Most restaurants that score badly are not doing anything wrong on purpose. These are the five problems the scanner finds most often, in the order that moves the score the most.
In our Lisbon pilot about a third of reachable sites refused GPTBot, ClaudeBot or PerplexityBot while serving a normal browser fine, and not one of them had a robots.txt rule saying so. The block comes from a hosting firewall or a "bad bot" list. Run the scanner; if it reports crawlers turned away at the door, ask your host to allow those three user agents.
Agents answer "what do they serve and what does it cost" from whatever they can read. A PDF is read badly, an image not at all, so the answer comes from a delivery app's copy of your menu. Publish the menu as a normal page with dish names and prices as text. Keep the PDF as a download if you like it.
A short JSON-LD block tells every assistant your name, address, phone, opening hours, cuisine and whether you take reservations. Most site builders can add it; the methodology lists the fields. It is the single cheapest fix on this list.
If the booking button goes to a reservation marketplace and the order button to a delivery app, every agent transaction carries their commission. Add a direct path on your own domain: booking software embedded on your site, a simple order form, even a confirmed-by-email request. Direct beats nothing, and it is what the agent will pick when it works.
Some builders render everything in the browser, so a crawler that does not run scripts sees a blank page. Check the scanner's "text without JavaScript" number; under 500 characters means the hours, address and menu need to be in the HTML itself.
Fixed something? Rescan from your restaurant's page; it is free and the badge updates.