How we score
What an AI shopping agent can confirm about a store from its public pages, measured the same way for every store.
Five pillars
Each pillar asks one question. Its weight is its share of the score. The score counts only the tests the agents could decide, passed or failed, so a pillar's score times its weight doesn't always add up exactly to the overall score.
| Pillar | Question | Weight |
|---|---|---|
| A Access and discovery | Can a declared agent reach and read the offer and policy pages? | 15% |
| B Offer identification | Can an agent tell what is sold, at what price, and whether it is available? | 25% |
| C Material terms | Can an agent find the full cost, delivery and return terms before buying? | 25% |
| D Observable operability | Does the page show how an agent would act on the offer? | 20% |
| E Evidence and recourse | Can the offer and its policies be cited and contested later? | 15% |
Two declared agent profiles
Declared HTTP client. A plain HTTP client with an identifiable user agent. It does not run page scripts, like Meta's discovery crawler.
Declared headless browser. Headless Chromium with the same identifiable user agent and no concealment. It renders pages.
Both honour robots.txt and send about one request per second. Neither passes a barrier, logs in, or opens a cart, checkout or login page. One offer page per profile is sampled. Results do not transfer between the two profiles.
Score, range and tests checked
Tests checked is the number of tests that could be decided, as passed or failed, out of all 38, each test counting once: 34 of 38 is 89.5%. It is also given per pillar: 7 of pillar B's 9 tests is 77.8%. A test that could not be decided is never counted as a failure. A test that does not apply to a kind of business, such as a product identifier for a restaurant or a hotel, is left out of the count and of the score.
Score is readiness among the tests that could be decided, from 0 to 100: the points of the passed tests over the points of the passed and failed tests. Tests that could not be decided are left out, so the score covers only what is known.
Score range is the score if every undecided test failed, up to the score if every undecided test passed. It describes missing information, not statistical confidence.
Two profiles, one figure. The score, its range and tests checked are the mean of the two agent profiles. A missing profile, or a missing value, counts as 0, so a store read under one profile only has at most 50% of tests checked. Each profile's own figures are on the store's fact sheet.
When the score is shown. Only for Agent Ready and Close, where enough is known for the score to mean something. Other stores show their label and how much was checked, overall and per pillar.
Fixed rules decide each test from typed data and page text. No AI model judges. Facts read from page text are published only after a two-rater check. Until then they show as Not yet confirmed.
The four labels
Every store has exactly one. The rules are checked in this order, and the first one that applies decides the label.
| Label | Rule | Stores | |
|---|---|---|---|
| 1 | Invisible to agents | The declared HTTP client couldn't read the home page: a bot wall, or a robots.txt that forbids crawling or couldn't be fetched. | 435 |
| 2 | Agent Ready | On the mean of the two agent profiles: at least 90% of tests checked, at least 80% of each pillar's tests checked, and a score of at least 70. | 63 |
| 3 | Close | The better agent profile has at least 50% of tests checked. | 789 |
| 4 | Not enough evidence | Under 50% of tests checked under both agent profiles. | 1262 |
Invisible is decided on the declared HTTP client alone, the simplest kind of agent. A store is Invisible when that agent meets a bot wall on the home page (a named bot-protection service such as Cloudflare, DataDome or Akamai), or a robots.txt that forbids crawling or can't be fetched, whatever the browser saw. Other refusals, such as a plain HTTP 403 or a home page that's empty without JavaScript, don't make a store Invisible: they leave that agent's tests undecided (see What a barrier means). Close is judged on the profile with more tests checked.
What a barrier means
A bot wall, an HTTP 403, a robots.txt rule or an empty page is an access result for that agent profile. It means the profile could not read the page. It says nothing about whether the store is honest or ready to sell.
A barrier leaves that profile's tests undecided, so it lowers tests checked and can keep a store out of Agent Ready. It never counts as a failed test. A bot wall or a robots.txt rule met by the declared HTTP client on the home page makes the store Invisible to agents.
Observation dates
Stores are observed in runs. 2253 stores were read on 20 October 2026, under rubric 0.3. 295 stores were read on 18 October 2026, under rubric 0.3. 1 store was read on 15 October 2026, under rubric 0.3. Each fact sheet states the dates its store was read on and the rubric version used. Each store has 1 observation so far, unreviewed. Site generator 3.0.
What this doesn't claim
A profile doesn't rate a store's honesty, product quality or prices. It shows only what an agent could confirm from public pages on those dates.
Add-to-cart, checkout, payment, delivery and refunds were not tested. A label is not a probability that a purchase will succeed.