Methodology and data

The questions

Two per city and trade, with no brand names in them. The first drives every published number; the second is a robustness check. Both are frozen for the life of this edition — a survey whose prompts drift between cells is measuring drift.

• Who is the best plumber in {city}?
I need an emergency plumber in {city}, who should I call?
• Who is the best HVAC company in {city}?
My AC broke in {city}, who is the best company to fix it?
• Who is the best electrician in {city}?
I need a licensed electrician in {city}, who do you recommend?
• Who is the best dentist in {city}?
I need a dentist in {city} taking new patients, who do you recommend?

The engines

chatgpt, claude-sonnet, gemini-flash, each called through its provider API with live web search enabled, 5 times per question. That is 15 answers behind every figure we publish.

These are the providers’ APIs with live search enabled, rather than their consumer apps. That is the closest reproducible instrument available, and reproducibility is what a survey needs.

The exact model version behind every answer is in the model_version column of the dataset, so two editions can be compared honestly.

Controls

• Fresh session for every run. No signed-in account, no memory, no prior turn.
• No location set on the request. Where an engine infers one, it infers it from our servers, which are in Northern Virginia — not Los Angeles.
• All runs collected inside one calendar week.
• The full response stored verbatim for every run, not just the names we parsed.
• Sampling left at each provider’s own default, except Gemini, which defaults to deterministic output and was explicitly set to sample so its runs vary like the others.

Limitations

• We do not show that being named produces revenue. We measured recommendation, not its value.
• Company names are matched by exact text. Where an engine writes a business’s name differently from our records, the match fails and the coverage figure is too low rather than too high.
• Provider APIs are not identical to the consumer apps. They are the closest reproducible instrument available, and reproducibility is what a survey needs.
• 90 cities is not all of Los Angeles, and 4 trades are not all local services.

What we publish, and what we do not

We name the companies AI assistants recommend — those appearing in at least 8 of 15 runs. We never publish a list of companies they ignore. Absence is visible by omission and any owner can check it in twenty seconds, but a public list of invisible businesses would be a shaming exercise rather than a scoreboard.