Committed to the record on June 9: Claude would cite AC companies' own websites under 5 percent of the time, and the web-searching engines would clear 35. Measured on June 11: 2 percent, and 47. The ways the forecast could fail were written down too. That is the standard your category gets measured to.
See what AI sees on your site →
Anyone can explain a number once it exists. A forecast committed to the record before the first run, with its own kill criteria attached, is a different kind of evidence. Here is the one we made for this category, and what the measurement returned.
The prediction was committed to a timestamped record on June 9, 2026, before a single HVAC query had run. The reasoning, also on the record: home services is an editorially thin local trade, so individual AC companies barely exist in Claude's training data. HVAC is the third unrelated category where that collapse held, after CPA firms across three geographies and med spas. Full account in teardown 09.
Why this matters to an AC company deciding whether to believe any vendor's numbers: a measurement practice that publishes dated forecasts with kill criteria cannot quietly move the goalposts when the data lands. You are not being asked to trust a story. You are being shown a prediction that was allowed to fail, and did not.
In the markets we measured, Hawaii hotels hold 11 percent of the web-searching citations and Honolulu real estate firms hold 14, because the portals and booking sites own the answer. Honolulu HVAC is different: across the five measured web surfaces (four web-searching engines plus the Bing organic control), AC companies' own websites carry 40 percent of the citations.
Review directories drew only 7 percent of mentions in this category, and the cohort was built with Angi, HomeAdvisor, and Home Depot excluded by rule, the way the teardown documents. There is no dominant national brand sitting on the answer. When a Honolulu buyer asks AI about AC repair or installation, the citations mostly land on actual AC companies' own sites or on editorial content, which means the citation is genuinely contested. Somebody in your market is going to hold it.
That is the practical difference between this page and our hotels and real estate pages: those categories are fighting a wall. Yours is a race. Cross-market comparison at the cross-category teardown.
The same eighteen questions, the same days, the same twelve firms, across six AI tools: four that search the live web and two that answer from model memory. The share of citations reaching an AC company's own site ran from 66 percent down to zero depending only on which tool the buyer opened.
| AI surface | AC-company-owned share | Third-party share | Mentions |
|---|---|---|---|
| ChatGPT search | 66% | 34% | 935 |
| Gemma (model memory) | 63% | 38% | 432 |
| Gemini grounded | 45% | 53% | 1,674 |
| Google AI Overviews | 45% | 43% | 302 |
| Perplexity | 38% | 59% | 1,477 |
| Claude (model memory) | 2% | 98% | 606 |
| Bing organic (control) | 0% | 91% | 792 |
Two readings worth taking seriously. Bing organic returned an AC company's own site 0 percent of the time across 792 results, so the gap is cohort-wide. Whether ranking there changes what any AI answer engine cites is not something this measurement tested. And Claude's 2 percent against Gemma's 63 on identical questions means even the two model-memory engines disagree completely. A company can be the answer on one surface and absent on the next, which is the reason a single-engine tool cannot tell you where you stand, and the reason we run all seven surfaces: six AI tools plus the Bing organic control.
Home services marketing is crowded with vendors who do the work and then report their own results. We are the other thing: the count that stays honest because we have no work in the race.
The questions are locked by hash before we look, the method is documented in full at /methodology/, and every published number is dated and tied to the run that produced it on the claims record. Your team or your marketing partner executes the punch list. The called shot above is what that discipline looks like when it is real.
Honolulu is where this cohort was measured. The method itself is geography blind. Lock the questions your buyers ask, run the surfaces, count the names. It reads the same from Oahu as it does from Phoenix, Dallas, or Tampa.
Paste your URL and see in seconds what the six AI tools can read from your site. Free, no signup.
Want your market and comp set measured? Book 15 minutes with Lance.