Can ChatGPT read your website? Why AI search is a lazy guest

We sent an AI to dinner at 27 restaurants

The restaurants were in Singapore and Hong Kong, the evening was 13 September 2026, and the guest had one job, the job a diner gives it every evening now: find out what each place serves, what it costs, and how to get a table.

Picture it at the door: polite, quick, and with no patience at all. It cannot look at pictures the way you do. It reads. What the restaurant wrote down, it takes in at once; what it did not, for this guest, does not exist.

At most tables, it sat down, asked for the menu, and waited.

Strictly, our guest was a text reader we built to see what such a crawler sees: it opened each homepage and the one menu page it linked, ran no scripts, used no model, and wrote down what was there. The scene is ours; every count below is what it found.

It could book at 21 tables and price a meal at 8

It cannot see. A menu posted as a photo of the printed card is lovely on a phone. To this guest it is a blank sheet: a picture with no words in it. Nine of the 27 homepages linked no menu page at all, so whatever menu they have, on a photo or a social post, the guest never met it in words.

It cannot read. Eighteen homepages did link a menu. Eleven of those pages opened as text the guest could read. Seven did not: the link led somewhere our reader could not open as a page of words. The usual suspects are a menu drawn by script after the page loads, a QR code that only a phone camera can follow, or a PDF to fetch and unpack. Some readers open a PDF and some do not; a menu typed on the page is read by all of them.

It cannot price the dinner. Only 8 of the 27 sites wrote a price for anything, a dish or a drink. Ask the guest what dinner costs, and at 19 tables it has nothing to say.

Booking, though, was easy. 21 of the 27 homepages link straight to a booking page. The guest could reserve a table without knowing what it would eat or pay.

So the guest does what any diner does with a closed menu and a blank bill. It walks down the street to a restaurant that makes it easy, one that wrote the menu and the prices on the page, and tells your customer about that one.

Why the guest moves on instead of asking

A person at a table with no menu calls the waiter. This guest has no waiter to call. It answers a question with a small table of its own, its context window, and it fills that table with whatever arrives in the HTML. It never goes into the kitchen to see what is cooking.

The kitchen is everything that exists but is never served: the menu drawn by JavaScript, the price list inside a photo, the hours inside an image, the document someone has to download. In the wider scan of the same batch, 11 of 28 published sites served their headings only through JavaScript. To a reader that runs no scripts, those pages are a doorway into a dark kitchen.

So are LLMs lazy? Not by design. They read under cost and time limits and take what is on the plate. And they have the whole street to ask. When one restaurant leaves the guest guessing, the next one along has usually written its menu down.

The easy fixes, first to last

None of this needs a new website. It needs the menu and the bill moved from the kitchen to the table. Google tells site owners they need no special style of writing for its AI features, and that markup is optional ¹. In the order I would do them:

1. Type the menu on a page. One heading per section, each dish as text. Keep the photo if you love it, and put the words beside it.
2. Write the prices next to the dishes and the drinks. A plain table of dish and price survives any reader. Even a range of main-course prices answers the budget question.
3. Link that menu page from the homepage, as an ordinary link. Not only a QR code, not only a PDF.
4. Keep the booking link where it is. Most of you already have this part right.
5. Answer the questions diners ask, in text. Vegetarian dishes, a private room, corkage, the last seating.
6. Add llms.txt last, knowing what it is. A curated Markdown index of your site, cheap to make. No major search engine says it reads one, and Google says Search does not ¹². Our llms.txt article has the full case.

What I would not pay for: a rewrite "for AI", or a second, machine-only copy of your pages. Markup names your restaurant; it does not replace the price in the text.

The restaurant article walks the full batch, and the dentist one found the same pattern in clinics.

Fewer diners click through to read your menu themselves

A good listing in the old results still fills some tables, but fewer each year. Across US Google searches, 68% ended with no click in early 2026, up from 49% in 2019 ³. More diners read an answer instead, built from whichever restaurant had its menu and prices written out.

You will not see the guest that walked past. There is no empty chair to count. The free scan on your own site shows what it can read when it sits at your table: about ten seconds, no sign-up.

Questions people ask

Are LLMs really lazy, or is that a metaphor?

A working metaphor. A language model reads under cost and context limits, and the crawlers that feed it often run no JavaScript, so it takes what is served in the HTML. The constraint is real; the word lazy is ours.

Was the guest in your test a real AI?

It was a text reader we built to see a page the way an AI crawler that runs no scripts sees it. It opened each homepage and the one menu page it linked, used no model, and recorded what was written there. The story is a picture; the counts are measured.

Why can it book a table but not read the menu?

A booking link is an ordinary link in the HTML, so the guest finds it at once: 21 of 27 homepages had one. A menu is often a photo, a PDF or a page drawn by script, and a price list even more so. The link is text; the menu often is not.

Is a QR code menu a problem for AI search?

On the table, no: your diners scan it with a phone. On the website, a QR code is a picture of a link, and a reader that reads text gets nothing from it. Put the same menu on a page and link it from the homepage as ordinary text.

How do I know what my site serves to the table?

Run the free scan: it reads your site the way a no-JavaScript reader does and returns both scores plus the three findings that matter most. About ten seconds, no sign-up.

Sources

1. Google Search Central, Optimizing your website for generative AI features on Google Search, 10 July 2026. Accessed 14 September 2026.
2. Google Search Central, Optimizing your website for generative AI features on Google Search, 10 July 2026. Accessed 14 September 2026.
3. SparkToro, In 2026, less than one third of Google searches still send a click, 9 June 2026. Accessed 13 September 2026.

Our own counts come from our restaurant batch of 13 September 2026, 30 sites, one scan and one text read each, scored with the published rubric. Not a market survey.