How a visitor is identified
Every label says what it rests on: verified, declared or inferred.
On this page
Every visit gets a category and, where one can be given, a name. Every label also says what it rests on, so a claim is never shown as a fact.
Categories
| Category | Examples | Meaning |
|---|---|---|
| AI assistant fetch | ChatGPT-User, Claude-User, Perplexity-User, Meta's fetcher | An assistant read your page for a person, mid-conversation |
| Browser agent | Gemini's agent, Claude's browser, Muse | An agent operating a real browser on your site |
| AI search indexing | OAI-SearchBot, PerplexityBot, Claude-SearchBot | Indexing, so an answer engine can cite you later |
| AI training crawler | GPTBot, ClaudeBot, Bytespider, CCBot | Crawling for model training |
| Search engine | Googlebot, Bingbot | Classic search crawlers |
| Other bot | curl, uptime monitors, link previews, scanners | Everything else that is automated |
| Person ("Humans" in the dashboard) | No sign of automation was seen |
Only the first two count as agent visits in the dashboard. Training crawlers and search engines are shown apart, so they never inflate those numbers.
Verified, declared, inferred
| Basis | What it rests on | Shown as |
|---|---|---|
| Verified | The request came from an address in the published ranges of the vendor its user agent claims, or it carried a valid Web Bot Auth signature | ✓ verified |
| Declared | The visitor's own User-Agent names it. A claim, reported as a claim | self-declared |
| Inferred | Parlox's reading of what the visit did: automation traces in the browser, missing browser headers, requests for paths only scanners ask for | inferred, with the observations behind it |
A person is an inference too: a visit in which no sign of automation was seen. Evidence is written as observations ("User-Agent says Chrome, but no Sec-Fetch headers were sent"), never as a score. Inferred labels can be wrong in both directions, so read single visits as indicative, never as proof.
Checking a claim against published addresses
A user agent is a claim; the address a request came from is a fact. The vendors publish the addresses their agents use, and Parlox refreshes its copy of those lists about every six hours:
| Vendor | Vouches for |
|---|---|
| OpenAI | ChatGPT-User, OAI-SearchBot, GPTBot |
Google-Agent (Gemini driving a browser), Google's other user-triggered fetchers, Googlebot and its family | |
| Anthropic | Claude-User, Claude-SearchBot, ClaudeBot |
| Perplexity | Perplexity-User, PerplexityBot |
| Meta | Meta's fetchers (from the addresses its network announces) |
| Microsoft, Apple | bingbot, Applebot |
- Inside the vendor's list: the visit is verified, under the name and the kind its user agent gives (a verified
GPTBotstays a training crawler). - Outside it: the name the visitor gave is kept, and the visit says outside published ranges. Lists lag behind new addresses, so Parlox never renames a visitor on the strength of one check. Anthropic does not present its list as complete, so an address outside it is noted, not called outside.
- No claim, but a vendor's or a cloud's network: a hint toward a browser agent, never enough on its own.
The address is compared in memory and discarded. Only the verdict is kept.
Where the address comes from. The browser part's beacons reach Parlox straight from the visitor's browser, so their address is the visitor's. Fetchers never touch Parlox: they reach your server, and the server part forwards the address for requests whose user agent names a bot (The client address). A server-reported visit is therefore judged on the address your server reports, and anyone holding one of your site's secret keys could send a report with a vendor's real address and get a visit marked verified. Keep secret keys secret, and revoke one you suspect.
Signed requests
Some agents sign their requests with Web Bot Auth: an HTTP message signature made with a key the agent publishes. The server part forwards the Signature-Agent, Signature-Input and Signature headers, and Parlox checks the signature against the agent's published key. The signed host must be your site's own, so a signature captured on another site cannot be replayed on yours.
A valid signature proves who sent the request, not what kind of client it is. Only signers known to be AI agents are labelled as one; any other verified signer is an other bot, under its own name.
Browser agents
Many browser agents send the same User-Agent as Chrome, so the browser part looks at how the browser is built and how it behaves:
- automation flags (
navigator.webdriver, the traces Selenium, Playwright and Puppeteer leave); - software graphics rendering, or none reported;
- a browser brand list with Chromium but no consumer brand;
- the notification permission already set to denied;
- clicks with no pointer movement before them, perfectly straight pointer paths, identical scroll steps.
Signals combine; a single weak one does not make a visit an agent. Locale and time zone are not signals: some agents report the user's own.
Unidentified browser agent. When the evidence says "automated browser" but nothing names the product, the visit is an unidentified browser agent. Parlox does not guess a name. A name inferred from evidence always says so ("Muse (likely)") and needs two independent facts, such as the browser's signature together with a matching fetch by the same vendor.
Vulnerability scanners
A request for a path no shopper or search engine asks for (/.git/config, /.env, /wp-login.php, /phpmyadmin, backup archives, private keys) is a scan, whatever its user agent says. It is labelled as a vulnerability scanner, an other bot.
What it cannot see
- Agents that run on the person's own computer, such as a browser extension, use the person's own address; only the in-page signals can catch them.
- An agent missing from Parlox's list of user agents shows as an "Unidentified bot", or as an unidentified browser agent if it drives a browser. A fetcher missing from the list that sends a browser's user agent and headers is judged a person on your server, and is not sent at all.
- A fetcher behind your CDN arrives with the CDN's address unless your server reads the CDN's client header.