19 restaurants tested. Not one could answer a question about itself.
Between July and August 2026 we ran a full AI visibility audit on 19 South African restaurants. Every one was scored out of 100 and tested live against ChatGPT and Gemini.
The average was 43.7 out of 100. The best restaurant scored 66. Not one reached the level where an AI assistant can reliably find, describe and recommend it.
The result that stopped us was the one that came back identical every single time. We score question coverage out of 10, meaning whether a site answers the plain questions customers actually ask. Do you take bookings? Are you open on Sunday? Is there parking? Can you seat a highchair?
Average question coverage across 19 restaurants. Not low. Zero. All nineteen of them.
The strange part
These are not neglected businesses. Reputation was their strongest category by a distance, averaging 7.9 out of 10, better than the wider group of 88 businesses we tested. These restaurants have real reviews, genuine press coverage, and in one case an international award.
So the picture is a restaurant that is well known, well reviewed, and completely silent the moment an AI is asked something practical about it.
Fourteen of the nineteen had the front door wide open to AI systems and nothing behind it. Seven did not have their opening hours, address or phone number written anywhere on the site as text a machine could actually read.
Why this costs you a table
Someone visiting the West Coast no longer opens twelve browser tabs. They ask an assistant where to eat tonight, and they get back one answer with two or three names in it.
To be one of those names, the assistant has to be able to say something concrete about you. If your site cannot confirm that you are open on a Sunday evening, you are a risky recommendation, and a risky recommendation gets left out.
Your reviews do not rescue you here. Reviews tell an AI that people liked you. They do not tell it whether you are open tonight, whether you take walk-ins, or whether the kitchen closes at nine.
It is not your website's fault
Eighteen of the nineteen were built on WordPress, and most of them looked good. This is not about a bad website, and it is not a criticism of whoever built it.
The rules changed recently and quietly. A site built to impress a human visitor is built differently to one built to be read and quoted by a machine. Almost nobody has caught up yet, which is exactly why an entire sector is sitting in the 40s.
The good news
The three weakest areas across all 19 restaurants are also the three most fixable. None of them require redesigning your site or changing how it looks to your customers. They are additions to what is already there.
The gap between 43 and 80 is not a rebuild. It is the difference between a site that looks good and a site that also answers questions.
This article draws on our wider study of 88 South African businesses tested for AI visibility. No business is named, and all figures are published as aggregates.