Tool
AMDT AI Traffic Analytics: see which AI bots read your site, and whether they are who they claim to be
AMDT AI Traffic Analytics is a free WordPress plugin. It shows you what your analytics cannot see: which AI bots read your site, whether they are who they claim to be, and how many people reach you from ChatGPT, Claude, Perplexity and other assistants. It also tells you whether an AI agent can use your pages.
Free · GPL licence · Your data stays on your server · Nothing is sent to AMDT
Open your analytics and you see people: sessions, pages, conversions.
Two things are missing from that picture. The first is who reads your site in order to feed a language model. That is not a session; it is a server request, and your analytics, which runs on JavaScript in the browser, never counts it. The second is how many people arrive because an assistant cited you. Those are real people, but they usually end up filed under direct traffic, or scattered across referrals that tell you very little.
Two different questions, two different measurements. This plugin handles both and, just as importantly, tells you when it cannot answer them.
The bots reading your pages
The plugin logs every request from a declared agent: which bot, which page, when, and from which address. The catalogue ships with twenty-three names, from OpenAI, Anthropic, Perplexity, Google, Microsoft, Apple, Meta, ByteDance, xAI and DuckDuckGo, and you can add your own. Each one states what it reads for: a search engine, a person’s request, model training, advertising, or several of these at once.
One distinction worth making, because almost nobody makes it: of those twenty-three names, twenty-one can appear in a request to your server. The other two, Google-Extended and Applebot-Extended, never visit pages. They are labels used in robots.txt to signal whether your content may be used for training, and according to Google and Apple they never appear in a request. That is why they ship switched off and marked as what they are: they are in the catalogue so you know what they mean, not because you will ever see them.
You will also find Googlebot and Bingbot there, which are classic search crawlers rather than AI agents. They are included because their index feeds AI products (AI Overviews and Copilot), so their crawling counts in this conversation too.
And there is one agent that does not introduce itself by name. ChatGPT agent arrives with the User-Agent of an ordinary Chrome browser, so no name tells it apart from a person. Since version 2.7.20, the plugin recognises it by the signature on its requests and logs it as “ChatGPT Agent”.
The people arriving from an assistant
When somebody clicks a citation inside ChatGPT, Claude, Perplexity, Gemini, Copilot, Grok, Mistral Le Chat, DeepSeek, Meta AI, You.com, Phind, Kagi Assistant, DuckDuckGo or Exa and lands on your site, the plugin records it and tells you which page they reached.
These visits are measured on cached pages too, because it is the browser, not your server, that reports where the visit came from. With one condition: the browser has to send that information (the referrer). Visits without it cannot be attributed to any assistant, and the plugin does not guess. There is a single exception, and it is documented: ChatGPT adds utm_source=chatgpt.com to the links it gives, and a visit without a referrer that carries that tag is attributed to ChatGPT.
Alongside that: a dashboard with totals and charts, which adds up bot reads by purpose and filters by any of them; a log you can filter and export to CSV; dated markers on the chart for changes in the AI sector that can move your figures, each with its source, alongside any you add yourself; and an alert when an operator’s agent that reads on behalf of people stops reading your site while its search crawler keeps going, which is the typical trace of a block somewhere along the way. None of what you see there leaves your server.
How it knows whether a bot is genuine
A bot identifies itself with a name, the User-Agent, and that name is simply text. Declaring yourself GPTBot costs exactly as much as declaring yourself anything else.
So the plugin does not take the name at face value: it checks it. A bot can prove who it is in four ways.
- The address lists operators publish. OpenAI, Anthropic and Perplexity publish the ranges their bots use.
- Reverse DNS. Google, Microsoft and Apple allow you to ask who an address really belongs to.
- Known operator networks. For Anthropic and Meta, the ranges ship with the plugin itself.
- The request signature. Some agents sign their requests with Web Bot Auth, an IETF draft: today Google-Agent does so on part of its requests, and so does ChatGPT agent. The plugin checks the signature against the agent’s public key. Anyone can write a User-Agent; only whoever holds the key can produce a valid signature, which makes it the strongest proof of the four.
Every request ends up in one of three states: verified, impostor or unverified. When the proof is the signature, the plugin marks it as “Verified by signature”, and the IP analysis shows you every signed request with its three headers.
