Homepage answers an agent
Every agent starts with an HTTP request. Anything other than a fast 200 with HTML ends the visit before it begins.
access.reachable
3/ 19 pts
Methodology
How Legible scores a store.
Twenty-five checks, four pillars, one number. Everything below is the same code that runs your scan, so the weights here are the weights you get.
Scoring
Each check returns pass (1), warn (0.5), fail (0) or info (not scored). A few checks return a partial score, for example Product schema completeness or the share of ACP prerequisites met.
A pillar score is the weighted average of its checks, from 0 to 100. The overall score is the weighted average of the four pillars:
Access
30%
Understanding
30%
Trust
20%
Transaction
20%
Caps
Some failures make everything else irrelevant. When one of these fails, the overall score cannot exceed the cap, however good the rest is. The report says when a cap applied and what the uncapped score was.
| Check | Condition | Max score |
|---|---|---|
| Homepage answers an agent | Fails because the homepage did not answer | 10 |
| No bot wall in front of the store | Fails because a bot wall turns agents away before they see any page | 29 |
| AI search and shopping agents are allowed | Fails because robots.txt blocks the main AI search agents | 44 |
| Product facts are in the raw HTML | Fails because the product page is empty without JavaScript | 59 |
Grades
A
90+
Agent-ready
B
75+
Mostly legible
C
60+
Partly legible
D
45+
Hard to read
E
30+
Mostly invisible
F
0+
Invisible to agents
01 · 30% of the score
Access
Can an agent reach your pages at all?
No bot wall in front of the store
Bot-protection that challenges every non-browser client blocks AI agents even when robots.txt welcomes them. Agents cannot solve JavaScript challenges or CAPTCHAs.
access.botwall
3/ 19 pts
AI search and shopping agents are allowed
Search agents (OAI-SearchBot, Claude-SearchBot, PerplexityBot, Googlebot, Bingbot) build the indexes assistants search before recommending a product. User-triggered agents (ChatGPT-User, Claude-User, Perplexity-User) fetch your page while a shopper waits. Blocking either removes you from the answer.
access.robots-search
5/ 19 pts
Training crawler policy
Training crawlers (GPTBot, ClaudeBot, Google-Extended, CCBot and others) collect content for future models. Blocking them is a legitimate business choice and does not affect AI search visibility. Legible reports your stance but does not score it.
access.robots-training
Not scored
robots.txt is valid
RFC 9309 says a robots.txt that errors with 5xx means "crawl nothing", and an HTML page served as robots.txt is noise. A clean file also points agents at your sitemap.
access.robots-file
1/ 19 pts
XML sitemap lists your products
Crawlers that build AI search indexes discover product URLs from sitemaps far more reliably than from links, especially on large catalogues.
access.sitemap
2/ 19 pts
Fast first byte
User-triggered agents fetch pages while someone waits for an answer, with short timeouts. A slow server is a skipped store.
access.ttfb
2/ 19 pts
Lean HTML
Agents download and tokenise your HTML. Megabytes of inline scripts and styles bury product facts and can hit size caps before the content.
access.weight
1/ 19 pts
Pages are not marked noindex
A noindex meta tag or X-Robots-Tag header tells search agents to drop the page from their index, even when robots.txt allows crawling.
access.indexable
2/ 19 pts
02 · 30% of the score
Understanding
Can it read what you sell without running JavaScript?
Product facts are in the raw HTML
Most AI fetchers do not run JavaScript. If the name, price and description only appear after scripts run, the agent sees an empty page and moves on.
understanding.raw-html
5/ 16 pts
Complete Product structured data
Schema.org Product and Offer JSON-LD is the most reliable way for agents and shopping indexes to read price, currency, stock, identifiers, shipping and returns without guessing from layout.
understanding.product-schema
5/ 16 pts
Organization and WebSite schema
Organization and WebSite markup tells an agent the store's official name, logo, URL and social profiles, so it can tell your shop apart from resellers and lookalikes.
understanding.site-schema
2/ 16 pts
OpenGraph product preview
Chat interfaces render link previews and many agents fall back to OpenGraph when JSON-LD is missing. og:title, og:image and og:type are the minimum.
understanding.opengraph
1/ 16 pts
Canonical URL and language
A canonical URL collapses filter and tracking variants into one product, and a lang attribute tells the agent which language and market the price and policies apply to.
understanding.canonical-lang
1/ 16 pts
One clear H1 with the product name
Extractors use the heading outline to split a page into sections. One H1 naming the product, then H2s for details and reviews, makes the page easy to chunk.
understanding.headings
1/ 16 pts
llms.txt guide for agents
llms.txt is a proposed markdown index of a site's most useful pages. No major assistant has confirmed it uses the file, so it carries little weight here, but it is cheap and gives agents a clean map of products and policies.
understanding.llms-txt
1/ 16 pts
03 · 20% of the score
Trust
Can it verify who you are and what the terms are?
HTTPS everywhere, with HSTS
Agents that hand a shopper off to checkout will not send them to an insecure page. HSTS proves the store never falls back to plain HTTP.
trust.https
3/ 14 pts
Policy pages are linked
Before recommending a purchase, assistants look for shipping costs, the return window and who to contact. In the EU, consumers have a 14-day right of withdrawal for most online purchases (Directive 2011/83/EU, Art. 9), so returns terms are a deciding fact.
trust.policies
4/ 14 pts
Business identity is machine-readable
Agents weigh whether a merchant is real. A legal name, postal address and a contact point in Organization schema are the strongest signals short of a marketplace listing.
trust.identity
3/ 14 pts
Visible price matches structured data
When the price a shopper sees differs from the price in JSON-LD, agents quote the wrong number or distrust both. Google Merchant Center suspends listings for mismatches, and under EU consumer law the price shown must be the price charged.
trust.price-consistency
4/ 14 pts
04 · 20% of the score
Transaction
Can it start a purchase through a protocol?
