I asked ChatGPT who the best plumber in Manchester is, 8 times in a row. Across the 8 answers it cited 75 sources. Not one of them was a plumber’s own website.
That’s the result that changed how I think about GEO for local businesses, so here’s the whole test, including the parts that make it less tidy.
How we ran it
On Sunday 18 October 2026 we sent the same question, “Who is the best plumber in Manchester?”, to ChatGPT 8 times. We used gpt-4o with web search switched on, through the API (via DataForSEO), so every run was a clean session with no history.
That’s worth saying up front. The API isn’t the ChatGPT app, and the app can behave differently. But it’s the closest repeatable version of “a stranger asks ChatGPT for a recommendation” that I know how to run.
Then we counted every citation in every answer and grouped them by domain.
Where the 75 citations came from
| Source | Citations | In how many of the 8 answers |
|---|---|---|
| Rated Traders | 32 | 8 of 8 |
| Checkatrade | 12 | |
| TrustATrader | 8 | |
| MyBuilder | 4 | |
| 3 | ||
| Trustpilot | 1 | |
| 8 other list sites and blogs | 15 | |
| Plumbers’ own websites | 0 |
Only 17 unique URLs across all 75 citations. ChatGPT kept going back to the same handful of pages.
23 plumbers got named across the 8 answers. Only 2 were named in all 8, and both of them were on Rated Traders’ list. I’m not naming them, because this isn’t about them.
Same thing with lawyers
One city and one trade is a small sample, so three days later I ran the same test on a lawyer query in Sydney. Same setup, 8 runs.
89 citations. Not one law firm’s own website.
61 of the 89 came from a single list site, sydney10.com, and it appeared in every answer. A second list site was in all 8 too. Three firms got named in all 8 answers, and all three were on that first list.
One detail from the lawyer runs bothered me more than the headline. In every run, the search ChatGPT actually sent out had “2023” tacked on the end. It went looking for a three-year-old list and built its answer on whatever ranked for that.
Why this happens
For a “who’s the best” question, ChatGPT doesn’t reason its way to an answer from your website. It searches the question, reads whatever ranks, and summarises it. And what ranks for “best plumber in Manchester” is lists, directories and review sites.
So your homepage calling you “Manchester’s leading plumber” is your own testimony. The lists are the witnesses, and the model goes with the witnesses.
This is also where link building and GEO turn out to be the same job. Getting onto the pages that rank for your “best X in Y” searches is outreach and digital PR, the same work we’ve been doing for link building for years. The difference is that a mention on the right list now does two things: it can send a link, and it puts you in the answer.
What to do this week
You can run a rough version of this yourself in half an hour.
- Write down the questions your customers ask. “Best [service] in [town]”, “who should I call for [problem]”, “[service] near me reviews”.
- Ask ChatGPT each one five to eight times, in fresh chats with web search on. One answer means nothing. You’re looking for what repeats.
- List every domain it cites. You’ll probably see the same three or four sites over and over.
- Check whether you’re on those pages. Some are free to join, some run on reviews, some are paid placements. Be honest with yourself about which ones are worth it.
- Track mentions, not just rankings. A rank tracker won’t tell you that you’re missing from the one list ChatGPT reads every time.
And look at the dates on those lists. If ChatGPT is pulling a list from 2023, whoever publishes the 2026 version that ranks gets read next.
The limits
Two questions, two cities, one model, one day each. The API isn’t the app. I’d expect broader questions (“how much does a boiler service cost?”) to cite very different sources, and big brands to look different from local trades. So treat this as a reason to run your own check, not as a rule.
But if a customer asked ChatGPT about your service today, who would it call as a witness?
First shared as a shorter post on LinkedIn.