It flags an impostor only when it can prove the claim: because reverse DNS contradicts it (Google, Microsoft, Apple), because the operator list does not include that address (OpenAI, if you have enabled the download), or because the address belongs to a different known operator.
And when it cannot prove anything, it says so. Unverified does not mean suspicious. It means that with the available evidence there is no way to confirm or deny the claim. Some bots will never leave that state, because there is no list, no reverse DNS and no known network to check them against: Bytespider, Grok, and any agent you add yourself. A tool that handed out verified badges to everyone would be giving you false comfort.
The bot catalogue also shows you how many addresses each list holds and the date its operator gave it, so you can tell an up-to-date list from one that has not been touched in a year.
llms.txt
llms.txt is a site content index written for language models. The plugin generates yours, counts who reads it, and lets you leave entries out from the editor itself. It also leaves out the entries your SEO plugin marks as noindex (Yoast SEO, Rank Math, All in One SEO and SEOPress) and, if you write it in the settings, publishes what your site is for. If you do not write it, the plugin does not make it up.
Once a day it also checks that your /llms.txt responds the way Lighthouse expects (a title, a link and at least fifty characters) and warns you on the dashboard if it does not.
If Yoast SEO or Rank Math are already serving theirs, the plugin does not generate another and cannot count those reads, and the dashboard tells you so. If your llms.txt comes from a different plugin, or is a file someone uploaded to the server by hand, it is not detected and you get no warning. That is a real limitation and we would rather you knew in advance.
The addresses it cannot identify
Not everything crawling your site announces itself. The plugin gathers the addresses it could not identify and hands you the blocking rule, already written, for Apache, Nginx, Cloudflare and Wordfence. That rule never includes the ranges of an operator that runs search bots or bots acting on a person’s request, and if you ask for one for such an operator, it starts with a warning.
You apply the rule, in your own firewall. The plugin blocks nobody. That is deliberate: deciding who gets in is a business decision, not a box a measurement tool ticks on your behalf.
Caching, where this gets complicated
A page served from cache without passing through WordPress cannot be measured. WordPress never runs, so there is nothing to record. This affects bots, not assistant referrals, which are counted on cached pages.
With WP Rocket, LiteSpeed Cache or W3 Total Cache, the plugin offers to exclude bots from the cache at the press of a button, and the problem goes away for the ones that announce themselves by name. It does not work for an agent that signs its requests but presents itself as a browser, such as ChatGPT agent: that one is only seen on pages the cache does not serve.
With WP Super Cache, or with a server-level cache (nginx, Varnish, your host’s own cache, or Cloudflare caching the HTML), it cannot be done. The plugin warns you when it detects one, although it does not always detect one: it tells you which header gives it away and where to act, and checks again when you ask it to.
Assistant visits are not lost either when WP Rocket (“Delay JavaScript execution”) or LiteSpeed Cache (“Load JS Deferred: Delayed”) hold scripts back until the visitor interacts with the page. The plugin lets its own through, so it also counts the visitor who reads and leaves without scrolling.
Whether an AI agent can use your pages
Some AI agents do more than read: they move through pages and act on them. They fill in forms, press buttons, open menus. To do that, the page has to be built in a way they can understand, and you cannot tell that by looking at it.
Agent-Readiness, included in the free version since 2.7.19, checks it. It reads the HTML your server returns for the home page, a sample post, up to eight pages and a sample product, or for any page of your site you paste in. For each one it gives you:
- A score from 0 to 100 for what such an agent needs: real buttons and links, the heading outline, the page landmarks, labelled form fields, image alternative text, and how much of the page is meaningless
div. - An approximation of Lighthouse’s “Agentic Browsing” audits, as far as they can be judged from the server: the 33 accessibility rules of its accessibility-tree audit, the WebMCP ones, llms.txt and the ARD manifest. Visual stability (CLS) needs a browser, so each result links to PageSpeed Insights to measure it.
- How much of the text reaches the initial HTML and how much needs JavaScript to appear.
- How to fix it with your page builder: Elementor, Divi, Bricks, Beaver Builder, WPBakery, the block editor or your theme.