UCP business profile
The Universal Commerce Protocol (UCP), announced by Google in January 2026, lets agents discover a store's capabilities from /.well-known/ucp and complete checkout through a standard API. It powers buying in Google AI Mode and Gemini. Current spec version: 2026-08-25.
transaction.ucp
5/ 11 pts
Machine-readable catalogue
Agent channels (ChatGPT shopping, Google AI Mode, Copilot) prefer a structured product feed they can refresh over crawling pages. A feed keeps price and stock current between crawls.
transaction.catalogue
3/ 11 pts
Agentic Commerce Protocol readiness (signal)
OpenAI and Stripe's Agentic Commerce Protocol (ACP) launched in September 2025 to let shoppers buy inside ChatGPT. There is no public discovery file: merchants are onboarded through OpenAI and submit a product feed. In March 2026 OpenAI moved checkout back to merchants' own sites, so discovery in ChatGPT now depends on that feed and on crawlable pages. Legible scores the prerequisites, not enrolment.
transaction.acp
1/ 11 pts
Add-to-cart works without scripts
Browser-driving agents (ChatGPT agent, Gemini in Chrome and similar) click through your real checkout. A plain form or link for add-to-cart in the server HTML is the most robust path when no protocol is available.
transaction.cart-path
2/ 11 pts
Platform and remediation path
What you can fix, and how fast, depends on the platform. Shopify merchants can switch most of this on; WooCommerce, Magento and custom stores need it built or configured.
transaction.platform
Not scored
AI agents evaluated in robots.txt
Legible applies RFC 9309 matching (named group first, then *, longest rule wins, allow wins ties) to the homepage and the product page for each agent. Search and user-triggered agents are scored; training crawlers are reported only. List checked against vendor documentation on 2026-10-07.
| Token | Vendor | Type | Purpose | robots.txt |
|---|---|---|---|---|
| OAI-SearchBot | OpenAI | Search / index | Indexes pages so they can appear in ChatGPT search and shopping results. | Honours robots.txt |
| ChatGPT-User | OpenAI | User-triggered | Fetches a page when a ChatGPT user or custom GPT asks about it. | May ignore robots.txt |
| GPTBot | OpenAI | Training crawler | Collects public content that may be used to train OpenAI models. | Honours robots.txt |
| Claude-SearchBot | Anthropic | Search / index | Indexes pages to improve the quality of Claude search results. | Honours robots.txt |
| Claude-User | Anthropic | User-triggered | Fetches a page when a Claude user asks a question that needs it. | Honours robots.txt |
| ClaudeBot | Anthropic | Training crawler | Collects public content that may be used to train Claude models. | Honours robots.txt |
| PerplexityBot | Perplexity | Search / index | Indexes and links pages in Perplexity answers. Perplexity says it is not used for model training. | Honours robots.txt |
| Perplexity-User | Perplexity | User-triggered | Visits a page live to answer a user question. Perplexity says it generally ignores robots.txt. | May ignore robots.txt |
| Googlebot | Search / index | Google Search crawler. Google AI Mode, AI Overviews and Shopping draw on its index. | Honours robots.txt | |
| Google-Extended | Training crawler | Control token, not a crawler: opts content out of Gemini training and grounding. No effect on Google Search. | Control token | |
| Bingbot | Microsoft | Search / index | Bing index crawler. Microsoft Copilot answers draw on the Bing index. | Honours robots.txt |
| Applebot-Extended | Apple | Training crawler | Control token: opts content out of Apple foundation-model training. Applebot itself still crawls for Siri and Spotlight. | Control token |
| Amzn-SearchBot | Amazon | Search / index | Improves search in Amazon products such as Alexa. Amazon says it is not used for model training. | Honours robots.txt |
| Amazonbot | Amazon | Training crawler | Improves Amazon products and services and may be used to train Amazon AI models. | Honours robots.txt |
| meta-externalagent | Meta | Training crawler | Crawls content for training Meta AI models and indexing for Meta products. | Honours robots.txt |
| CCBot | Common Crawl | Training crawler | Builds the open Common Crawl dataset that many models are trained on. | Honours robots.txt |
| Bytespider | ByteDance | Training crawler | Collects content for ByteDance AI models. | Honours robots.txt |
Limits and ethics
- One product page. Legible reads the homepage and one product page. Category pages and variants are not crawled. Point the scan at your best seller for the most useful result.
- No JavaScript. Deliberate: it mirrors how most AI fetchers read pages. Google's crawler renders JavaScript, so Gemini may see more than Legible reports.
- From the outside. ACP enrolment, Merchant Center feeds and private APIs cannot be verified from public URLs. Those checks are signals, and they say so.
- Polite by design. At most 15 requests per scan, never more than two at a time, 8-second timeouts, 2.5 MB per response, an identifying user agent, and a rate limit per visitor.
- Safe by design. Only public http(s) addresses on standard ports. Every redirect hop is re-validated and its DNS checked; private, loopback, link-local and reserved addresses are refused.