HVAC AI Visibility Report: 66 AI Answers, 12 US Metros

The two prompts were sent exactly as written, with {metro} substituted and nothing else added:

1. P1 — "Who is the best HVAC company in {metro}?"
2. P2 — "My AC stopped working in {metro}. Which company should I call?"

Each run is a fresh single-turn request: no system prompt, no chat history, no supplied search results. Fifteen fields were recorded per run, including the DataForSEO task id, the full verbatim response text, the ordered list of every business named, every citation URL returned, whether the answer pointed the reader to a directory or registry, and any company the answer cautioned against.

Business names were read off the verbatim response text by hand, not by a classifier, and brand variants were merged only through an explicit alias list — never by fuzzy matching. That is why this dataset counts "Legacy Air Conditioning and Heating" and "Legacy Heating, Cooling, Plumbing & Electrical" as two brands: same-entity identity could not be verified in session, so we did not assume it.

What went wrong, and where the design is weak

One design gap belongs on the record before anyone cites this. The panel's build note recorded location as "not settable on this endpoint." That is wrong. DataForSEO's documentation for the ChatGPT llm_responses endpoint lists both web_search_country_iso_code and web_search_city, and the Perplexity endpoint exposes web_search_country_iso_code for Sonar models. Those fields existed and we left them unset, so the honest statement is "geolocation was available and unused," not "geolocation was impossible." A homeowner physically in Phoenix may see different answers than an unlocated API call that merely says "Phoenix." Version 2 sets those fields.

Sampling parameters were also left at endpoint defaults, which are not zero: DataForSEO documents a default temperature of 0.94 and top_p of 0.92 on the ChatGPT endpoint, and 0.77 / 0.9 on Perplexity's. One qualifier belongs with that ChatGPT figure — the same documentation adds that temperature is "not supported in reasoning models," and our ChatGPT runs used a reasoning model, so the 0.94 default cannot be assumed to have applied to them. Non-zero sampling is one obvious contributor to the run-to-run movement reported below, and this design cannot separate that contribution from index churn.

The full limitation set:

• Single day. All 66 runs sit inside one ~5.5-minute window. Nothing here describes a trend or a stable state.
• n=1 in 42 of 48 cells. Each main-grid number rests on one observation. The repeat subset shows single observations are unstable, so read every main-grid cell as one draw, not a measurement.
• n=4 in 6 cells. Four observations demonstrate that variance exists; they cannot estimate it.
• Not a ranking-factor study. No attribute of any recommended company was measured, so no causal or correlational claim about what earns an AI recommendation is supported.
• Two engines only. Google AI Overviews, Gemini, Claude and Microsoft Copilot were not tested, though the endpoint exposes claude and gemini families. Scope stayed at two engines so the repeat subset could run properly.
• APIs, not apps. gpt-5.5 via API is not the ChatGPT app; sonar-pro via API is not perplexity.ai in a browser. App-level personalization, memory, ad units and UI ranking are absent.
• No dated Perplexity build. The endpoint exposes no dated build for Sonar models, so the version behind "sonar-pro" on 2026-08-17 cannot be pinned.
• Analyst-assigned labels. Business-type and citation-class buckets are judgment calls from in-session evidence. No filings, franchise disclosure documents or ownership registries were checked; another analyst would bucket some domains differently and move the percentages.
• Citation counts are not volume-comparable across engines. ChatGPT repeats a URL across claims, so its per-run lists were de-duplicated; Perplexity's count is the source list it returned, and that list came back at 20 URLs in 30 of its 33 runs and 19 in the other three — a ceiling on what the engine reports, not a count of what it consulted.
• Nothing the models asserted was verified. Star ratings, review counts and awards inside the answers were recorded as model output only. Several are likely wrong.
• Absence proves nothing. A company unnamed in these 66 runs may still be recommended on another day, phrasing, or engine.

The prompts, model ids, parameters and task ids are all recorded, so the calls are re-issuable. The answers are not reproducible: these endpoints are non-deterministic and the underlying web index changes daily.

Read the run column first in the table below, which is counted from the panel's raw run ledger. Phoenix, Atlanta and Chicago carry 10 runs each because they were the repeat-subset metros, so their brands had more chances to be named. Cross-metro comparison of these counts is not valid.

The right-hand column follows one fixed rule and breaks no ties: it lists every brand named in 3 or more of that metro's runs, and where no brand cleared 3 it lists every brand named twice instead. Brands named exactly twice are otherwise omitted — between one and six of them per metro — because picking a subset of tied names would be an editorial choice this dataset cannot justify.

