Open data
What AI assistants actually recommend to local homeowners
We asked ChatGPT, Gemini and Perplexity the same 12 questions a homeowner types, in 3 cities across 4 trades, and recorded every answer word for word. That is 12 markets in all. Free to read, no login.
- 2.77 questions the average company is named in, out of 12 n = 151 companies
- 40% are never named at all 61 of 151 companies
- 432 answers recorded word for word read on August 8, 2026
- 6 of 18 companies with 200+ reviews are named in nothing in both readings, most-reviewed of them: 872 reviews
Read on August 8, 2026. One reading per market, not an average across runs, and it says so on every page. Trades covered: Electrical, HVAC, Plumbing, Roofing.
By trade
Every figure carries the number of companies behind it.
| Trade | Companies | Named in, average | Never named |
|---|---|---|---|
| Electrical | 35 | 3.43 of 12 | 13 (37%) |
| HVAC | 48 | 2.6 of 12 | 21 (44%) |
| Plumbing | 29 | 3.31 of 12 | 11 (38%) |
| Roofing | 39 | 2 of 12 | 16 (41%) |
What the assistants read
Domains cited most often, by how many of the 12 markets cited them and how many distinct questions they appeared under.
| Source | Markets | Questions |
|---|---|---|
| angi.com | 12 of 12 | 116 |
| bbb.org | 12 of 12 | 81 |
| yelp.com | 12 of 12 | 79 |
| homeadvisor.com | 12 of 12 | 70 |
| reddit.com | 12 of 12 | 62 |
| facebook.com | 12 of 12 | 47 |
| youtube.com | 12 of 12 | 36 |
| expertise.com | 11 of 12 | 55 |
| thumbtack.com | 10 of 12 | 69 |
| homeguide.com | 10 of 12 | 49 |
Raw citation counts are tracked separately and never mixed with these: one answer can cite the same domain three times, which is one question, not three.
Every market we have read
Each page lists the questions, who was named, and what the answers cited.
Method, and what this is not
- How is AI visibility measured here?
- Each market is a city and a trade. We ask 12 fixed questions a homeowner would actually type, such as "best HVAC company in Chino Hills", to ChatGPT, Gemini and Perplexity, and record every answer word for word along with the pages it cited. No question names a business, so one reading measures every company in that market at once.
- How often is a local home service company named by an AI assistant?
- Across 151 companies in 12 markets read on August 8, 2026, the average company was named in 2.77 of 12 questions. 61 of them (40%) were named in none at all.
- Which websites do AI assistants cite when recommending a local contractor?
- Mostly the contractors' own websites. Of 3801 citations across these readings, 61% pointed at a home service company's own site and only 20% at a directory or review site. Directories still show up in the most markets, which is a different thing: angi.com, bbb.org, yelp.com and homeadvisor.com appear almost everywhere but account for a small share of what was actually read.
- Does having a lot of reviews mean AI assistants recommend you?
- Barely. Across the 78 companies whose review count was visible, the correlation between review count and how many questions a company is named in is 0.118, which is close to none. Measured across both readings, 6 of the 18 companies carrying 200 or more reviews were named in none of the twelve questions either time, and the most reviewed of those carries 872. A strong reputation on review sites and being named by an assistant are close to separate problems.
- How much does AI visibility move on its own?
- A lot, which is why no reading here is presented as a verdict. We read the same 12 markets twice, 1 day apart, changing nothing in between. 55% of the 121 companies present in both readings had a different count the second time, by an average of 1.84 questions out of 12, and the largest single move was 5 questions. Anyone showing you a single confident AI visibility score, including us, is showing you one sample of a moving number.
- Does guidepost need access to my accounts?
- No. It reads public AI answers, the same ones any homeowner sees. There is nothing to install, no permission to grant, and no connection to accounting or field service software.
Answers were collected through each assistant's own public API. Consumer apps are never scraped. Findings that contradict what we expected are published anyway.