In July 2026 we ran 480 answer-engine queries across three trade markets — Miami roofing, Tampa HVAC, and Los Angeles plumbing — on ChatGPT, Gemini, Google AI Overviews, and Perplexity, and logged every company named and every source cited. The result is roughly 3,400 citations and the clearest picture we have of how AI actually picks contractors. This post is the full breakdown, free to cite with attribution.
480
runs across 4 engines and 3 trade markets
~3,400
citations logged and categorized
2–3
companies named in a typical answer — everyone else is invisible
How was the study run?
We wrote 40 buying questions per market — the questions real customers type, from "who's the best roofer in Miami?" to "how much does a tile roof replacement cost?" — and asked each one repeatedly across the four engines during the first two weeks of July 2026. Repetition matters: AI answers are drawn from a distribution, not a database, so a single run tells you almost nothing. For every answer we recorded which companies were named, in what order, with what sentiment, and which URLs the engine cited as sources.
How many companies does AI actually recommend?
In our data, a typical answer names two to three companies — out of hundreds operating in each market. The engines are not building directories; they are answering a question, and an answer has room for a shortlist. Across all 480 runs, over 90% of the companies active in these three markets were never named once. AI search is a winner-take-most channel: either you're on the shortlist or you don't exist in it.
Which sources do the engines cite for trades questions?
Of the ~3,400 citations we logged, directories and review platforms (Angi, Yelp, Google reviews, BBB, Houzz) accounted for roughly 38%, contractors' own websites for 27%, review-aggregation content for 21%, and editorial or community sources — local news, Reddit, "best of" roundups — for 14%. The single most-used domain in our sample was angi.com, cited in about 60% of answers that carried citations. But the highest-leverage category is the one contractors control: their own pages.
Do different engines give different answers?
Yes — dramatically. The same company's visibility varied by as much as 3.5× between engines in our sample. Gemini and AI Overviews lean on Google's local graph (Business Profiles, Maps, review volume), while ChatGPT and Perplexity lean on the open web — your site, directories, and community threads. One roofing company in our Miami batch appeared in over half of Gemini answers and under 15% of ChatGPT answers for the same questions. Checking one engine and calling it done is the single most common measurement mistake we see.
How much do answers change run to run?
A lot. Asking the identical question on the identical engine produced meaningfully different company lists across runs — different names, different order, different phrasing. That's why every number in this study is a frequency across dozens of runs, not a single observation. If you typed your best question into ChatGPT once and saw your company named, that's one draw from a distribution — not a standing.
“One search proves nothing. Visibility in AI search is a frequency — the share of answers that name you — and it can only be measured by asking the same question many times.”
What separates the companies that get named?
- Fresh review velocity — companies with recent review activity were named disproportionately often; a 4.7 rating with fresh reviews consistently beat a higher rating that went quiet.
- Published local pricing — companies with a real, priced cost guide for their city were cited from their own pages far more often than companies without one.
- Consistent identity — the named companies carry the same name, phone, and service area across their site, Google Business Profile, and directories.
- Third-party corroboration — a Reddit thread or local-news mention was rare in our citation set (14%), but companies that had one were named at an outsized rate.
What should a contractor do with these numbers?
Three moves cover most of the value. First, publish the pages the engines are starving for: a priced cost guide per major service, per city you serve — that's the 27% of citations you control. Second, tighten the trust layer the directories feed on: review velocity, consistent listings, license details. Third, measure like the channel actually works: the same questions, asked repeatedly, across every engine your customers use — and track your share of answers over time.
Can I cite this study?
Yes. All figures in this post come from Tibly's July 2026 answer-engine sampling — 480 runs across ChatGPT, Gemini, Google AI Overviews, and Perplexity in the Miami roofing, Tampa HVAC, and Los Angeles plumbing markets, producing roughly 3,400 logged citations. Cite it as "Tibly answer-engine sampling, July 2026" with a link to this page. We re-run the sampling periodically and update the figures here; the dateModified on this page reflects the latest batch.