These are the panel's canonical merged labels, not the companies' legal names, and we verified nothing about any of them. Moncrief and TE Certified were each named in 9 of Atlanta's 10 runs, which tells you the engines were consistent about those two names inside one five-minute window on one day. It does not tell you either company did anything to earn it.

Concentration matters more than any single name: 125 of the 194 distinct brands were named in exactly one run. Roughly two-thirds of the companies these engines produced appeared once and never again. Position needs the same care — ordered lists reflect the order names appear in the answer text, and the models rarely state a strict ranking, so position 1 means "named first," not "ranked first."

Two of the 66 runs did the opposite of recommending: the answer text cautioned the reader against a specific named company. We are not reprinting those claims, because the panel records them as unverified model output and no rating, review count or award in any answer was checked against Yelp, Angi, BBB or the company itself. The behaviour is the finding — these engines will occasionally warn a homeowner away from a named contractor, on evidence nobody has audited.

A separate measurement pass on 2026-08-17 did inspect 94 of those brands — Google listing, homepage reachability, robots.txt and homepage JSON-LD — and its findings are published as an observational profile of what the AI-named HVAC companies actually look like. That file measures companies, not answers, and it establishes no cause either.

What the panel can describe is the shape of the answers and the evidence surfaces the engines reached for.

The named companies were overwhelmingly small and local-looking. Of 194 distinct brands, 190 were labelled apparent single-market contractors, 3 were multi-market brands (Cool Today, Goettl, One Hour Heating and Air Conditioning) and 1 was a retailer. By mentions: 323 single-market, 10 multi-market, 1 retailer. A brand earned the multi-market label only where a cited page on its own domain used a per-metro location path, proving multi-metro operation. "Apparent" means exactly that — brands labelled single-market may be roll-up or franchise owned, since no ownership registry was checked.

Third-party review platforms dominated the citations. The 66 runs returned 767 citation URLs, 465 of them unique, spanning 200 unique domains.

The two engines behaved nothing alike as citers. ChatGPT returned 110 citation URLs across its 33 runs — 84 unique, from just 22 domains, a mean of 3.33 per run, with a per-run range of 2 to 6. Perplexity returned 657 across its 33 runs — 381 unique, from 191 domains — but its per-run count barely moved: 20 URLs in 30 of the 33 runs and 19 in the other three, a floor of 19 and a ceiling of 20. A list that comes back at exactly 20 URLs in 30 of 33 runs is a capped list, not a measure of how widely the engine searched, so the 110-against-657 gap is not evidence of citation breadth and should not be quoted as if it were. The contrast that survives the cap is domain diversity: 22 domains for ChatGPT against 191 for Perplexity.

Reddit as the most-cited domain is the finding most worth dating, because it inverts the closest prior benchmark. BrightLocal's November 2024 study of 800 manual ChatGPT local searches reported that "Forums like Reddit and Quora did not appear as sources throughout this study," alongside a mix of 58% business websites, 27% business mentions and 15% directories. Twenty-one months later, on a different engine pair and a different unit of measurement, reddit.com contributed 66 citation URLs and appeared in 44 of our 66 runs. BrightLocal classified source types across manual searches; we counted API-returned citation URLs — so treat these as two dated snapshots pointing opposite ways, not a trend line. Expertise.com appears in both: 18% of BrightLocal's directory sources then, third most-cited domain here now.

Contractor-owned "best HVAC company in {city}" pages did get cited — rarely. Thirty citations across 18 distinct URLs on contractors' own domains matched the best/top/compare/ranking path rule, including four pages from one San Antonio contractor and three from one Orlando contractor. That class is 3.91% of all citations. Whether publishing such a page changes whether an engine names you is not something this panel can test, and the small share should temper confidence either way.

Directory pointers came only from ChatGPT, and they supplemented named contractors rather than replacing them. Seven of the 66 runs (10.61%) told the reader to go to a directory, review platform or state registrar as an action — Yelp, Angi, BBB, Consumers' Checkbook. All seven were ChatGPT runs: 7 of 33 ChatGPT runs (21.21%) against 0 of 33 Perplexity runs, a split the panel-wide 10.61% hides. In zero of those runs did the directory stand in for a named company. The pattern is a shortlist plus a verification step, as in run r005 (ChatGPT, Phoenix, P1): "There isn't one objectively 'best' HVAC company in Phoenix, but my short answer would be: start with North Valley Mechanical HVAC & Plumbing or Larson Air Conditioning, then get at least 2–3 quotes before authorizing major work."

Overlap here is the Jaccard index: brands both engines named, divided by all distinct brands either named.

A cell at 0.0 means the two engines produced entirely separate lists of HVAC companies for the identical question, minutes apart.

