Every business owner we speak to lately asks some version of the same question. “Is AI going to name my business?”
So we tested it. We asked ChatGPT to recommend plumbers, electricians, attorneys and dentists across five South African towns in four provinces, and we counted what came back.
It named 46 real businesses. Most of them small independents, not national chains. On the face of it, that is good news.
Then we looked at where it sent the customer, and the answer turned out to be the thing that actually matters.
What we tested
On 14 September 2026 we ran nine queries through ChatGPT with web search switched on. Every query used the same wording: “Who is the best [trade] in [town], South Africa? Name specific businesses.”
The towns were Paarl and George in the Western Cape, Sandton and Centurion in Gauteng, and Umhlanga in KwaZulu-Natal. The trades were plumber, electrician, attorney and dentist.
For every business the model named, we recorded two things. Whether it linked to that business’s own website, and whether the business was actually in the town we asked about.
Half of them had nowhere to send the customer
Of the 46 businesses recommended, 23 were linked to their own website. That is 50%.
The other half were linked to a generic Google Maps search, a directory listing, or nothing at all.
This is the part worth sitting with. If the AI recommends you and then links to a Maps search for your trade in your town, the customer does not land on your website. They land on a page showing you next to every competitor you have, with their review counts visible alongside yours.
You were recommended, and you still lost the customer at the last step.
Two queries produced no business websites at all. Asked for electricians in Centurion, the model named eight businesses and linked every single one to a Maps search. The same happened with plumbers in Umhlanga. Sixteen businesses named across two provinces, and not one of them had a site the model could point at.
More than a quarter were in the wrong town
13 of the 46 businesses recommended were not in the town the question asked about.
We asked for the best plumber in Sandton. The answer included a plumber in Cosmo City, Roodepoort, and another in Germiston. Both roughly 30km away, in a city where 30km can be an hour of traffic.
The same thing happened in Paarl. Asked for electricians, the model returned a business in Wellington. Asked for plumbers, it offered one based in Tafelsig, Cape Town, most of an hour down the N1.
The mechanism is not mysterious. The model widens its search until it finds something it can describe with confidence. If nothing in your town gives it a strong enough signal, somebody a district away gets the recommendation that should have been yours. Neither of you ever finds out it happened.
The gap is not geography. It is your trade.
We expected the big cities to do better than the small towns. They did not. George, population well under 200,000, returned three plumbers and linked all three to their own websites. Sandton managed one in four.
The split runs along a different line entirely.
Attorneys and dentists were linked to their own websites in eight of nine cases. Asked for the best attorney in Paarl, the model named five firms, linked four of them, and placed all five correctly in Paarl. Asked for a dentist in Centurion, it named four practices, linked all four, and got every location right.
Plumbers and electricians managed fifteen out of thirty-seven.
That is the finding we did not expect and the one that matters most, because it says this is not about how big your town is or how much you spend. It is about whether your sector has a habit of building proper websites.
Why this happens
An AI answer is assembled, not remembered. When it makes a claim about your business, it wants a source to attach to that claim. It links to whatever backs it up.
If you have a real website with a page about the service being asked about, that page is the obvious thing to cite, and the customer arrives on your site. If all you have is a Google listing, the only citable thing left is your Maps entry, and that is where the customer goes.
A law firm almost always has a website with a page for conveyancing and a page for litigation, because that is simply how firms present themselves. A great many plumbers and electricians run on a Facebook page, a Google listing and word of mouth. That was completely survivable when people searched, scrolled and phoned.
It is not survivable when a machine is doing the choosing and it needs something to cite.
What to do about it
None of this needs a new strategy. It needs the same fundamentals that win Google Maps, because AI answers are assembled from the same signals.
Give each service its own page. One page called Services, listing eight things you do, gives an AI nothing specific to cite when someone asks about one of them. Eight pages gives it eight chances.
Write in the words customers use. Not your industry’s words. Somebody types “burst geyser”, not “emergency hot water cylinder failure”.
Finish your Google Business Profile and keep the reviews coming. It is the thing the model falls back on, and it is how it decides you are real. If you want the detail on that, we have written about how many reviews you actually need.
Then go and check. Open ChatGPT, type your own service and your own town, and read what comes back. It takes thirty seconds and most owners have never done it. Either you are in the answer with a link to your own site, or you have just found out what your next job is.
Method and limits
Nine queries run through ChatGPT, model GPT-4o-mini, with web search enabled and temperature set to zero, on 14 September 2026. Every business named in each response was counted, along with whether the response linked to that business’s own domain, a directory, a Google Maps search, or nothing.
This is a first sample, not a survey. Nine queries and 46 recommendations is a small base, results differ between AI models, and answers shift as the underlying web index changes. We are publishing the method so it can be repeated, including by anyone who thinks we are wrong. We intend to run it again across more towns and more trades.
Frequently asked questions
Does AI actually recommend small local businesses in South Africa?
Yes. In our test ChatGPT named 46 real South African businesses across five towns, most of them small independents rather than national chains. Being named is not the problem. Half of those businesses had no website the AI could link to, so the customer was sent to a Google Maps search showing them alongside every competitor in the area.
Why does AI link to Google Maps instead of my website?
Because it could not find a page of yours worth citing. AI answers link to whatever source backs up the claim. If your business has a Google listing but no real website, or a website with nothing on it about the service being asked about, the only citable thing left is your Maps entry. You get the mention and your competitors get the shop window.
Why did AI recommend businesses from a different town?
It widens the net when the local signals are too weak. In our test, a question about the best plumber in Sandton returned businesses in Roodepoort and Germiston, roughly 30km away. If nothing in the town gives the model enough to work with, it will reach further out until it finds something it can describe confidently.
Why do plumbers and electricians do worse than attorneys and dentists?
Website norms differ by sector. In our sample, attorneys and dentists were linked to their own sites in eight of nine cases, while plumbers and electricians managed fifteen out of thirty-seven. Professional practices almost always have a proper site with service pages. A lot of trades run on a Facebook page and a Google listing, which was survivable when people searched and scrolled, and is not when a machine is doing the choosing.
How do I get my business cited by AI instead of just mentioned?
Give it something to cite. That means a real website with a separate page for each service you sell, written in the words customers actually use, plus a complete Google Business Profile and reviews that confirm what you do. AI answers are assembled from the same signals that decide Google Maps rankings, so the work that wins one tends to win the other.
Is this test definitive?
No, and we would not present it as one. Nine queries and 46 business recommendations is a first sample, not a survey. Results vary between AI models and change as the web index changes. We published the method and the raw numbers so anyone can repeat it, and we intend to run it again across more towns and trades.