The problem is what happens away from the table. Someone who has not chosen where to eat yet searches, or asks an assistant. In both cases your menu is not part of the conversation.

What a search engine actually sees

Google does index PDFs — that much is true, and it gets quoted as though it settles the matter. But it indexes them as documents: it pulls out a title, some text, and stops.

The difference is between "this file exists" and "this restaurant serves carbonara at £14, no nuts, until 11pm daily". The second is information an engine can answer a question with. A PDF does not carry it in a machine-readable form — even when the letters are right there.

And if your menu was scanned or exported as an image, even the letters are gone. What is left is a several-megabyte file that means nothing to a machine.

What the guest loses

Search aside, a PDF is a poor thing to hand a phone. It opens in a separate viewer, the page is A4 on a screen that is not, and every guest's first gesture is to pinch.

A menu that is readable on arrival is not a cosmetic improvement. It is the difference between a guest who browses the whole list and a guest who orders the first thing they managed to read.

What a readable menu looks like

Three things, in this order:

1. Its own URL. The menu is a page — yourrestaurant.com/menu — not a file. A page can be linked, shared, indexed and updated.
2. Real HTML text. Dish names, descriptions, prices and allergens are text in the page, not pixels in a picture.
3. Schema.org markup. Restaurant, Menu, MenuItem and Offer tell a crawler which part is a dish and which part is a price, instead of leaving it to guess.

The third is what opens up results for a specific dish rather than only for the name of the place.

And then there is updating it

There is a more mundane reason too. You run out of something. With a PDF: open the document, edit, export, upload, wait for the cache, and the old file carries on living in somebody's phone.

With a page: change the item's status, and the next person to open the menu sees it. The QR code on the table never changes, because it points at an address rather than at a file.

If you are working out how to lay yours out, look at a live menu and then how the platform works.

That is a different game. A search engine ranks sources; an assistant writes an answer and then decides whom to credit. If it has not read you, it cannot mention you.

What an AI crawler actually reads

Briefly: HTML. That is all.

Googlebot executes JavaScript — late, but it does. The crawlers behind most AI assistants do not. They request the URL, take the response, and move on. If your site paints its content in the browser, what they saw was an empty page.

So the first question is not which keywords to use. It is whether the words are in the server's response at all.

Why short sentences win

An assistant assembles an answer out of fragments. A fragment that is true on its own is safe to quote. A fragment that depends on the paragraph above it is not.

Compare:

• "As we saw above, this applies to the lunch menu too."
• "The lunch menu at [the restaurant] is served from noon to 4pm on weekdays."

The second can be lifted with nothing around it. The first cannot.

In practice: phrase the heading as a question, and answer it in the first sentence underneath. Everything after that is detail for the people who want it.

The three files that cost an afternoon

1. A robots.txt that lets the right crawlers in. GPTBot, ClaudeBot, PerplexityBot, OAI-SearchBot, Google-Extended. A blocked agent does not stop answering questions about your neighbourhood — it just answers them without you.
2. llms.txt. A short markdown map written for an agent: what this place is, which pages matter, how to cite it. The convention is new; the file is fifteen lines.
3. FAQPage and Restaurant markup. The questions and answers you already wrote for humans become machine-readable without writing anything new.

And the dullest thing, which works

Put an author's name and a last-updated date on the page. Visible, in the body, not only in a meta tag.

A person and a model, each facing two equally good sources, both take the more recent one with a name under it. That is not a trick — it is the same judgement you would make.

If you are still assembling the foundations, start with your menu — it is the page an assistant will quote first.