloudflare confirmed the identity and reviewed behavior | Very hard | Only on Cloudflare; a policy label, not a guarantee of intent |
| Card-network agent programs | The agent is registered with Visa, Mastercard or Amex | Very hard | Mainly at checkout; still rolling out |
The user-agent string
The user-agent is a text header the visitor writes itself. Cloudflare puts it plainly: "user agent headers alone are easily spoofed and are therefore insufficient for reliable identification."
Use it to group traffic and to make a first guess. Never use it alone to grant access, lift a rate limit or trust an order.
Published IP lists
Many AI operators publish the IP ranges their bots use, often as JSON files. If a request claims a bot name and comes from inside that bot's published ranges, it very likely comes from that operator.
Two limits matter. The lists change, so fetch them on a schedule instead of copying them once. And Cloudflare notes that one address "might be shared by multiple users or multiple services within the same company", so an IP match tells you the network, not the product or the person.
Reverse DNS
Some operators, Google and Microsoft among them, document a DNS check. You look up the hostname behind the visitor's IP, confirm it ends in the operator's domain, then look that hostname up again and confirm it points back to the same IP. Cloudflare lists reverse DNS as one accepted way for a bot to identify itself.
Our guide to catching fake GPTBot, ClaudeBot and Googlebot requests walks through the exact steps for each vendor.
Web Bot Auth signatures
A signed request carries a digital signature made with the operator's private key. Your site or CDN checks it against a public key the operator publishes. A copied header is useless without that private key.
OpenAI documents this for ChatGPT agent. Signed requests carry a Signature-Agent header set to "https://chatgpt.com", and OpenAI says the guidance "applies to signed cloud browser requests, not web crawlers." Google says it "is also experimenting with the Web Bot Auth protocol, using the https://agent.bot.goog identity" for Google-Agent.
Signatures expire quickly, so they must be checked within minutes of the request, usually as it arrives; most logs don't keep the headers needed to check them later. The full mechanics are in our Web Bot Auth explainer, and the ChatGPT-specific checks are in how to tell if a visitor is really ChatGPT.
How do Cloudflare verified bots and signed agents fit in?
Cloudflare keeps a list of verified bots, and since 1 July 2026 that list includes signed agents. Its documentation says simply: "Signed agents are now Verified."
To be verified, Cloudflare requires two things. The bot must identify itself honestly, through a Web Bot Auth signature, a published IP list with a stable user-agent, or reverse DNS. And it must behave: follow robots.txt and crawl directives, keep request rates reasonable, and not evade site owner preferences. Cloudflare says breaches can get a service removed from its allowlist.
The July 2026 change also added a field Cloudflare calls Direct versus Intermediary:
- Direct means a single operator runs the bot, and only that operator can send requests that present as it.
- Intermediary means an agent service that many end users can operate. The operator runs the software, but a different person starts each action.
This matters for trust. With an intermediary agent, you can trust the operator without trusting every person using it. Cloudflare calls this "transitive trust", and says it is testing the Forwarded header to pass on information about the end user.
Cloudflare also groups verified bots by behavior, including Search, Agent, Training and Transact. That lets you write rules such as "allow search crawlers, limit training crawlers" without listing every bot by name.
How are Visa, Mastercard and American Express identifying agents?
The card networks are building agent identity into checkout, and they are using the same signature approach. Cloudflare said in October 2025 that Visa's Trusted Agent Protocol and Mastercard's Agent Pay both use Web Bot Auth as their agent authentication layer. It added that American Express would also build its agentic commerce offering on Web Bot Auth.
In this model, an agent registers with the network, and its public keys sit in a directory the network hosts. Cloudflare says Visa and Mastercard will each host a directory for their registered agents. A merchant, or the merchant's CDN, can then fetch those keys and check that a shopping agent is one the network knows.
Cloudflare's write-up also describes separating browsing from paying, so a site can tell whether an agent is looking at products or trying to buy. These programs are still rolling out, and the exact details have changed over time, so treat them as a direction of travel rather than something every agent supports today.
Network registration answers "is this a known agent?" It still does not prove what the shopper approved. That is the job of signed instructions such as an agent mandate, which sit closer to the payment than to the web request.
What should you do in practice?
Match the strength of the check to what the visitor is allowed to do. Reading a public product page needs little proof. Placing an order needs a lot.
- Sort traffic by claimed name first. Use the user-agent to group visits into crawlers, user-triggered fetchers and everything else. Treat this as labeling, not trust.
- Verify the bots you care about. For crawlers and fetchers, check IP lists or reverse DNS. Record the result with each request: verified, failed or unverified.
- Turn on signature checks where you can. If you are on Cloudflare, its verified bot data is already available to your rules. Elsewhere, look for a plugin or server module that checks Web Bot Auth live.
- Write rules by verified identity. Allow, limit or block based on what was proven, not what was claimed. A request that claims to be a known bot but fails verification should get no special treatment.
- Expect agents you cannot name. Some agentic browsers look like ordinary Chrome. Rate limits, bot challenges and checkout rules still apply to them, whatever they call themselves.
- Keep "which agent" separate from "which person". Identity checks stop at the operator. Decide separately what an agent may buy, for whom, and when a person should review it.
- Revisit often. Operators add new bots, change IP ranges and start signing. Check the AI agent directory for current names and verification methods.
If you are not sure where you stand, the free readiness scan shows how your site looks to AI agents today.
How Sill helps
Sill shows you which AI agents and bots visit your site, the pages they read, and what agents ask for and buy. It verifies signed agent requests (Web Bot Auth) on Cloudflare and WordPress installs. When an AI agent places an order, your rules approve it, decline it or hold it for a person to review, and every decision is signed and recorded.
Questions people ask
What is the most reliable way to identify an AI agent?
A cryptographic signature checked live as the request arrives. Web Bot Auth lets the operator sign each request with a private key, and your site or CDN checks it against the operator's published public key.
Does a verified AI agent mean the order or request is safe?
No. Verification tells you which company or service sent the request. It does not tell you who the end user is or whether you want that action, so you still need your own rules.
Why do some AI agents show up as normal Chrome visitors?
Some agents drive a real browser, either in the cloud or on the person's own computer. Cloudflare notes that agents often send the Chrome user agent, because there is no clear way for a semi-automated browser to identify itself in that header.
Do I need Cloudflare to verify AI agents?
No, but you need something that checks requests as they arrive. IP and reverse DNS checks can run on any server. Signature checks need software at your CDN, server or plugin that understands Web Bot Auth.
Sources
- Cloudflare docs: Verified bots (signed agents) · checked 16 Nov 2026
- Cloudflare blog: Forget IPs: using cryptography to verify bot and agent traffic · checked 16 Nov 2026
- Cloudflare blog: Securing agentic commerce: helping AI agents transact with Visa and Mastercard · checked 16 Nov 2026
- Google: User-triggered fetchers · checked 16 Nov 2026
- OpenAI: Help ensure cloud browser access to your site · checked 16 Nov 2026
- Anthropic: Claude in Chrome is generally available · checked 16 Nov 2026