The results export to Markdown or CSV. Agent-Readiness only analyses your own site and never changes your pages: it tells you what to fix, and fixing it is up to you.
Privacy: what is stored and what is not
Everything is stored in your own site database. This is what goes into it:
- From bots and from whoever reads llms.txt: IP address, User-Agent and page.
- From people arriving via an assistant: the page, the User-Agent and the address they came from. Of their IP, only the network is stored, never the full address: for IPv4, up to the third block; for IPv6, the first forty-eight bits.
- From automated browsers posing as an iPhone: the request with its full IP address.
- From requests signed with Web Bot Auth: also their three signature headers (
Signature-Agent,Signature-InputandSignature) and the result of checking them. - From an ordinary visitor: nothing.
That last line carries one exception worth stating out loud: the automated-browser detector could, in theory, mistake a person for one. It needs two matching signals to do so, which makes it unlikely, but not impossible.
Nothing is stored in your visitors’ browsers: no cookies, no local storage. Pages carry a small script that, only when the visitor arrives from an AI assistant, reports it to your own server.
Data is kept for thirty, sixty or ninety days, as you choose. The diagnostic log ships switched off. If you switch it on, it also writes IP addresses and User-Agents to a file with a random key in its name, inside its own folder in your uploads directory (amdt-ai-traffic-analytics/), which denies direct access on servers that read .htaccess, such as Apache.
On uninstall, data is not deleted unless you untick “Keep data on uninstall” in the settings, and never while the other edition is still installed. If you choose to delete it, the diagnostic log is deleted too.
The plugin also adds suggested wording to the WordPress privacy guide (Settings → Privacy), for you to review and fold into your own policy if appropriate.
What leaves your server
Nothing reaches AMDT. There is no telemetry. The plugin only connects to the following:
- The IP address lists. Once a day, the public lists from OpenAI (four of them, including OAI-AdsBot’s), Anthropic and Perplexity. This ships switched off until you enable it: when you activate the plugin, a notice asks whether you want it. Until you enable it, OpenAI and Perplexity bots will show as unverified, and an impostor using the name of an OpenAI bot will not be caught; Anthropic bots are still checked by reverse DNS and by their networks, which ship with the plugin.
- The signing keys. Under the same switch, the public keys of Google-Agent and ChatGPT agent, and only from their two addresses (
agent.bot.googandchatgpt.com), never from one a visit points to. They are downloaded only when a signed visit from one of them arrives and the stored key has expired. - DNS. The plugin asks your server’s DNS about the addresses of bots and of whoever reads llms.txt. It does not do this for people arriving from an assistant.
- Your own site, when you open the dashboard or the settings. The home page, at most twice a day, to check whether there is a cache sitting in front of WordPress, and your
/llms.txt, at most once a day, to check that it is served correctly. - Your own site, when you run Agent-Readiness. The pages it analyses, your
/llms.txt, your/robots.txtand the addresses where an ARD manifest may live, always checking the certificate. The PageSpeed Insights link next to each result is an ordinary link: the plugin sends nothing to Google.
What it doesn’t do
We’d rather tell you this before you find out for yourself:
- It does not block bots. It tells you who came and hands you the written rule; applying it is yours to decide.
- It does not improve your AI visibility or make anyone cite you. It measures. What you do with the measurements is a separate conversation.
- It does not see everything. It misses pages served by a server-level cache, bots presenting themselves as ordinary browsers without signing their requests, and assistant referrals that arrive without a referrer, except ChatGPT’s when they carry its tag.
- It does not change your pages or analyse other people’s sites. Agent-Readiness only reads your own site and tells you what to fix; fixing it is up to you.
- It carries no API, no MCP server and no WebMCP. The free version measures; it does not serve content to agents.
Requirements and installation
You need WordPress 6.9 or later and PHP 8.1 or later, and it has been tested up to WordPress 7.1. The plugin is free and distributed under GPL v2 or later. The panel is in English, with a Spanish translation included in the plugin itself.
It installs like any other plugin: from the Plugins screen of your WordPress, search for “AMDT AI Traffic Analytics”, install it and activate it. On activation, a notice asks whether you want to download the operators’ IP lists and signing keys; you can change this later in the settings.
