How do diners choose a restaurant now? Restaurant SEO for AI search
By describing the evening they want to Google AI Mode or ChatGPT and letting the AI do the comparing. One sentence carries the cuisine, the party, the day, the budget and the district, and a short list comes back. A restaurant makes that list only when its own site states the menu, a price, the hours, the address and a way to book, as text.
Mr Ibrahim had six people, a Sunday and no booking
Mr Ibrahim had six relatives landing on Saturday and nowhere to feed them on Sunday. Six places at one round table is not a booking he could afford to get wrong, so he skipped the search box and described the evening to ChatGPT:
Where can six of us eat good dim sum in Mong Kok on Sunday, under 400 a head, and can we book?
Mr Ibrahim wanted a table he could announce to the family by lunchtime. ChatGPT gave him a few names, drawn from sites that had written the answers down, and he booked one.
Somewhere in Mong Kok that day, a restaurant was open, had a round table free and served exactly that. If its site never said so, ChatGPT never mentioned it, and nobody called to ask. Would your dining room have been named? Are the kitchens nearby seating those families?
Before, he would have typed three words and rung the top few. Now the diner describes the whole evening. We read 27 restaurant sites to see how many would make Mr Ibrahim's list.
The whole request: six people, Sunday, dim sum, a budget
Google reports that in the US, AI Mode questions run to roughly triple a classic search in length, and the ones that plan around a day and a budget grew 80% faster than AI Mode overall 1. Gemini in Chrome is also available in Singapore 3.
For each clause in Mr Ibrahim's sentence, the AI looks for a site that answers in words. Each row pairs a clause with the fact the site must state and the rubric check involved. Hours and prices are read as text only.
| What the diner wrote | What the site must state | How we read it |
|---|---|---|
| "good dim sum" | The cuisine in the headings: a menu page or heading for it | R2: headings served without JavaScript |
| "on Sunday" | Opening hours by day, in text and in markup | Markup counted; not scored yet |
| "in Mong Kok" | The district, in the address of the business node | EN1: a business declared in markup |
| "under 400 a head" | A price or a range, on the menu | Read as text only |
| "can we book" | A reservation link or a phone, on the page | EE4: more than one contact route |
| "or deliver" | Delivery stated, and the platform, if any | Read as text only |
| The AI itself | The citation crawlers allowed by name in robots.txt | AI2: citation crawlers allowed |
No special dialect is needed. According to Google, its generative AI features call for no special way of writing 2.
Opening hours sit in markup on 11 of 27 sites
"On Sunday" is the clause an owner thinks is obvious and a machine cannot guess. 11 of the 27 sites put their opening hours in markup, none of them in Hong Kong. For a table of six, a wrong guess about the day spoils the evening. I would want every day and its times written out, one line each.
Booking is mostly there; a second way to reach you is not
"Can we book?" is the clause restaurants handle best: 21 of 27 sites offer a booking route. The gap is in the redundancy. On the contact check (EE4), 7 sites pass with two or more routes, 14 warn with one and 7 fail with none. A phone link beside a WhatsApp or booking link is two. Delivery is weaker: 6 of 27 sites say the word.
Singapore declares the restaurant; Hong Kong serves the menu
The two markets fail in opposite places, on a small Hong Kong sample of 8. 18 of 19 Singapore sites carry a Restaurant node, yet only 5 of 19 serve a menu as text. In Hong Kong no site declares a Restaurant node, and 6 of 8 serve the menu as text. Median AI score: 74 in Singapore, 69 in Hong Kong.
The most common failure was R2: 11 of 28 served headings only through JavaScript, so a reader that skips scripts sees an empty page. 25 of 28 declare a business node (EN1), and those had a median AI score of 73 against 58. Every site allowed the citation crawlers (AI2).
What to change, in the order a hungry person would check
- Put the menu in the HTML. Dishes and prices as text (R2).
- Write the hours by day. Sunday and holidays, in text and as openingHoursSpecification.
- Declare the restaurant. One Restaurant block with name, district address, hours and phone (EN1).
- Give every page a booking route. A phone link and a booking or WhatsApp link together (EE4).
- Say what you deliver. Name the platforms and districts.
"Dim sum near me" still works, a little less every year
Ranking for the short search still fills some tables, and each year it fills fewer. US Google searches that ended without a click were 68% in early 2026, against 49% in 2019 4. The family around Mr Ibrahim's table now reads an answer, built from whichever restaurant wrote its facts down.
Reworking a ranking site until it can answer a six-person Sunday request is the work we do. Start with the free scan of your own site, which tells you which clauses it can already answer.
Questions people ask
Does the AI use delivery platforms instead of my website?
It can use both; what changes is who wrote the answer. In our batch only 4 of 27 sites stated their delivery at all, so most leave it to the platforms.
Is my Google Business Profile enough?
It holds your hours, address and reviews, and we cannot measure it from outside. Your own site is the one source you control that any AI assistant can read, so it should state the facts too.
Do I need special markup for AI?
No. Google says structured data is not required for its generative AI features 2. What counts is having the facts stated on the page, in plain words, where the AI can read them.
Should a restaurant publish prices?
Your menu has prices already; only 8 of the 27 sites we read state one in text. Someone asking for "under 400 a head" gets an answer from whichever restaurant gives a figure.
Will this get my restaurant cited by ChatGPT or Google AI Mode?
Nobody can promise a citation, and we do not. The checks show whether an AI can read the facts an occasion needs. Sites that declared a Restaurant node scored a median of 73 on the AI axis against 58. That measures readiness, and says nothing about ranking.
Next: Do you need an llms.txt file?
Sources
- Google, How AI Mode is changing and expanding the way people search, 19 May 2026. Accessed 13 September 2026.
- Google Search Central, Optimizing your website for generative AI features on Google Search, 10 July 2026. Accessed 13 September 2026.
- Google, We are expanding Gemini in Chrome to users in Asia Pacific, 20 April 2026. Accessed 13 September 2026.
- 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.