New: sites with an MCP server are now scanned and scored through it.See what’s new
Methodology

How the Axidort Score is calculated

The score measures one thing: whether an AI purchasing agent can find, price and buy from your website. It is based on what agents do on your site, not on how your site looks.

The agents

Every scan runs four agents and a scorer.

  • Static check agent. Fetches robots.txt, sitemap, llms.txt, agent and MCP manifests and the home page, and runs twelve checks. It also lists the agent standards the site publishes (Link headers, DNS-AID records, Markdown for agents, Content Signals, Web Bot Auth, API catalog, MCP server card, A2A Agent Card, Agent Skills, ARD, Poppy, OAuth discovery and auth.md, ACP, UCP, AP2, MPP and x402; WebMCP tools come from the shopper’s browser). These are reported for information and do not change the score. A request the site refuses is reported as “couldn’t verify”, not as missing, and a file counts only when it has the fields its specification requires.
  • Site classifier. Reads your home page as a browser shows it, with its menu links, and decides whether the site is an online store, a digital product or subscription site, or an information site. A site that offers a free tool but sells paid plans is a digital product site. It also chooses a realistic task, such as finding a specific product, a paid plan, or a page the menus link to.
  • Shopper agent. A browser agent that searches or navigates, opens a product or offering, reads the price and stock, adds to cart and goes as far as the checkout or sign-up page. It is limited to fifteen steps and a few minutes. On the product page the scan also reads the rendered page itself (structured data, and the price and stock next to the product’s name), and asks the agent once to record the price and stock if it has not, so these facts do not depend on how one AI model behaves.
  • MCP agent. When a site publishes an MCP server (a server card or MCP manifest under /.well-known/), the agent shops through the server’s tools instead of the website, the way an AI agent would: it searches, reads price and stock, and adds to a cart or opens a checkout. It never calls a tool that pays, places or confirms an order or changes an account, and never sends personal or payment details. If the server cannot be used (it needs a login, offers no shopping tools, or leads to nothing to buy), the browser agent shops the website instead and the report says why.
  • Reader agent. Fetches the same page without a browser, the way many AI assistants read the web, and checks what is machine-readable.
Safety. The agent never places an order, pays, logs in or creates an account, never enters personal or payment details and never submits contact, quote or newsletter forms. Clicks and requests that could do any of those are blocked before they run.

The seven measures

The agent-measured components carry 90% of the score. Static signals carry 10%, because a file on a server matters far less than whether an agent can actually buy.

MeasureWeightMeasured byHow it is scored
Product discoverability20%Browser or MCP agent100 when the agent reaches a product in three steps or fewer, through site search or through the site's links. 70 in four to eight steps. 40 when it takes more than eight steps, or the product can only be reached through the sitemap. 0 when no product is found.
Checkout reachability20%Browser or MCP agent100 when the agent reaches checkout without needing a login. 75 when login or an account is required first, or when the agent reaches a digital plan’s buy button and stops there for safety, so the steps after it are not tested. 50 when the item is in the cart but checkout is blocked. 20 when a buy option exists but cannot be used. A CAPTCHA on the way to checkout costs 20 points. An order summary counts as checkout only when it asks for an address or payment. Left out when the agent ran out of steps or time before it could try to buy, since nothing was learned about checkout. Through an MCP server: 100 when the agent opens a checkout with the server’s tools, 90 when it reaches the purchase tool and stops there. A store’s direct Buy now button, and a crypto wallet’s Connect wallet step, count as reaching the purchase step (75). Paying: when every payment option needs a person to approve it, checkout is capped at 50 for a crypto wallet and 60 for wallets such as Apple Pay or PayPal with no card or invoice option; an agent payment protocol (x402, AP2, ACP, UCP or MPP) lifts checkout to at least 90.
Product data quality15%Reader agentReads the page without a browser. Counts machine-readable name, description, image, brand, price, availability and SKU or GTIN, shipping cost and returns (in structured data or stated on the page), and whether the content is readable without JavaScript. On a store or digital product the name must come from the product's own structured data, not from page-level data. Only the parts that apply to the kind of site are counted. When the agent bought through an MCP server, the offering data its tools return (name, price, availability, ID, description, image) counts as machine-readable.
Pricing visibility15%Both agents100 for a clear price with currency and unit. 60 when the currency, unit or billing period is unclear; a price that states its unit or period ("$29 / month", "£12 / box of 10") is never counted as unclear. 30 for B2B sites that gate price behind login or a quote. 0 when no price is found. A price counts only when it is on the product page: next to the product's name, never a promotion such as "$10 off", and a price the agent reports must appear on the page it was reading, written as a price: a rating, a size or a review count is not one.
Stock availability10%Both agents100 when stock appears as text and in structured data. 70 for either one. 40 when it is shown only by an icon or colour. 0 when it is not shown. Stock text counts when it is next to the product's name, or when the agent reports it and it appears on the page. It must say something about stock: "Available in 3 colours" does not.
Agent accessibility10%Both agentsStarts at 100. Loses 50 for a bot-protection challenge, 30 when the site rejects an add-to-cart request, 20 for a cookie or pop-up wall, and 15 each for content that needs JavaScript or a very slow load (the page content takes more than eight seconds). A barrier counts only when the page shows it, never on the agent's word alone. On a store's product page it also loses 20 when the buy button is not a real button or link, 20 when it has no name an agent can read (an icon alone), and 10 when more than three buttons or links on the page have no name.
Static signals10%Static checksShare of the scored checks passed: robots.txt rules for AI search crawlers and for the agents that browse for a person (blocking AI training crawlers is reported, not scored), sitemap, llms.txt, structured data, MCP or agent manifests, API documentation, HTTPS and semantic HTML. The product feed check is reported only: stores send their feeds to ChatGPT and Google privately, where no scan can see them. A check the site refused to answer shows as "couldn't verify" and is left out, since the refusal already counts under agent accessibility.

