Which HVAC Companies ChatGPT Recommends: 94 Profiled

The first honest finding is how much could not be measured at all. "Listing" below means a Google Maps business listing found by brand name within 60 km of the metro centroid and carrying an HVAC or contractor category — not a verified identity. Every gap is recorded with its literal reason and never estimated.

Every listing figure on this page is vendor tool data rather than a Google-reported metric. DataForSEO's business_data_business_listings_search endpoint returned the ratings, review counts, claimed status, categories, addresses and coordinates recorded here on 2026-08-17 [VENDOR — DataForSEO]. That endpoint's is_claimed parameter defaults to true and was left at its default, so an unclaimed listing surfaces here as no confident match. The records it returned carry is_claimed: true on 89 of the 93 matches and omit the field entirely on the other four, which the artifact stores as null rather than false — on those four rows, "claimed" rests on the search filter alone.

Read each row against its own denominator, not against the top row. These are independent checks rather than a funnel, and the column is not meant to shrink monotonically: robots.txt and homepage JSON-LD were attempted for every brand that published a website URL, so those rows sit above the full-body row rather than below it. AI-crawler counts later on this page are denominated on the 60 and 18 robots.txt files that were actually readable — never on 94, and never on 69.

What the repeatedly-named group did share, on 2026-08-17:

1. A matched Google listing. All 69 matched one, and 66 of those records came back flagged is_claimed: true; on the other three the field was absent from the response. That is partly a method artifact: the vendor search ran with the endpoint default is_claimed=true, so an unclaimed listing would appear here as "no confident match", and "listing found" means "a listing came back under a claimed-only filter", not "this business has a Google presence".
2. A high star rating. Of the 67 rated listings, 63 sat at 4.7 or above; the median was 4.9 and only four fell below 4.7 (3.7, 4.0, 4.3, 4.6). The 25-brand comparison group was tighter still: all 23 rated listings at 4.8 or above.
3. A website reachable enough to matter — usually. Sixty-seven of 69 had a website URL on the listing; 57 returned a full 200 response to our single automated fetch.
4. Almost nothing else. Review counts spanned 6 to 35,404. Homepage structured data was present on fewer than half. Forty-nine of the 60 readable robots.txt files named no AI crawler whatsoever.

Two cautions belong with the star ratings before anyone builds a target around them. Rating and review values belong to the single best-matching listing, not to the brand — several of these companies run multiple Google listings, and where more than one qualified we recorded the one with the most reviews. And a listing match is a name-plus-metro-plus-category match, not an identity verification.

The clearest illustration is the row most likely to be wrong. The panel's brand label "Estes Services" was named in 8 of Atlanta's 10 runs; our matcher attached a listing titled Estes Heating & Air Conditioning Inc in Lawrenceville, GA, showing 4.0 stars from 25 reviews and no website URL, which is why nothing else in that row could be measured. That record is also one of the four whose is_claimed field came back absent. Whether that listing belongs to the business the engines meant is exactly what a name match cannot settle. Matching error runs in both directions: a real company can be missed, and a wrong same-named company can be attached. Every matched title, address and coordinate is recorded so individual rows can be rejected, and this one deserves scepticism.

Review profiles clustered hard at the top of the star scale — 63 of 67 rated listings at 4.7 or better — while spreading across three orders of magnitude on volume. Nine listings carried under 100 reviews; 20 carried under 500; 11 carried 10,000 or more. A rating floor and a review-count floor are very different claims, and only the first is visible here.

Websites told a messier story than the "AI reads your site" pitch suggests:

• 57 of 69 homepages returned a full 200 response to our fetch.
• Three returned HTTP 403 and two returned 202 — bot-protection challenges. Those sites exist and answer humans; they refused one automated request from one IP with one user agent.
• One timed out entirely: TE Certified, named in 9 of Atlanta's 10 runs, the joint most-named brand in the whole panel.
• Four returned 200 with a body under 2,000 bytes, including a 0-byte response from Island Breeze, named in 5 of Phoenix's 10 runs.

