How Kupeta tests

What the tests do, how the scores are made, and what they can't tell you. We'd rather you knew the limits than trusted a number too much.

Last updated 19 November 2026

1Two kinds of test

The free test only looks around a store: what robots.txt allows AI agents, whether machines can read the products' name, price and availability, whether a sitemap lists them, and what stands in the way, such as a captcha or a cookie bar over the page. It takes a few minutes and needs no account.

The full test sends test shoppers through real purchases, several times each, on a website its owner has verified. They search, choose a variant, put it in the cart and go through checkout with made-up details, then stop before the last confirmation. Full tests also look for what could trick an AI shopper.

How Kupeta behaves on websites is described in detail on the page about Kupeta's agent.

2The test shoppers

AI assistants differ: some see a page as a picture and click by sight, others read its text and structure; some are careful, some fast, some cheap and less capable. Kupeta's test shoppers are built on leading AI models to behave the way these kinds of assistants do.

They are Kupeta's own test shoppers. They imitate the behaviour of popular AI assistants, but they aren't the assistants themselves and aren't official products of the companies that make them.

Each shopper and its settings are a version. When a shopper changes, it becomes a new version, and earlier results stay tied to the one that produced them.

3Results come from the record

Whether a purchase succeeded is never taken from what the AI says about itself. Kupeta records every step of every attempt: what the shopper saw, what it did and what happened. The result is decided from that record by fixed rules, without AI.

An attempt can end as:

  • Reached payment, with the right product, variant, quantity and price
  • Wrong item: it reached payment, but with the wrong size, colour or quantity, which counts as a failure
  • Blocked by a captcha, a bot wall or a required login
  • Stuck: it couldn't find the product, choose a variant or get through a form or a pop-up
  • Price mismatch between the product page and checkout
  • Gave up or ran out of time
  • Our error, which doesn't count and is tested again

4The Agent-Ready score

A score from 0 to 100 for how well AI shoppers can buy from a store. In a full test, it's made of:

  • the share of test shoppers that reached payment, which weighs the most; every shopper counts the same, because a store should work for all of them
  • how far the others got
  • how quickly and in how few steps they did it
  • the results of the automatic checks
  • how reliably it works when they try again

A free test has no purchases, so its score comes from the checks alone. Every score is shown with the parts it's made of.

ScoreRating
90–100Excellent
70–89Good
50–69Needs work
Under 50AI shoppers have serious trouble

5The security score

A score from 0 to 100 for how well a store resists what could manipulate AI shoppers: hidden instructions, hidden text, scam links, prices that change at checkout and paid extras ticked in advance. It starts at 100 and loses points for every finding, more for more serious ones. With any critical finding, it can't be higher than 49.

Finding hidden instructions and the like is done by fixed rules, not by AI, so the same page always gives the same findings.

6Scores don't change behind your back

The way scores are calculated is versioned. When we change it, it becomes a new version: old results aren't quietly recalculated, and every result says which version made it.

The gain in points shown with each fix, and any amount of money, is an estimate and is always marked as one. An amount is only shown when the store has given its monthly sales.

7What the tests can't tell you

  • AI varies. The same shopper can act a little differently each time. That's why full tests repeat attempts, and why a single result is never the whole story.
  • Real assistants differ from ours. Test shoppers imitate kinds of assistants; a particular assistant may do better or worse on your store.
  • The last step isn't tested. Shoppers stop before the final confirmation, so what happens after it (the payment itself, the order confirmation) isn't part of the test.
  • A test is a moment. Stores and AI assistants change; a result says how it was on the day of the test. Watching repeats tests regularly.
  • Made-up details only. Shoppers don't log in or create accounts on your website, so purchases that need an account aren't tested.
  • Captchas end the attempt. Kupeta never gets around them, so whatever lies behind one stays untested and is reported as a finding.