How do you measure how AI agents perform on your website?
Measure agent performance the way you would measure a checkout funnel: give an agent a real task, then record whether it finished, how many steps it took and where it stopped. Combine that task test with a fetch test (what an agent reads without a browser), your server logs (which agents visit and what they are served) and a standards check.
By the Axidort team · Updated
Why page speed and SEO tools do not answer this
Lighthouse, Core Web Vitals and rank trackers measure what a person or a search crawler experiences. An agent has different failure modes: it is challenged as a bot, it cannot find the search box, the price only appears after a script runs, or the add-to-cart request is refused. None of those show up in a speed report.
The metrics worth tracking
| Metric | Question it answers | How to get it |
|---|---|---|
| Task completion | Did the agent reach the goal (for a store, the purchase step)? | Agent task test |
| Steps to goal | How much work was it? Fewer steps means less cost and fewer errors. | Agent task test |
| Blocker type | What stopped it: bot challenge, login wall, rejected request, pop-up? | Agent task test |
| Fact accuracy | Did the agent read the same name, price, currency and stock a person sees? | Task test compared with the page |
| No-JavaScript readability | Are the core facts in the HTML the server sends? | Fetch test |
| Agent traffic and status codes | Which agents visit, and are they served 200, 403 or a challenge page? | Server or CDN logs |
| Standards coverage | Which machine-readable entry points exist? | Standards check |
1. Run an agent task test
This is the measurement that matters most. Give a browser agent a task a customer would give it, for example "find a 10 mm drill bit, tell me the price and whether it is in stock, and take me to checkout", and watch what happens. Run it logged out, from a clean session, and stop before payment.
You can do this by hand with any agentic browser. To make it repeatable, an Axidort scan runs the same journey with a declared agent, records each step with screenshots and scores the result on seven measures. It never places an order. Because agents are language models, the same site can score slightly differently between runs, so judge a change over more than one scan.
2. Run a fetch test
Many assistants never open a browser. They fetch a URL and read the response. Check what that response contains:
# What does an AI assistant receive? Look for the price, currency and stock in the output.
curl -sL -A "Mozilla/5.0 (compatible; ExampleAgent/1.0)" https://www.example.com/products/some-product \
| grep -i -E 'price|availability|in stock|application/ld\+json'
# Is a declared AI agent served the page or a challenge?
curl -s -o /dev/null -w "%{http_code}\n" -A "ChatGPT-User" https://www.example.com/If the facts are missing from that output, they exist only after JavaScript runs, and a reader agent will not see them. A 403, a 429 or a page titled "Just a moment" means bot protection answered instead of your site.
3. Read your logs
Filter server or CDN logs by the user agents AI companies publish, such as GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, Claude-User and PerplexityBot. Track three things per agent: requests per day, the pages requested, and the share of responses that are not 200. A high share of 403 responses to user-triggered agents such as ChatGPT-User means real customers’ assistants are being turned away. The user agents and what each one does are listed in robots.txt for AI agents.
4. Run a standards check
A standards checker confirms which machine-readable entry points you publish: AI bot rules in robots.txt, a sitemap, Markdown responses, an MCP server card and so on. Cloudflare’s scanner reports that most of these are rare: about 4% of the 200,000 most visited domains declare AI usage preferences and under 4% serve Markdown to agents (Cloudflare, April 2026). A standards check is quick and useful, but it does not tell you whether a task can be completed.
What the numbers look like today
When Axidort ran the task test on 50 leading commerce sites, the median score was 11 out of 100. An agent reached a cart or buying flow on 21 of them, and 31 served a bot challenge to a declared agent (scanned October 1, 2026). The full results are in the Axidort Index.
Frequently asked questions
01Can Google Analytics show AI agent traffic?
Only partly. Agents that fetch pages without running JavaScript never fire an analytics tag, so they are invisible there. Server or CDN logs filtered by user agent are the reliable source.
02How often should I measure?
After any change to product pages, pricing, stock display, checkout or bot rules, and on a regular schedule otherwise, because bot-protection vendors change their defaults without telling you.
03What is a good task completion result?
For a store, an agent should find a product, read price and stock correctly and reach the purchase step without a challenge. In the Axidort Index that corresponds to grade A or B, which 7 of 50 sites reach.