A company whose homepage our crawler could not read was still among the most-named brands in the dataset. That does not make site health irrelevant — it makes a single automated fetch a weak instrument.

Where the engines actually looked is a question about answers, not companies, and it belongs to the parent report: the citation mix, the domain concentration and the directory behaviour are all published there with the raw ledger. Consumers, meanwhile, are not treating AI output as final. BrightLocal's 2026 Local Consumer Review Survey of 1,002 US adults found the share using AI for local business recommendations climbed from 6% in 2025 to 45% today, and that 97% of AI users sometimes double-check AI recommendations against real reviews — 88% of AI users checking either that a review is legitimate (51%) or the source (37%). An AI mention is a shortlist entry that a homeowner then verifies — which is the strongest practical argument for treating reviews and site quality as one job rather than two competing ones.

The run column is not comparable across metros. Phoenix, Atlanta and Chicago each carried 10 runs because they were the repeat subset; the other nine metros carried 4. A brand in Atlanta simply had more chances to be named than a brand in Orlando.

Inside a single metro the comparison holds, and it is the striking one. All Temp was named in 9 of Chicago's 10 runs with 8 reviews on its best-matching listing. Morris-Jenkins, whose listing carries 35,404 reviews — the largest count in the study set — was named in 3 runs in Charlotte, out of that metro's 4. Both facts are per-listing snapshots from one day, and neither is evidence about the other.

Across groups, the medians do differ: 1,220 reviews for the 69 repeatedly-named brands against 282 for the 25 single-run brands. That gap is reported for description only. The two groups are unmatched on metro, size, business age and category; the comparison group draws from four metros while the main group spans twelve; no significance test was performed, and none should be inferred downstream. A metro-composition artifact explains the gap as easily as anything about the companies. How the panel classified single-market against multi-market brands is a property of the answers, and it stays in the parent report.

1. A Google listing exists and is findable by name in the right metro. True for 93 of 94, with is_claimed: true returned on 89 of those 93 and absent on four — under the claimed-only search caveat above.
2. The listing carries a star rating at or above 4.7. True for 63 of 67 rated listings in the repeatedly-named group, and 23 of 23 in the comparison group.
3. A website URL is published on the listing. True for 67 of 69 and 24 of 25. Two of the repeatedly-named brands had none, and for them nothing else could be measured.
4. The homepage answers a request. True for 66 of the 67 attempted in the repeatedly-named group; only one failed to resolve at all.
5. The robots.txt file exists and is fetchable. True for 60 of 69. Nine were unreadable: two 403s, two 202s, two 200s returning something other than a plain-text robots file, two brands with no website URL to check, and one non-response.
6. The robots.txt does not block search-side AI crawlers. No file in either group disallowed OAI-SearchBot, PerplexityBot or Claude-SearchBot outright.
7. Homepage structured data. Optional in practice: 28 of 57 fully-fetchable homepages carried LocalBusiness or HVACBusiness JSON-LD, and 29 did not.

Item 6 is where this profile earns its keep, because the robots.txt findings contradict a common sales line. Across the 60 readable files in the repeatedly-named group, only 11 named any of the nine AI crawlers we checked; in the comparison group, 1 of 18 did. Silence, not policy, is the norm.

Counts are stated directives among the 60 readable robots.txt files in the repeatedly-named group, 2026-08-17. The chart above collapses full and partial disallows into one "blocked" segment; the table separates them. A directive records a policy at one moment; it does not prove any crawler obeyed it, was blocked, or ever visited.

Three companies published User-agent: GPTBot / Disallow: / and were named anyway — Hero Air (4 Chicago runs), Horne Heating and Air Conditioning (4 Charlotte runs) and Super Service Cooling, Heating & Plumbing (2 Las Vegas runs). All three also disallowed ClaudeBot, CCBot and Google-Extended. None of the three disallowed OAI-SearchBot.

