Methodology
Up to 35 scored checks on any one store, across 4 categories, plus one gate. This page is generated from the same registry the scanner runs, so it cannot describe a version of the tool we no longer ship.
Which stores it reads
Any online store. We look for a commerce platform on your homepage, for product pages that offer something for sale, and for a cart and buy buttons. If we find none of them, we stop and say so, and nothing is scored.
Shopify publishes its catalog and its agent endpoints, so a Shopify store is read from those. A store on any other platform is read from its sitemap and its product pages, which means its product checks describe the pages we sampled. If we cannot find its product pages, those checks are untested and the score is withheld.
5 checks test something particular to one platform (Shopify, Salesforce Commerce Cloud). On any other platform they are not run and do not appear in the report, so they touch neither the score nor coverage. A store is not marked down because its platform has not shipped a protocol, so the same score means the same thing wherever the store runs.
2 more test an open protocol that Shopify stores publish by default. On another platform they are reported when the store publishes it, and left out when it does not.
How the score is built
Each check returns a score between 0 and 1. A category score is the weighted mean of its checks; the overall score is the weighted mean of the categories.
| Category | Weight | Checks |
|---|---|---|
| Basic | 25% | 13 |
| Site-wide | 20% | 8 |
| Product pages | 35% | 12 |
| Long-tail pages | 20% | 3 |
Bands
| 85–100 | AI-ready |
| 65–84 | Mostly ready |
| 40–64 | Gaps to close |
| 0–39 | Invisible to agents |
Two rules that keep it honest
A check we could not run leaves the denominator. It does not score zero. “We could not measure your page speed” must never reach you as “your page speed is terrible”, and a check that fails because of a bug in our code is our problem, not a mark against your store.
Below 60% coverage we publish no score at all. A number assembled from the fraction of a store we could see is an artifact of what we could not. You get the findings we did establish, and no headline figure.
Absent is not the same as prevented
| Your store answers | We report |
|---|---|
| 404 / 410 | Fail: the thing is genuinely missing |
| 429 / 403 / 5xx / timeout | Untested: we were prevented from looking |
A store that defends itself well is not a store with missing policies. This distinction is the one the whole report rests on.
Every check we run
Weights are relative within a category. A weight of 0 means the check is a gate or is reported for information and does not move the score.
Basic · 25% of the score
Whether an agent can reach you at all, and whether the machine-readable entry points it looks for are there.
| An online store was found at this address | gate |
| robots.txt is reachable and parseable | 2 |
| AI crawlers are allowed on product, collection and content pages | 5 |
| robots.txt declares a sitemap | 1 |
| The sitemap is segmented, current and lists each product once | 2 |
| llms.txt is published at the domain root | 2 |
| llms.txt follows the spec and its links resolve | 2 |
| Shopify publishes an agent instruction file at /agents.md· Shopify only | 3 |
| Storefront MCP server responds and advertises its tools· Shopify only | 4 |
| Agents get a real answer to "what is your return policy?"· Shopify only | 3 |
| Catalog search answers the questions shoppers actually ask· Shopify, and elsewhere when published | 4 |
| UCP merchant profile is published at /.well-known/ucp· Shopify, and elsewhere when published | 3 |
| Products meet Shopify Catalog completeness requirements· Shopify only | 4 |
| The store answers quickly and shows its content promptly | 3 |
Site-wide · 20% of the score
The pages an agent consults before it will recommend a purchase: shipping, returns, who you are, how to reach you.
| Shipping policy is published and substantive | 3 |
| Return/refund policy is published and substantive | 3 |
| Privacy policy is published | 2 |
| Terms of service are published | 1 |
| An About page explains who the business is | 2 |
| Contact options are published and machine-readable | 3 |
| An FAQ answers pre-purchase questions | 3 |
| FAQ content is marked up as FAQPage structured data | 2 |
Product pages · 35% of the score
The product detail pages themselves: structured data, identifiers, description depth, and whether any of it survives without JavaScript.
| Product pages carry complete Product/Offer structured data | 6 |
| Products carry GTIN/barcode identifiers | 5 |
| Product descriptions contain specifications, not just marketing copy | 4 |
| Product descriptions answer what a shopper needs to decide | 4 |
| Reviews are exposed as structured data | 3 |
| Review text is in the page an agent receives | 2 |
| Shipping and returns information appears on product pages | 3 |
| Product markup states shipping and return terms as data | 2 |
| Product images have meaningful alt text | 3 |
| Price, title and description are in the server-rendered HTML | 5 |
| Variant options are named meaningfully | 2 |
| Product markup is free of the storefront's default defects· Salesforce Commerce Cloud only | 3 |
Long-tail pages · 20% of the score
Editorial and collection content that answers the specific questions buyers ask before they are ready to buy.
| High-intent long-tail content exists beyond product pages | 5 |
| Buyer questions an AI assistant would ask are answered somewhere on the site | 4 |
| Blog posts are marked up as articles | 2 |
The 14 crawlers we check for
We read your robots.txt and work out which of these you allow on product, collection and content paths. Blocking an answer crawler removes your store from AI shopping results today. Blocking a training crawler is a defensible business choice, so we note it rather than mark you down for it.
| OAI-SearchBot | OpenAI | answer |
| ChatGPT-User | OpenAI | answer |
| GPTBot | OpenAI | training |
| Claude-SearchBot | Anthropic | answer |
| Claude-User | Anthropic | answer |
| ClaudeBot | Anthropic | training |
| PerplexityBot | Perplexity | answer |
| Perplexity-User | Perplexity | answer |
| Bingbot | Microsoft | answer |
| Applebot | Apple | answer |
| Applebot-Extended | Apple | training |
| Google-Extended | training | |
| CCBot | Common Crawl | training |
| Meta-ExternalAgent | Meta | training |
What a scan costs your store
We fetch public pages only. No login, no forms, no cart, and nothing is changed.
- At most 80 requests for a whole scan, and typically far fewer.
- At most 4 at a time, with at least 200 ms between them.
- One retry, with backoff, honoring
Retry-After. - After 5 consecutive refusals we stop requesting your store entirely.
- We read
robots.txtfirst and obey it. - If your firewall refuses our crawler, the person scanning may read your store through their own browser instead. The same limits and the same robots.txt rule apply, and the report says it was read that way.
More detail, including how to block us, is on our crawler page.
Limits worth knowing
Product checks run against a sample of your catalog, not every product, so those results are an estimate. The sample is deterministic: scanning twice gives the same products.
Some checks have a language model read your pages: which buyer questions your pages answer, which facts your product descriptions state, and what your About page says about who you are. The model can only give credit for what it can point to. Every page it cites as an answer is one we opened and checked, and every fact it credits is a phrase we found on your page. A citation we could not check is left out of the score, not counted for or against you.
Where we cannot identify your content pages or read your language, those checks are marked untested rather than scored zero.
Page speed comes from Google PageSpeed Insights and is skipped when no API key is configured: skipped, not zero.
Read our privacy notice, about our crawler, or ask us to stop scanning your store.