Before they visit, they ask the AI

1The scene: a question you never heard

Composite example, not a specific customer

Wednesday night. Noura is organising dinner for ten friends from university. This time she doesn’t post in the family group or call anyone. She opens ChatGPT and types: “Quiet restaurant in north Riyadh, outdoor seating, room for ten, open after midnight.”

She gets three names. Your branch isn’t one of them, even though you have exactly what she asked for: outdoor seating, a table for ten, open until 2 am.

In a worse version of the same night, your name does appear, but the answer says you close at 11. Those were last Ramadan’s hours, and nobody updated them.

Either way, you’ll never know. No missed call, no cancelled order, no angry review. A customer who didn’t come and left no trace.

2The number: 45.2%

According to the Saudi Internet Report 2025, released by the Communications, Space and Technology Commission (CST) in July 2026, AI-tool adoption among internet users in Saudi Arabia reached 45.2%, more than double the previous year.¹ Among 20–29-year-olds it reached 55.7%, and the most-used apps were ChatGPT, then Gemini, then DeepSeek.²

On 20 May 2025, Google brought AI Overviews to the Middle East and North Africa, and in Arabic.³ Someone searching Google in Arabic may now read the answer before they see a single link, including yours.

I want to be precise here. That number measures AI-tool use in general, not restaurant choice. I found no published Saudi figure for how many diners ask an assistant before deciding where to eat, and I won’t invent one. But when close to half of internet users use these tools, the question is no longer “do they ask it?” It’s “what does it say about you when they do?”

3The image: “Ask the one who’s tried it”

There’s an old Arabic saying: ask someone who has tried it, not someone who is wise. It still holds, and your customer still lives by it. What changed is who “the one who’s tried it” is.

It’s no longer the neighbour who ate at your place last week. It’s a machine that has read what thousands of those people wrote in reviews, and has read what you published too: your hours, your menu, your photos and the category you picked for yourself on the map. Then it sums it all up in three lines.

Nobody outside these companies knows exactly how the answer is chosen. But it most likely doesn’t make up your hours. It takes them from what’s published. If that’s old, incomplete or contradictory, it repeats the old or incomplete version, or leaves you out.

So the saying still holds, but it now points the other way. The one who’d tried it used to talk about you, and you had no say in what they said. Today, half of what the new one says comes from you.

4The map: where does your branch data live?

Take one branch and draw this map on paper.

In most groups I’ve seen, every cell in the last column holds a different name, or a question mark. The doctrine has a rule: “a metric with two owners has no owner.”⁴ Data works the same way. Hours that five people edit belong to no one.

The fix isn’t a new tool. It’s a management decision. Branch data is one asset, in one master file, with one named owner. Every change starts in the file and flows out to every other place, never the other way round. Ramadan and Eid hours go into the file two weeks before the season, not on the night.

5Where the money leaks

In the first episode of season one we named “the closed door”: an ad that brings a customer to a branch that can’t serve them.⁵ Wrong hours in an AI answer are a worse kind of closed door. It shuts before the customer leaves home, and it shows up in no report.

The second leak is category. If your branch is listed only as “restaurant”, and your customer asks for “mandi for twenty” or “family Friday breakfast”, you don’t exist for the occasion where you sell the most. Customers ask by occasion, not by your name.⁶ These are the entry points we covered in the occasions episode.⁷

The third leak is one I expect, and this is my reading, not published news: the day may come when an assistant does more than answer, and books or orders on the customer’s behalf. If that day comes, the assistant won’t ask whether your menu is delicious. It’ll ask whether your menu is machine-readable, your prices right and your hours accurate. A branch whose data can’t be read won’t get the order in the first place.

6The Sunday question

The twenty-question test

The test is a thermometer, not a lab. Answers vary by account, location and day, so use the same device, the same wording and the same day each month. Month one is a baseline, not a verdict. And don’t trust a global figure for “what percentage discover restaurants through AI”. Your own number, after three months, beats any imported one.

8Credit, and what’s next

The number is CST’s, and the Arabic launch is Google’s. The idea that customers enter a category through an occasion rather than a brand name comes from the Ehrenberg-Bass Institute’s research.⁶ My part was sitting between the three and drawing the map for a branch in Riyadh.

Which leaves one question: who in your group owns the decision to appoint that owner? If the answer is “me, but I haven’t found the time”, next episode is a letter to you.