GregM/ai-assistant-hvac-picks

AI assistant picks for HVAC, 2,383 cities

When a homeowner asks an AI assistant "My furnace won't turn on. Who can fix it this week in Plano?", which company does it name, which does it pass over, and why?

This dataset records one AI assistant's answers to six heating and cooling questions in each of 2,383 cities (2,068 US, 88 Canada, 227 UK), September 29 to October 1, 2026: 34,907 picks with the reason given and the page it read, 22,056 companies it considered and passed over with the reason, and 21,167 companies with their result.

It is published by Picked by Agents, which tracks who AI assistants recommend for local purchases. Every row links to its public page, for example HVAC city tests and the Picks Index.

Files

File	One row per	Columns
data/picks.csv	company picked for a question	city, state, country, date, question, rank (1-3), company, domain, why, source_url, page_url
data/passed_over.csv	company considered and not picked	city, state, country, date, question, company, domain, why_not, page_url
data/companies.csv	company	company, domain, city, state, country, date, verdict, times_picked, times_first, page_url
data/city_tests.jsonl	city	the full record: engine, summary, and each question with its picks and passed-over companies

verdict is picked_first (ranked first at least once), picked, passed (considered, never picked) or unmentioned (a local HVAC company found on Maps that the assistant never named).

from datasets import load_dataset
picks = load_dataset("GregM/ai-assistant-hvac-picks", "picks", split="train")

How it was made

For each city an LLM agent with web search played the assistant. It ran one Google Maps search and one web search for the city, as an assistant would, then answered the six questions (furnace repair this week, best company, new system quotes and cost, AC out today, book a tune-up online, upfront pricing; UK versions ask about boilers). It was told to read the candidate companies' own sites before recommending, to rank up to three companies with a one-sentence reason, and to cite the page each reason came from. The engine field records the model where the run noted it (mostly Claude Sonnet with web search).

Limits, read before citing

• One run per city, one month. A city's result is a single answer set, not a rate. Assistants vary run to run.
• One kind of assistant. This is an LLM agent with web search, not the ChatGPT, Gemini or Siri apps. Our monthly Picks Index asks several consumer assistants the same questions.
• The method shapes the sources. The agent was told to read companies' own sites, so source_url is almost always the company's site. Don't read that as how consumer assistants pick sources.
• Reasons are the assistant's words, quoted as given, not verified claims about the companies.
• Google review counts and ratings are not included.

License and citation

CC BY 4.0. Use it for anything, with credit and a link:

Picked by Agents, "AI assistant picks for HVAC, 2,383 cities" (October 2026), https://pickedbyagents.com/hvac

A business that wants its record corrected can ask through pickedbyagents.com.