The score adapts to the site

Not every site sells physical goods with a cart, so not every measure applies to every site. Steps that do not apply are shown as N/A, left out of the score, and their weight is shared by the measures that do apply. Static signals always stay at 10%.

  • Online stores (B2C and B2B) are scored on all seven measures. A store with a cart is scored on adding to it and reaching checkout; a store with only a direct Buy now button is scored on reaching that button. If a store has no cart, buy button, sign-up or quote flow at all, checkout is left out.
  • Digital products and subscriptions (software, plans, downloads, courses, tickets, and tokens or other digital goods paid for with a crypto wallet) have no physical stock, so stock is left out. Reaching the sign-up, purchase or wallet step counts as reaching checkout.
  • Information sites sell nothing online, so price, stock and checkout are left out. They are scored on whether an agent can find and read the key content. A quote or contact form does not count as selling. If the scan finds a working cart, buy or sign-up button on the site, price and checkout are scored after all. Otherwise the score is labelled for content (for example “Easy for agents to read”) and is not ranked against the commerce sites of the Axidort Index.

A cart or buying flow means a real button, link or form: a page that only writes about carts or checkouts does not have one. If the agent finds a cart or buying flow anywhere on the site, checkout is scored. If the agent is blocked or never finds anything to judge, the measures stay in the score at zero, so a failed scan cannot look like a site with nothing to buy. If a site refuses to ship to the scanner’s location, checkout is left out and the report says why.

Paying

Reaching checkout is not the same as being able to pay. The scan records how the site takes payment, from its payment buttons, logos and payment-provider scripts on the product, checkout and wallet pages. A card (agents can use delegated or virtual cards) and invoice or purchase order are routes an agent can complete. A crypto wallet, Apple Pay, Google Pay, PayPal and similar wallets ask a person to approve the payment on their own device, so when they are the only options, an agent cannot finish the purchase and checkout is capped. An agent payment protocol such as x402 or AP2 lets an agent pay without a browser, and counts in the site’s favour.

Grades

AAgent-ready85 to 100
BMostly ready70 to 84
CPartially ready50 to 69
DAgent-hostile30 to 49
FInvisible to agents0 to 29

How the Axidort Index is compiled

The Axidort Index ranks well-known B2B and B2C commerce sites by Axidort Score. Each site was scanned with the same agents and the same scoring as a scan you run yourself.

  • Declared agent. The scanner identifies itself as AxidortAgentCheck. Sites that treat declared agents differently from people will score lower, which is part of what the score measures.
  • Two scans. The Index published now shows, for a site that scored under 70, the better of two scans. From its next rebuild every site is scanned twice and the average is shown, so it compares fairly with a report, which is one scan; the two scans also give the run-to-run range shown on every report.
  • One location. All scans ran from a single data-centre location. Sites that vary by region may behave differently elsewhere.
  • A point in time. Scores are dated. Sites change, and so do their bot defences. Scan your own site for a current result.
A low score is not a verdict on a business. It means an agent met a barrier on that site at that time, for example a bot challenge or a rejected add-to-cart. If you represent a listed company and believe a result is wrong, contact us through the details on the agent page and we will rescan and correct it.

Limits

  • The shopper agent is a language model, so the same site can score slightly differently between runs. Rescan to confirm a change, and use the movement over several scans rather than a single run.
  • Pages behind a login are not scanned. B2B portals are measured on what an agent can see and do without an account.
  • The score does not measure price, product range, delivery quality or customer service.

See your own score

It costs nothing, and the first call with our team is free.

Free scan Results in about three minutes Never places an order Sign in with Google to run your scan