If a heating and cooling company asks an AI assistant, “What is the best HVAC company?”, it may get a very different answer than it would from “Who should I call for a no-heat emergency in East Tennessee this weekend?” or “Which local HVAC company is honest about when I need a new system instead of a repair?” Those are not slight wording changes. They are different customer decisions.
That is the heart of a careful briefing published on October 29, 2026 by Ian Ash at AEO Updates. Before anyone claims a brand is winning in AI search, they should ask which questions are being used to keep score. This brief explains why that point matters if you want to be recommended when a real customer asks for help.
Why the question set matters more than the dashboard
AI visibility tools send a chosen list of prompts to assistants such as ChatGPT and Gemini, then record whether a brand is named or cited. A mention means the answer named the company. A citation means a source was referenced. Neither one proves a person saw that answer or hired the business.
Ash notes that published research from measurement firms supports covering distinct customer needs, keeping useful context in the question, and repeating tests. It does not prove there is one magic number of prompts that represents “the market.” A dashboard built around flattering “best provider” questions will tell a different story than one built around price, urgency, comparisons, or a customer thinking about switching.
Start with the job the customer is trying to get done
The useful starting point is not “HVAC” or “SEO.” It is a narrow decision: the service, the audience, the place, and what the person is trying to do. For a local contractor, that might mean separately watching emergency repair, replacement estimates, maintenance plans, and “should I switch from my current company?” Those are different jobs, and they deserve different questions.
Those questions should come from evidence of real demand, not from a list an AI invented because it sounded plausible. Customer calls, sales objections, website searches, reviews, and the way people describe a no-heat night or a high summer bill are better raw material than a generated list of “best HVAC company near me” variations. An AI can help organize that material. It cannot turn guesswork into proof that customers ask those questions.
Do not let easy questions overweight the score
If ten slightly reworded fee or price questions sit next to one question about trust or emergency response, a simple average will make price look ten times more important than it may be. Ash’s warning, drawn from Peec research reported in Search Engine Journal, is to test meaningful differences in customer needs and keep near-duplicates together instead of letting cosmetic rewrites dominate the score.
Write the questions the way a customer would ask them. Add the details that change the decision, such as location or urgency. Do not load the prompt with praise for your own company, and do not tell the assistant to “list 20 providers” just to make a report look fuller. Peec’s wording research found that list-style and keyword-style prompts can produce more brand names than an open-ended question. If you change the format, you may change the chance of being mentioned.
Being named in the question is not a visibility win
An unbranded question asks AI to come up with the options: “Who should I call for a furnace that stopped working overnight in Knoxville?” A branded question already supplies a name: “How does ABC Heating compare with XYZ Cooling for a weekend no-heat call?” Both are useful, but they measure different things. If your company appears after the prompt already named you, that is not evidence the assistant discovered you on its own.
Ash offers an illustrative 100-question layout for one category and market: about 80 unbranded questions covering discovery, needs, trade-offs, and the decision to act, and about 20 branded questions covering comparisons and existing-customer advice. He is clear that this is a planning example, not a scientifically proven standard, and not an estimate of how often people actually ask each type of question.
Repeat the test, and do not overclaim the number
The same question can produce different answers on different runs, and answers collected through a company’s software interface may differ from what a customer sees in the public chatbot. Graphite’s research, as described by Ash, treated roughly ten responses as a possible starting point for a rough estimate, not as proof that a small change is meaningful. Daily collection over a month describes that month. It is not the same as proving the result would hold under unchanged conditions.
Real customers also talk in conversations, not one-off survey questions. A person may start with “My heating bill jumped,” add details about the house and the current contractor, and only then ask whether to repair, replace, or switch. Ash points to vendor research suggesting earlier context can change later recommendations. That kind of path is worth testing separately, with made-up or consented details, and without pretending a scripted scenario is a real customer chat.
What a local business should take from this
Freeze a core set of questions, write down why each one is there, and do not quietly drop a question because the answer is unflattering. If the set has to change, mark the break so older scores are not compared as if nothing happened. Treat a failed collection as missing data, not as proof you were absent. And never read “the brand appeared in 32 percent of these test answers” as “32 percent of customers saw us.” Ash uses that kind of figure only as an example of careful wording, not as a finding about consumer behavior.
For a shop owner, the practical version is simple. Make a short list of the questions a real customer would ask before they call you. Separate the ones that do not mention your name from the ones that do. Run them more than once. Read the full answers, not just a score. Then ask whether AI search is sending the right jobs your way, or only answering questions you already look good on.