Type "best HVAC company in Wilmington NC" into ChatGPT and you'll get back a short, confident list. Two to five businesses, a sentence or two about each, sometimes with links to the sources it read. For the businesses named, that's a referral from the most-used AI tool on the planet. For everyone else in town, it's a customer they never knew existed.
So how does ChatGPT pick? That's the question this post takes apart. Part of the answer is observable and well-supported. Part of it is genuinely unknown to anyone outside OpenAI, and I'll be clear about which is which, because there's a cottage industry forming around pretending to know more than anyone does.
QuickBooks is an accounting system, not a financial intelligence system, and that distinction matters more than most business owners realize. The same kind of distinction applies here: ChatGPT is not a directory with a ranking algorithm you can reverse-engineer. It's a language model that, for questions like this, goes and does research. Once you see it that way, the playbook gets a lot clearer.
Two modes: memory and search
When ChatGPT answers a question about businesses, it's working from one of two places, and often both.
Mode one: what the model already knows
The model was trained on a huge snapshot of the public web. Businesses that have been written about, reviewed, listed, and discussed over years exist somewhere in that training data. For well-known national brands, this memory is rich. For a twelve-person plumbing company, it's thin and possibly stale: an old address, a service you discontinued, or nothing at all.
You can't edit a model's memory. You can only influence what the next snapshot absorbs, by making the public record about your business broad, accurate, and consistent. That's slow-acting work. The fast-acting work is in mode two.
Mode two: live, search-backed answers
For current, local, and specific questions, ChatGPT increasingly runs a real web search, reads what comes back, and composes its answer from those results, often with citations you can click. OpenAI has documented ChatGPT's web search capability publicly at openai.com, and you can watch it happen: ask a local recommendation question and you'll frequently see it browsing, then citing sources.
This changes the strategic picture completely. If ChatGPT is reading search results to write its answer, then the question "how do I get recommended by ChatGPT" largely becomes "how do I show up well in the sources ChatGPT reads." Those sources are web pages: directory listings, review sites, local roundup articles, and, crucially, your own website. ChatGPT's search has historically drawn on Bing's index, which is worth sitting with for a second, because almost no small business has ever paid attention to Bing. Verifying your site with Bing Webmaster Tools is free and puts you in the index that may be feeding the answers, in a queue where most of your competitors never bothered to stand.
What appears to make a business nameable
Watch enough ChatGPT answers about local businesses and patterns emerge. I say "appears to" deliberately. These are observations, not documentation. But they're consistent observations, and they line up with how the system plausibly works.
It names businesses with a clear, consistent identity
To recommend you, the system has to be confident you're a single real entity: one name, one phone number, one location or service area, corroborated across multiple independent sources. A business whose name appears three slightly different ways across the web, with two phone numbers and a dead address on an old directory, is harder to be confident about. When you're writing a three-name answer and confidence is the currency, fuzzy entities get skipped.
The fix is unglamorous: pick the canonical version of your name, address, and phone, and make every listing on the web match it. Your Google Business Profile should be the anchor copy that everything else agrees with.
It leans on third-party corroboration
ChatGPT's answers about businesses tend to echo what review sites, local press, and roundup articles say. That makes sense: a search-backed answer is built from search results, and those are the pages that rank for "best roofer in" queries. Your own website saying you're excellent carries some weight. Five independent sources saying it carries far more. Review volume, review recency, mentions in local media, and presence on the directories that rank in your market all feed the evidence pool the model reads from.
It draws on businesses whose websites answer questions
When the question is informational rather than purely "who should I call," like "what does a mini-split install cost" or "do I need a permit to replace a water heater in North Carolina," ChatGPT builds the answer from pages that actually address it. A contractor whose site has a genuinely useful page on exactly that topic can end up cited as a source, name and link included. A site that's five paragraphs of "quality service since 1998" gives the model nothing to use. This is the same content bar that good SEO has always rewarded; the difference is that the payoff now includes being quoted in the answer itself, not just ranked under it.
