How ChatGPT decides which businesses to recommend (and what you can influence)

Ask ChatGPT "who is a good emergency electrician in Sheffield?" and you get a short, confident list. No ads, no ten blue links, no page two. For a business owner the obvious question is: how did those names get there, and why not mine?

There is no published ranking algorithm, and anyone who tells you otherwise is guessing. But the mechanics are well enough understood to be useful. ChatGPT builds a recommendation from two layers, and each layer responds to different things you control.

Layer one: what the model already believes

Every version of ChatGPT is trained on a large snapshot of the web, books and licensed data, frozen at a cut-off date. From that it forms an impression of each entity it has seen often enough: what the business is, where it is, what it is known for, and how people talk about it.

Three things determine whether you exist in that layer at all.

Repetition across independent sources. A business mentioned on its own site and nowhere else is a weak signal. The same business described consistently on its site, Google Business Profile, Companies House, a trade body register, two review platforms and a local newspaper is a strong one. The model is not counting links; it is learning that these facts co-occur.

Consistency of facts. If your trading name is "Northern Sparks Ltd" on Companies House, "Northern Sparks Electrical" on your website and "N. Sparks Electricians" on a directory, the model may treat those as three weak entities instead of one strong one.

Age. Anything that happened after the training cut-off is invisible to this layer. A rebrand, a new branch or a new service line can take months to be reflected, which is why the second layer matters so much.

Layer two: live retrieval

When a question looks like it needs current or local information, ChatGPT runs a web search, reads a handful of pages, and writes its answer from those. This is the layer most businesses can influence quickly, because it works more like a search engine.

The retrieval step tends to favour:

• Pages that are reachable by OAI-SearchBot and ChatGPT-User (check your robots.txt and firewall).
• Pages whose first paragraph states the answer plainly: who, what, where, for whom, what it costs.
• Sources the system already treats as reliable for that kind of question: review sites, directories, local press, industry bodies.
• Pages that render as HTML rather than relying on client-side JavaScript.

If retrieval finds a directory listing that says you are closed on Saturdays and your own site says you are open, the answer may well repeat the directory. The model cannot tell which is right; it picks what is corroborated or what it read first.

How the two layers combine

The answer you see is a blend. The model drafts a list of candidates from what it believes, checks and extends it with what it retrieved, and then writes a short justification for each name. Those justifications are a useful tell: they usually echo the phrases the model found. "Highly rated for same-day call-outs" means a review source or your own page used language close to that.

Because the process involves sampling, the same prompt can produce a different list on a second run. Any one answer is a draw from a distribution, which is why Seoptist asks each prompt three times per engine every week and reports the proportion of answers that mention you rather than a single yes or no.

What you can influence

Roughly in order of effort against effect:

1. Let the bots in. Allow OAI-SearchBot and ChatGPT-User in robots.txt and confirm your CDN is not challenging them.
2. Make one page per service that answers the question in its first 100 words. "We are a NICEIC-registered electrician covering Sheffield and Rotherham, available for same-day emergency call-outs, with a fixed call-out fee of £X." Specific, factual, quotable.
3. Align your entity facts everywhere. Same name, address, phone, hours and services on your site, Google Business Profile, Companies House, directories and review sites.
4. Add structured data. LocalBusiness or Organization JSON-LD with the same facts, plus FAQPage where you have real FAQs.
5. Earn presence on the sources the model already cites. Look at which domains appear in answers for your prompts and make sure you are listed, reviewed or mentioned there.
6. Keep third-party information current. Stale directory entries are the most common cause of wrong recommendations.

What you cannot influence

• The training cut-off. You cannot make the model learn a fact faster than its next release.
• Which sources the system trusts. You can be present on them; you cannot change the list.
• The sampling. You can raise your probability of being mentioned; you cannot guarantee it.
• Paid placement. There is none in the organic answer, and claims to the contrary are usually a scam.

A quick self-check

Before you buy anything, do this in an afternoon:

[ ] robots.txt allows OAI-SearchBot and ChatGPT-User
[ ] curl -A "OAI-SearchBot" returns 200 and real HTML for your top 5 pages
[ ] Name, address, phone identical on site, GBP, Companies House, top 3 directories
[ ] Each service page answers who/what/where/cost in the first paragraph
[ ] LocalBusiness or Organization JSON-LD validates with no errors
[ ] Asked ChatGPT 5 realistic prompts 3 times each; noted who it named and which sources it cited

The last item is the one most people skip, and it is where the surprises are. The free AI Visibility Check does that part for you across five prompts and two engines, and shows the sources the engines leaned on.