If there is a CDN or external proxy in front of your site other than Cloudflare, you will need to declare it in the settings. If you do not, every visit will show the proxy’s address and the measurement will be worthless. A proxy on the same server or on a private network is accepted automatically.
AMDT AI Traffic Analytics
Frequently asked questions
How can I tell whether ChatGPT or Claude are reading my site?
Language model bots identify themselves by name on every request to your server (GPTBot, ClaudeBot, PerplexityBot and others), but your usual analytics does not record them, because it runs on JavaScript in the browser and a bot does not execute JavaScript. You have to measure at server level. AMDT AI Traffic Analytics logs those requests, checks whether each bot is genuine, and tells you which pages were read.
Does it block AI bots?
No. It tells you who came in and hands you the blocking rule, already written, for Apache, Nginx, Cloudflare and Wordfence. Whether you apply it, and applying it in your firewall, is yours to decide.
Does it work if I use a caching plugin?
It depends what is being measured. Human visits from assistants are counted on cached pages too, provided they arrive with a referrer (or, from ChatGPT, with its utm_sourcetag), and also when WP Rocket or LiteSpeed Cache hold scripts back. Bot requests are only recorded when the page passes through WordPress; with WP Rocket, LiteSpeed Cache or W3 Total Cache the plugin offers to exclude bots from the cache at the press of a button. With WP Super Cache or a server-level cache it is not possible, and neither is it for an agent that signs its requests but presents itself as a browser, such as ChatGPT agent.
I already have the llms.txt from Yoast SEO. What happens?
Yoast keeps serving it. This plugin does not generate its own and cannot count who reads it, and the dashboard says so. If you have written in the settings what your site is for, it adds that section to Yoast’s file. If your llms.txt comes from a different plugin or from a file uploaded to the server, it is not detected.
What data does it store, and where?
In your own site database. From bots: IP, User-Agent and page, and for signed requests, also their signature headers and the result of checking them. From people arriving via an assistant: the page, the User-Agent, the originating address, and only the network portion of their IP, never the full address. From an ordinary visitor, nothing, and it sets no cookies in their browser. Data is kept for thirty, sixty or ninety days, as you choose.
Does it send anything to AMDT?
No, and it carries no telemetry. It connects to the public IP lists of OpenAI, Anthropic and Perplexity and to the public signing keys of Google-Agent and ChatGPT agent, both switched off until you enable them. It also queries your server’s DNS and makes requests to your own site to detect caching, check your llms.txt and, when you run Agent-Readiness, analyse your pages.
Do I need Cloudflare?
No. If you have a CDN or external proxy in front of your site other than Cloudflare, you simply declare it in the settings.
Is data deleted on uninstall?
Not by default. Only if you untick “Keep data on uninstall” in the settings, and never while the other edition of the plugin is still installed.
Why do GPTBot or PerplexityBot show as “unverified”?
Because the download of the operators’ IP lists is switched off, which is how the plugin ships. Enable it in the settings. ClaudeBot, Googlebot and Bingbot are still checked by reverse DNS even if you do not.
Does it tell me whether an AI agent can use my site?
Yes. Agent-Readiness analyses the HTML of your site’s pages, scores them from 0 to 100 on what an agent that fills in forms and presses buttons needs, and explains how to fix it with your page builder. It only analyses your own site and changes nothing on it.
Support
If you find a bug or behaviour that does not add up, open a topic in the plugin’s support forum on WordPress.org or write to us through the contact form.
AMDT AI Traffic Analytics is a tool developed by AMDT, a digital marketing consultancy in Madrid.
If what you need is the next step
Measuring is where it starts. Once you know who is reading you, the question stops being how many bots came in and becomes on what terms you want them to.
AMDT AI Traffic Analytics Pro is the complete edition of the plugin, for WordPress and WooCommerce, and it takes that step: alongside measurement, it serves your content to AI agents through an API and an MCP server, publishes the files that announce them, declares AI usage permissions, protects that API and analyses your content. On platforms that are not WordPress (PrestaShop, Shopify, Magento or bespoke builds), AMDT implements the same layer as custom work rather than as a plugin.
Both editions share the same tables and settings, so moving from one to the other loses nothing you have already measured. The Pro edition is not available for download: if you are interested, talk to us.