How ChatGPT decides which electrician to recommend

Ask an assistant for an electrician in your city and you get two or three names and a sentence about each. Not ten links. Not a page you can scroll. Three names, and if you are not among them there is no second page to be found on.

That is a different competition from ranking, and it is won differently.

The model is not ranking you, it is quoting whatever it can defend

This is the part that takes longest to accept, so it is worth stating plainly: an assistant is not scoring businesses and returning the best one. It is assembling an answer it can state without qualifying, from sources it can reconcile.

Which means the question is not "am I the best electrician in this city?" It is "can a machine describe my business in one confident sentence without contradicting itself?"

For most independent trades the answer is no, and the reason is boring.

Contradictory facts are the whole problem

Take a typical electrical business. Its website says it covers the whole valley. Its Google profile lists a five-kilometre service radius. A directory entry from four years ago has an old mobile number. A lead-generation site lists it under "electrical contractor" with a business name that is slightly different — a trading name, or the owner's name, or a spelling nobody corrected.

None of these is a scandal. Every one of them is a reason for a model to hedge.

When it hedges, it names somebody else. Not because the competitor is better, but because the competitor's facts agree with each other, and agreement is the only thing the model can convert into a confident sentence.

The highest-value work here is therefore not clever and not creative. It is picking one accurate description of your business and making every source on the internet say the same thing: the same trading name, the same number, the same service area, the same list of what you do.

What "extractable" actually means on a trades site

The second problem is that most trade websites are written to reassure, not to inform, and reassurance does not extract.

Compare these.

Buried: "With over fifteen years of experience serving the community, our fully qualified team delivers a professional service you can rely on, with customer satisfaction our number one priority."

Answer-first: "We are a two-electrician business covering Kathmandu inside the Ring Road, with a licensed electrical contractor's registration. We do domestic rewiring, fault finding, inverter and solar installation, and same-day emergency callouts until 8pm. Callout is charged at a flat rate, quoted before we start."

The first is worth nothing to a model: it contains no fact that can be lifted into an answer. The second can be quoted almost as it stands, and it also happens to answer the questions a customer was going to ask on the phone.

Being specific is not a stylistic preference here. It is the mechanism.

Reviews are evidence, not decoration

A review saying "great service, highly recommended" is invisible to this process. A review saying "came out the same evening to find a fault two other electricians had missed, and charged the quoted callout rate" is a citable fact about what you do.

You cannot script reviews and should not try. You can ask better: when the job is done, ask the customer to mention what the work actually was. "If the same-day callout was what you needed, mentioning that helps other people find us" gets you specifics, and specifics are what a model has something to say about.

Structured data, briefly

Mark up the site so the facts are machine-readable rather than inferred from prose — the business type, the area served, the services, the hours, the contact details. LocalBusiness with the appropriate subtype, validated rather than assumed to work.

This is not a ranking trick and it will not rescue a site whose facts contradict each other. It removes ambiguity from facts that are already consistent, which is a real but narrow job.

What you cannot control, and should not be sold

Nobody can guarantee an AI citation. There is no submission form, no dashboard, no ranking report, and no relationship to buy. Any offer that implies otherwise is selling access that does not exist.

What is controllable is whether the evidence exists, whether it agrees with itself, and whether it is written in a form something can quote. That is the work, and it is measurable only in the crude way described below.

How to measure it

Pick twenty queries a customer would realistically type. Not keywords — sentences. "Who should I call for rewiring in [your area]", "best electrician for solar installation near me", "emergency electrician [your city] evening".

Run them through ChatGPT, Perplexity, Gemini and Google's AI Overviews. Log who gets named, and what those businesses have in common. Repeat monthly.

The first run is usually uncomfortable and always instructive. It tells you who the model considers your competition, which is frequently not who you thought.

If you want that run properly, with the entity consistency work behind it, that is what our GEO and AEO work covers.