Is your restaurant website losing orders?

16 things in three groups. Ordering (4 checks): whether guests can order without a marketplace, whether there is an order button, whether the menu is a web page and whether the phone number can be tapped. AI assistants (6 checks): whether the page's words are in its HTML, and the five files and endpoints assistants use to read a menu and place an order. Website basics (6 checks): https, phone layout, title, description, structured data and how fast the page answers.

Each check carries points: 27 for ordering, 30 for AI assistants and 23 for website basics. The score is the share of points passed out of 100. A check we could not see is left out of the sum rather than failed. A is 85 or more, B 70, C 55, D 40, and anything lower is F.

Guests can ask ChatGPT, Claude and other assistants where to eat and what to order. An assistant can only recommend and order from what it can read. Vercel's study of AI crawler traffic (December 2024) found that ChatGPT's and Claude's crawlers do not run JavaScript, so a page that builds itself with script reads to them as blank.

A plain text file at the root of a website that tells an AI assistant what the site is and where things are. It was proposed by Jeremy Howard on 3 September 2024 and is documented at llmstxt.org. llms-full.txt is the longer version, with the whole content in one file.

MCP, the Model Context Protocol, is an open-source standard for connecting AI applications to outside systems (modelcontextprotocol.io). A restaurant ordering page with an MCP server lets an assistant read the live menu and build an order through tools, instead of guessing from the page.

UCP, the Universal Commerce Protocol (ucp.dev), describes itself as a common language for platforms, agents and businesses. A UCP profile at /.well-known/ucp tells a shopping agent where a business's ordering tools are. When we read ucp.dev on 12 October 2026 it said a food specification was coming soon.

Some home pages arrive as an empty shell and draw everything with JavaScript after they load. Their buttons and phone numbers are not in the HTML we read, so we mark those checks unknown and leave them out of the score. The shell is itself a finding, because AI crawlers read the same empty HTML.

It names a system only from a link or script on the page that goes to that system's own website, and shows the link. It recognises 46 ordering and POS systems from our rankings and the main delivery marketplaces. An ordering site it does not recognise is still counted as direct ordering, named by its address.

Because that is where an order happens. When your home page links to an ordering page, the five AI checks read that page, which is run by your ordering system. A low result there is set by the system's setup, not by your website builder.

No. It reads the same public pages a guest would, plus robots.txt, llms.txt, llms-full.txt and /.well-known/ucp, and tries to open an MCP connection without placing an order. It takes a few seconds and around a dozen requests.

Yes. The report address carries the website, so a copied link opens the report for anyone, checked again when it opens. A result may be reused for up to ten minutes, so a link opened by a whole team does not re-read the site each time. We do not publish a list of the sites checked.

No. No vendor pays for a result here, and a system on the report keeps the place it earned on our rankings from its published prices and terms. Some links to vendor websites may earn us a referral fee; the disclosure page lists which.