Specoria
Menu
TR Free test

InsightsMeasurement

Beyond the last click: measuring the agent funnel from fetch to repeat order

Last-click reports only see the shopper who clicked through from an assistant. A five-step chain, from fetched to repeat order, shows where the agent channel really wins or loses you sales, and which signal you already have for each step.

By 7 min read

A shopper asks an assistant for a waterproof hiking jacket under a budget. The assistant reads a dozen product pages, names three stores, and the shopper clicks one of them a day later, from a bookmark. In your analytics, that order is "direct". The assistant did the work; the report gives the credit to nobody.

Last-click attribution was built for a funnel where every step happened on your site or in an ad. With agents, much of the journey happens before the click, and some of it never produces a click at all. In this article, we split the agent channel into five steps, show which signal you already have for each one, and explain what to do with the gaps.

Why the last click sees so little

Two published measurements show how much happens before anyone reaches your checkout.

HUMAN Security examined the page paths of AI agents in its traffic data. From January to August 2025, about 87% of the pages agents browsed were product pages; checkout made up 2.2%. Agents mostly research. If you only count completed orders, you miss nearly all of their activity.

When the click does come, it is worth more than you might think. Adobe reported that in March 2026, AI traffic to US retail sites converted 42% better than non-AI traffic, a reversal from March 2025, when it converted 38% worse. That is US data and an average across retailers, but it means a small number of visits can carry real revenue.

So the channel is large at the top, thin in the middle and valuable at the bottom. A single last-click number flattens all of that.

The five steps

Think of the agent channel as a chain. Each step has its own question and its own evidence.

  1. Fetched. Did an assistant's crawler or user agent read your pages?
  2. Mentioned. When a shopper asks a real buying question, does the answer name you?
  3. Clicked. Did a person follow a link from the assistant to your site?
  4. Reached checkout. Could a shopper, or an agent working for one, get from the product to the checkout step?
  5. Bought and came back. Did the visit become an order, and did that customer order again?

A weak step limits every step after it. A store that is never fetched can't be mentioned; a store that is mentioned but whose checkout an agent can't reach turns recommendations into abandoned carts.

Step 1: fetched

Signal: your server or CDN logs. Assistants send their own user agents. OpenAI, for example, documents OAI-SearchBot for its search index and ChatGPT-User for pages opened on a user's behalf. Count requests by bot and by page type.

What to watch: which pages they read. If assistants mostly fetch your blog and rarely your product pages, the catalog may be blocked, slow or hard to reach from your navigation. Our bot policy by page type covers which pages to open.

Limit: some browser agents look like ordinary browsers and leave no bot name in the log. We described them in the agent map. Fetch counts are a floor, not a total.

Step 2: mentioned

Signal: none in your analytics. A conversation where you weren't recommended leaves no visit. The only way to see it is to ask: run real buyer tasks against the assistants many times and count how often you are named.

That count is the selection rate, and one screenshot is not enough to estimate it. We explained why in measuring the agent channel: the same question rarely returns the same list twice.

What to watch: selection rate per task, with its confidence interval, and whether the answer links to you or only names you.

Step 3: clicked

Signal: the Referer header and the utm_source parameter. OpenAI says ChatGPT adds utm_source=chatgpt.com to links in its search results, so those visits can appear as their own source. Other assistants send a Referer with their own domain.

What to watch: one ratio ties step 1 to step 3. Crawl-to-refer divides an AI company's bot requests by the human page views its assistant sent you. A rising ratio means the assistant reads more of your site without sending more people. You can work it out from your own log with our crawl-to-refer calculator, which runs in your browser.

Limit: a shopper who copies a link, or comes back later from a bookmark, shows up as direct. Clicks are also a floor.

Step 4: reached checkout

Signal: your funnel reports, filtered to sessions from step 3. For agents that act inside a browser, you need a test: send an agent through your store and see where it stops.

In our own field study we did that 40 times on five live stores. The strongest models reached checkout on all nine completable tasks, the lighter ones on seven of nine. The stops were concrete: a checkout button hidden where the agent couldn't see it, a cart link that was hard to find, a "sign in or continue as guest" screen. Forty runs is a small sample; read it as field observation, not a success rate. The details are in we sent AI agents shopping in a real browser.

What to watch: for each product type you sell, can an agent go from product page to checkout without guessing a URL?

Step 5: bought and came back

Signal: your order system. Store the first known source on the order (the utm_source or Referer from step 3) so it survives past the session. Then you can compare average order value, return rate and repeat purchase for assistant-referred customers against everyone else.

What to watch: repeat orders. A customer who found you through an assistant and then ordered again directly is the clearest sign that the channel built a relationship, not just a sale.

What you can set up this month

  1. Split your access log by AI user agent and by page type (product, category, policy, checkout). Note what share of their requests reach product pages.
  2. Create a separate channel in your analytics for assistant referrers and utm_source values, instead of leaving them inside "referral".
  3. Calculate crawl-to-refer for each AI company once a month and keep the series.
  4. Write five real buyer tasks for your best categories and measure selection rate with repeated runs, not one screenshot.
  5. Walk your checkout path as an agent would: no hover menus, no guessed URLs, no sign-in wall.
  6. Save the first source on every order and report repeat purchase by source.

The bottom line

The agent channel doesn't fit in one number. It is a chain: fetched, mentioned, clicked, reached checkout, bought and came back. Each step has a signal you can collect, and two of them, mentions and fetches by unnamed browser agents, need measurement outside your analytics. Look at the whole chain and you can see which step to fix first.

Specoria measures the steps your analytics can't see: whether agents can read and use your store, how often assistants name you on real buyer tasks, and where a browser agent stops on the way to checkout. Even the free plan's project panel includes an agent traffic report built from your own logs. Start with a free readiness test.

Sources