San Antonio's P2 cell is the high-water mark at 0.6667 — four shared brands out of six distinct. Las Vegas produced zero overlap on both prompts. With one run per engine per cell, no individual number here is a measurement; the distribution is the point.

The same engine, asked four times

The repeat subset re-ran prompt P1 four times per engine in three metros, all inside roughly six minutes. Only one of the six cells kept the same first-named brand across all four runs.

The engines differ in kind, not degree. ChatGPT's mean pairwise overlap across the repeat cells was 0.2519; Perplexity's was 0.7980. Phoenix on ChatGPT is the extreme: 17 distinct brands surfaced across four identical requests, 15 of them in only one of the four, and one brand present in all four. Perplexity behaved like a narrow, stable shortlist that reshuffled its order; ChatGPT behaved like a wide, shifting pool.

Rewording the question changes the answer too

Prompt framing moved the output as much as engine choice did. Across the 48 main-grid runs, P1 ("Who is the best HVAC company in {metro}?") produced 131 brand mentions and 117 distinct brands, a mean of 5.46 per run, pointing to a directory in 4 runs. P2 ("My AC stopped working in {metro}. Which company should I call?") produced 110 mentions and 95 distinct brands, a mean of 4.58 per run, pointing to a directory in 2 runs. Comparing P1 against P2 within the same engine and metro, mean overlap was 0.1733, with 8 of 24 pairs sharing no brand.

Multi-market brands drew more mentions from the emergency-framed prompt: 2 mentions under P1 against 5 under P2. That is five mentions in a 334-mention dataset — an observation to test properly, not a finding to act on.

None of this is unique to HVAC. SISTRIX's Johannes Beus, whose team tracked 82,619 prompts across 17 weeks and six countries, measured weekly source replacement of 5% in AI Overviews, 56% in Google AI Mode and up to 74% in ChatGPT Search, and reached the conclusion this panel independently walked into: "a single citation placement is not a reproducible result but merely a snapshot." His companion line belongs on every AI-visibility proposal: "GEO is a continuous process, not a one-off optimisation with permanent results."

One quirk shapes how these answers read. Counting every Perplexity response whose recorded text contains the exact phrase "you provided", 8 of the 33 Perplexity runs credited search results or sources the caller never supplied: r002, r006, r008, r020, r024, r050, r056 and r062. Run r006 (Phoenix, P1) contains the phrase "Based on the results you provided, the strongest contenders are". Every run was a fresh single-turn request; nothing was provided. That is the model's own framing of its retrieval, recorded verbatim and uncorrected — and it is why the run ids are published, so the count can be re-run against the ledger rather than taken on trust.

1. Stop chasing a single answer. With 8 of 24 matched cells at zero overlap and only 1 of 6 repeat cells holding its first-named brand, any screenshot of "ChatGPT recommends us" — or of a competitor — is one draw from a distribution.
2. Measure with a panel, not a query. Fix the prompts, fix the engines, log every run, repeat on a schedule. Our 66 runs cost $3.044076; the scarce resource is the discipline of recording verbatim output, not the API budget. And set the geolocation fields your endpoint exposes — the gap we published above is the one to avoid.
3. Fix eligibility before anything clever. Google's own documentation on AI features is unambiguous about the floor: "To be eligible to be shown as a supporting link in AI Overviews or AI Mode, a page must be indexed and eligible to be shown in Google Search with a snippet, fulfilling the Search technical requirements. There are no additional technical requirements." The same page notes AI Overviews "are only shown when our systems determine that it is additive to classic Search." Start by confirming crawlers reach you — test whether these engines' crawlers can even fetch your site in about a minute.
4. Take third-party review surfaces seriously, on evidence. Third-party directory and review platforms carried 59.58% of the 767 citations here, and consumeraffairs.com appeared in 48 of 66 runs. That describes where these engines were looking on one day, not proof that a profile edit changes an answer.
5. Expect the answer to move, and budget for it. Between SISTRIX's up-to-74% weekly churn in ChatGPT Search and the within-six-minute movement recorded here, AI visibility is a monitoring commitment, not a project with an end date.
6. Refuse guarantees, including ours. Nobody can promise you a slot in an AI answer, and this dataset is why: one engine, one question, four times in six minutes, 17 different companies in a single metro.

That last point is the basis of how we work — measuring one contractor's AI answers month over month against a fixed prompt panel, reporting what moved and what did not, with the raw runs attached. That panel is not sold separately: it sits inside the $2,500/month retainer on our pricing page, because a measurement you can switch off is a measurement nobody runs. For the engine-specific mechanics of being findable at all, the companion piece is the tactics guide for ChatGPT specifically.