What ChatGPT actually reads when it recommends an HVAC contractor (and how to get into the dataset)

A practitioner posted online: "I asked ChatGPT to recommend a roofer in Austin. It named 3 companies. None of them were paying for ads. None of them had the most Google reviews. None of them were number one on Google Maps." That is not a glitch. It is how a meaningful share of homeowners will find a residential HVAC contractor in 2026 and 2027.

I write the content side of HVAC Know It All in 2026 and I run AI-visibility tracking through Teal Maker across the four major assistants for Full Stack HVAC's own surfaces. The pattern that comes back, week after week, is the same: the contractor or brand that wins the local-service citation is the contractor or brand whose name appears in the source corpus the model already trusts.

The discovery shift is already here

Industry surveys from late 2025 and early 2026 (SparkToro, Profound, Backlinko) put the share of consumers using AI tools for local-service discovery in the 40-50% range, up from single digits twelve months earlier.¹ For a residential HVAC shop, this means you can be number one on Google Maps with 500 reviews at 4.9 stars and still be invisible to a meaningful fraction of the homeowners searching for a contractor this week.

What ChatGPT actually reads

Google Maps ranking is driven by proximity, Google Business Profile completeness, review velocity, and link signal. LLM recommendations are driven by what is in the model's training data and what gets pulled in by its retrieval layer. Concretely, when an assistant names contractors, it is drawing from:

• Reddit threads where real homeowners and tradespeople name specific companies. SparkToro's 2025 analyses estimate Reddit at roughly 40% of citation volume across the major LLMs.²
• Local "best of" listicles on city blogs, neighborhood publications, and home-services aggregators. These rank in Google for the head terms the LLMs also indexed.
• Industry publications and trade content: ACHR News, Contracting Business, RSES Journal, HVAC School, HVAC Know It All. These are slow to update but high-trust in the corpus.
• The text body of Google reviews, not the star count. A review that names the technician, the equipment, and the outcome is extractable; a bare 5-star tap is not.

What the assistant is not weighing: paid ads, base Maps rank, or star average in isolation.

Full Stack HVAC tracks brand-visibility scores across ChatGPT, Claude, Perplexity, and Google AI Overviews for the contractors and vendors in the catalog. The single most consistent finding in twelve months of tracking is that visibility for a given brand can swing from near zero to high coverage across model versions, sometimes within a quarter. The contractors who stay visible across model swaps are the ones with broad, named third-party coverage rather than concentrated mentions on a single platform.

Entity clarity is the input the corpus rewards

Three behaviors separate the contractors who get cited from the ones who do not.

1. Consistent brand description across every property. The Google Business Profile, the website About page, social bios, every directory listing. The model's training process and retrieval layer reward a single, unambiguous identity over a scattered one. If three sources call the company three different things, none of the three reliably consolidates into a recommendation.
2. Named co-occurrence in independent third-party content. The shop name appears alongside the relevant service term ("ductless retrofit in Lehigh Valley", "heat-pump installation in Halifax", "Manual J in Phoenix") on pages the model has indexed. A handful of high-trust mentions outweighs a hundred self-published pages.
3. Answer-first content on the site itself. Short, direct answers to the questions homeowners actually ask, sitting near the top of the page. Forty to sixty words for the answer, with the supporting detail underneath. LLMs extract the lead; they do not read the marketing-copy preamble.

What schema markup does and does not do

There is a recurring claim in the SEO press that LocalBusiness schema improves AI-search visibility. The evidence I see in the tracking data does not support that.

Schema does help Google's knowledge graph disambiguate the entity: it is how Google understands that "Smith Heating" is an HVAC contractor in Denver and not a metalworking shop in Portland. That disambiguation is real value, and the schema cost is essentially zero, so ship it. LLMs read the visible text on the page, not the JSON-LD block underneath. Schema is hygiene. The growth lever is the content the assistant can quote.

The HVAC playbook for the next ninety days

1. Get named in two to three local "best of" listicles. Reach out to local home-services bloggers, neighborhood newsletters, and city-specific business directories that publish "best HVAC in [city]" pages. These pages are what the assistants quote when a homeowner asks for a contractor in your service area.
2. Publish three or four answer-first posts on your own site. Pick the questions you actually answer on the phone ("what does a new ductless mini-split cost in [city]?", "should I repair or replace a 16-year-old gas furnace?", "what does a Manual J cost and why does it matter?"). Lead with a forty-to-sixty word direct answer. Add the substantive detail beneath.
3. Normalize the brand description. One sentence describing the shop, consistent across the Google Business Profile, the website, the Facebook page, every directory. The model reads consistency as identity.
4. Open the robots.txt to AI bots you actually want. Allow OAI-SearchBot, ClaudeBot, PerplexityBot, and Google-Extended. Block them, and the assistant cannot retrieve your content even if it would otherwise quote you.
5. Show up in the communities the audience actually uses. The local subreddit, the neighborhood Facebook group, the city's Nextdoor. Helpful, sourced answers under a real name and business signature generate the named-entity mentions LLMs trust.

What not to chase

Do not optimize for a single AI platform. Google AI Overviews citation behavior in particular is unstable from week to week; tracked panels at Profound and similar tools consistently describe Overviews citations as inconsistent.³ Build the third-party footprint that several assistants can find rather than tuning for one.

Traditional local SEO is not going away. Google Maps still drives the largest single share of residential HVAC calls. AI search is an additional channel running alongside Maps and organic, and the contractor who shows up in both is the one whose schedule fills faster.

For the Google Business Profile work that pairs with this (the part Google Maps still ranks on), see the GBP completeness analysis. For the content infrastructure that produces the third-party mentions the LLMs read, see the 6-month content cliff post.