It can read structured data, and clarity costs you nothing
Whether ChatGPT's pipeline directly consumes schema markup isn't publicly documented, and I won't claim it is. What's true is that structured data using the schema.org vocabulary makes your facts unambiguous to any machine reading your pages, that Google's systems demonstrably use it, and that it takes a competent developer very little time to add. When uncertainty is high, you place the cheap bets that pay off in every scenario.
What nobody outside OpenAI knows
Here's the honesty section, and I'd encourage you to hold any vendor in this space to it.
Nobody outside OpenAI knows how results get weighted inside an answer, why it names three businesses instead of five, or why it picks one over another when both look similar on paper. The behavior shifts between model versions, sometimes noticeably, without announcement. The same question asked twice can produce different lists. There is no submission form, no ranking report, and no "ChatGPT optimization certification" that means anything.
That has two practical implications. First, anyone promising guaranteed placement in ChatGPT answers is promising something no one can deliver. Second, your strategy has to be robust to change, which means building the durable evidence trail (consistent identity, real reviews, useful content, crawlable site) rather than chasing whatever trick worked in a screenshot last month. The durable work also improves your Google rankings, your conversion rate, and your referrals, so it pays off even in the scenario where AI referrals stay a small channel for years.
Also worth saying plainly: today, far more local customers still find businesses through Google than through ChatGPT. This channel is growing and worth getting ahead of precisely because your competitors aren't there yet, but it's an addition to the fundamentals, not a replacement for them.
Recommendation questions vs. research questions
One more distinction that sharpens the playbook: customers ask AI assistants two different kinds of questions, and you win them with different assets.
The first kind is the direct recommendation: "who should I call for a roof leak in Wilmington." Here ChatGPT is essentially compiling a shortlist from the evidence pool, and the assets that matter are your identity, your reviews, and your presence on the sources it reads. Your website matters mostly as corroboration.
The second kind is the research question: "how do I know if my roof needs replacing or just repair." Here ChatGPT is synthesizing an explanation, and the asset that matters is whether your site contains a page worth drawing from. Win enough of these and something compounding happens: the customer first meets your business as the expert source behind the answer, days before they're ready to hire anyone. By the time they ask the recommendation question, yours is a name they've already seen.
Most competitors in a given trade and town are contesting neither question. You can contest both with one website, built once, maintained honestly.
What to do this month
- Ask ChatGPT the questions your customers would ask: "best [your trade] in [your town]," plus two or three informational questions about your services. Save the answers. That's your baseline, and it's more informative than anything an agency report will tell you.
- Fix every name, address, and phone inconsistency you can find across the web.
- Verify your site in Bing Webmaster Tools and confirm it's indexed.
- Look at which sites ChatGPT cited in your baseline answers. Those are the sources that matter in your market. Make sure you're present and accurately represented on them.
- Write or upgrade one page on your site that thoroughly answers a question customers actually ask. Repeat monthly.
We work with a lot of trades businesses, and the pattern holds across plumbing, HVAC, and cleaning and restoration: the market leader in AI answers is rarely the biggest company in town. It's the one with the cleanest evidence trail.
If you'd rather have this done for you
This is the work we build into every Omnyra site. We're a veteran-owned shop in Wilmington, NC, with 1,500+ small business sites built in the last 90 days, including airsupporthvac.com, sanosteam.com, and ramartrans.com. The process is done-with-you: we build your site live on a call, first draft in 24 hours, live in 7 days, guaranteed.
AEO is built into our Standard tier at $2,000 plus $200/mo. Tiers run from $500 Minimal up to $3,500 Max with a 24/7 AI receptionist, and from $6,000 for Super Max. Pay-in-4 and Klarna available. Compare tiers at /pricing or book a call and we'll pull your ChatGPT baseline together, live.