How do browsing AI agents read a store?
A browsing AI agent loads a store's pages in a real browser and works from what they show. It scrolls rather than using in-page menus, opens products up to its price limit, and reads specifications, delivery and returns. When a fact is missing, it keeps searching, then refuses. How much it reads depends on the model behind it.
Last updated
What does a browsing agent actually see?
The rendered page, in most of what was observed. The bench agents worked from screenshots through their vendor's computer-use interface. Muse ran the page's scripts like any browser. Grok's agent also fetched some pages as plain text, and once read text that was hidden from view.
On the bench, each run was one isolated browser, driven by screenshots. Three model families were tested this way: Anthropic's, OpenAI's and Google's.
In the lab, on 2026-09-24, Grok's agent applied a discount code taken from page text that was not displayed. It told its person that no code was shown. That run was excluded from the lab's findings. It still shows that an agent can read more of a page than a person sees.
How does it move through a page?
It scrolls. An "On this page" menu placed on every product page of the bench store was used in 0 of 24 runs, across three model families. In one model family, pop-ups were closed and ignored. When a link named where a fact was, the agent followed it straight away.
In a study of newsletter pop-ups (24 runs, one model family), every agent closed the pop-up and carried on. None gave an email address or took the discount.
In another study (one model family), the same fact was placed at five distances from the product page. When the navigation had a labelled Help entry, the agent took it at once. Without one, it read the whole product page first, which took longer.
What does it read, and how closely?
Specifications, reviews, delivery and returns, in amounts that depend on the model and on the job. Agents asked to report back to their person read far more than agents allowed to buy. And the model family showed in behavior more clearly than the mission did.
In the first bench study (168 runs), agents that had to report back spent about 36 seconds on specifications, 22 on reviews and 26 on returns. Agents allowed to buy spent about 5, 0 and 0.
The three families read differently. One read the most. One read specifications but never reviews. One took the shortest path. Use of the store's compare tool ranged from 29% of runs for two families to 87% for the third.
| Behavior | Anthropic | OpenAI | |
|---|---|---|---|
| Used the compare tool | 29% | 29% | 87% |
| Set a filter or a sort | 100% | 100% | 100% |
| Opened a product page | 100% | 100% | 100% |
| Median product pages opened | 4 | 3 | 5 |
| Made no misclick | 83% | 88% | 70% |
What happens when a fact is missing?
The agent searches, and the search is long. In one bench study (one model family), a missing fact roughly doubled the visit: 63 steps against 34.5, and 23.5 pages against 10. When the fact did not exist at all, the agent searched everywhere and then refused to order.
Refusing was the norm, not inventing. In a later study on three families, the store stated no warranty period and the mission required one. Eleven of twelve runs refused. One run, from one family, wrote down a two-year warranty the store never stated, and ordered on it.
Three clicks of distance was not a wall for two families, which found the fact 4 times out of 4. It was for the third: 3 of its 4 runs ran out of steps still searching.
What did Muse and Grok's agent do on a test store?
They read it closely. On 2026-09-24, on Cromanion's test store, both waited for reviews that loaded late and read them correctly. Both were blocked once by color swatches without labels. Grok's agent scanned the whole catalog before choosing, and went through checkout three times.
Both noticed that the cart did not show the chosen color. Each said it could not verify the variant.
In one Grok run, pressing Enter in the discount-code field submitted the whole checkout form, and an order was placed by accident. The agent said so in its report.
Common questions
Do agents read a page the way a screen reader does?
Not necessarily. The bench agents worked from screenshots, like a person looking at the screen. Some agents also fetch pages as plain text. Agents that read the page's code directly were not tested on the bench, and may behave differently.
Is this how every AI agent behaves?
No. It is what three model families did on one fictitious store, and what two real agents did in a handful of dated lab runs. The observations are anecdotal for the real agents. Agents change from one release to the next.
See it on your own site
Paste one tag. Cromanion crawls your site, watches real sessions in Learn mode, and only acts when you switch it on — with a permanent 10% holdout proving what it caused. Free to start, no credit card: a 14-day or 1,000-session live trial, then keep measuring for free.