That split is documented behaviour, not an accident. OpenAI publishes GPTBot as the crawler for content "that may be used in training our generative AI foundation models" and OAI-SearchBot as the one "used to surface websites in search results in ChatGPT's search features", and states plainly: "Each setting is independent of the others". Perplexity draws the same line — PerplexityBot "is not used to crawl content for AI foundation models" — and Anthropic separates ClaudeBot from Claude-SearchBot, which "navigates the web to improve search result quality for users". Google-Extended is a training-and-grounding token only: Google states it "does not impact a site's inclusion in Google Search nor is it used as a ranking signal in Google Search".

The mechanism that keeps those blocks narrow is group precedence. Two of the three files pair their AI-crawler blocks with a wide-open wildcard group — User-agent: *, Allow: /. That wildcard does not soften the GPTBot block: Google's robots.txt documentation states that "Only one group is valid for a particular crawler", chosen by most specific user-agent match, and that "User agent specific groups and global groups (*) are not combined". GPTBot obeys its own group and ignores the wildcard; OAI-SearchBot, unnamed, falls back to the wildcard and is allowed. Blocking training while permitting search is one file away — and worth testing rather than assuming, which is what our free crawler-reachability check for HVAC sites does.

Two of those same files carried a Content-Signal: line reading search=yes,ai-train=no,use=reference. The first two tokens map onto Cloudflare's published Content Signals vocabulary, which defines exactly three signals — search, ai-input and ai-train; use=reference is not one of them. Cloudflare's own page warns that "Not all automated systems honor robots.txt files, and some may ignore these Content-Signal directives", and Will Allen, who announced the policy on 24 September 2025, framed the underlying gap this way: "This allows you to specify which crawlers are allowed and what parts of your site they can access. It does not, however, let them know what they are able to do with your content after accessing it." A stated preference is not an enforcement mechanism, and neither is a robots.txt line.

Item 7 deserves the same deflation. Nine repeatedly-named companies whose homepages loaded completely carried no JSON-LD at all — including PV Heating (6 Atlanta runs), Benefit Air (5 Phoenix runs), Hero Air and Morris-Jenkins. HVACBusiness remains the correct schema.org type for an HVAC contractor, and 23 homepages in the repeatedly-named group carried it. Ten carried LocalBusiness, five of those alongside HVACBusiness — which is why the union is 28 homepages rather than 33, and why only five used LocalBusiness on its own. But Google's own guidance is unambiguous that "Structured data isn't required for generative AI search, and there's no special schema.org markup you need to add", while remaining worth keeping for rich-result eligibility. Our observation matches Google's statement rather than the schema-will-get-you-cited pitch. If you want the version that separates markup that still earns something from markup that does not, the structured-data types worth an HVAC contractor's time covers it.

The wider benchmark points the same way. SISTRIX found weekly source churn of 5% in AI Overviews, 56% in Google AI Mode and 74% in ChatGPT Search [VENDOR — SISTRIX] across 82,619 qualified prompts over 17 weeks. Read that 74% as the top of the range, not a US reading: its ChatGPT breakdown is Germany 74%, UK 60%, France 42%, with no US ChatGPT figure published — the US number, 54%, is for AI Mode. Every attribute on this page was recorded on one day, in one pass: ratings move, review counts move, and robots.txt files get rewritten by whoever ships the next site template.

The practical consequence is a measurement habit, not a checklist sprint: fix your prompts, fix your engines, log every run verbatim, repeat on a schedule, and profile the companies that appear beside you rather than screenshotting the one answer that flattered you. That standing measurement loop is what we run for HVAC contractors month over month, with the raw runs attached and no promise about which names an engine will produce next. For the engine-specific mechanics behind being findable at all, our ChatGPT-side breakdown for HVAC owners covers the pathway in depth. Recurring profiling of this kind carries no separate AI-visibility price: it sits inside the $2,500/month SEO retainer, which includes the GEO and AI-search work rather than billing it as an add-on.