What Sources Does ChatGPT Use for Local Recommendations?

ChatGPT pulls local recommendations from two places: its training data and, when search is enabled, live web retrieval. Retrieval-backed answers draw on indexed web pages — business directories, review platforms, best-of listicles, local press — and cite them. The exact mix varies by market, and you can inspect it directly by reading the citations.

Two places ChatGPT gets local answers

Every ChatGPT answer about local businesses comes from one of two systems, and they behave very differently.

Training data is the static option. The model was trained on a snapshot of the web that is months old by the time you talk to it. When ChatGPT answers a local question without searching, it recalls aggregate patterns from that snapshot: businesses that were widely written about, linked, and reviewed before the training cutoff. It cannot know that you opened last spring, moved offices, or doubled your review count in March.

Live retrieval is the dynamic option. When search is enabled — and it often is by default for local queries — the model runs a web search, reads a handful of the pages that come back, and composes its answer from what those pages say. These answers carry citations you can open and read.

The distinction matters because only one of these is a lever you can pull this quarter. You cannot edit training data. You can change what live retrieval finds, starting this week.

The domains that dominate retrieved local answers

When ChatGPT searches for “best HVAC company in Mesa,” the pages it reads are usually not the companies’ own homepages. They are pages that rank for that query and describe many businesses at once:

Source type	Why it gets retrieved
Business directories	Structured, complete listings that rank for “X in [city]” queries
Review platforms	Ratings and review volume map directly onto “best X” questions
Best-of listicles	Their format matches the question format exactly
Local press	Trusted domains with genuinely local coverage
YouTube and Reddit	First-person experience on high-authority domains

Notice what is mostly absent: individual business websites. Your site matters — it is where retrieved pages point, and where the AI confirms details like services and service area. But for the “who should I call” question itself, the model leans on pages that compare and rank.

Review platforms deserve a specific note. They are heavily retrieved for “best X in Y” questions because they are, structurally, exactly what that question asks for: many businesses in one category, in one place, with ratings attached. When ChatGPT tells a customer that a plumber is “highly rated with strong reviews for emergency work,” it is usually paraphrasing a review page it just read. Your review count, rating, and the actual words customers wrote are all inputs to that sentence.

What our tracking lab sees

This is not a guess. Our weekly tracking lab runs 90 answers per week in one consumer niche and records every citation. The most-cited domains in those AI answers were YouTube, Reddit, and major review and news sites — not the businesses’ own websites.

One niche is not every niche, and we would not pretend otherwise. But the same shape shows up in the local-business audits we run: the pattern that predicts visibility best is presence on the sources the AI answers already cite — directories, review sites, local press — plus consistent name/address/phone data and a complete Google Business Profile.

The practical implication is uncomfortable for anyone who has spent years polishing only their own website: the pages that decide whether ChatGPT names you are mostly pages you do not own. You can influence them — claim them, complete them, earn reviews on them, get covered by them — but you cannot skip them.

How to find the exact sources for your market

The generic list above is a starting point. The list that matters is the one for your category, in your city. There are two ways to get it.

Do it manually. Open ChatGPT with search enabled and ask a local question the way a real customer would: “best mortgage broker in Plano,” “who should I call for a burst pipe in Mesa,” “compare dentists near downtown Raleigh.” Read every citation. Write down the domains. Then repeat the same questions on two or three other days, because retrieval-backed answers vary from run to run. The domains that keep reappearing are your market’s source list.

Or read it off an audit. Every BizWhiz audit produces a citation-source inventory: the exact domains the AI answers drew on in that market. The audit battery asks 20 localized customer questions — four intents, multiple phrasings — against six AI systems, producing 100+ timestamped answers per audit. Every response is stored raw and quoted verbatim. The inventory falls out of that data; nothing is inferred. You can see the full method on how it works.

Either way, the output is the same kind of thing: a short, concrete list of domains where being present, complete, and well-reviewed changes what the AI reads about you.

The work: get present on those domains

Once you have the source list, the job stops being mysterious. It is presence work:

• Claim and complete your profiles on every directory that appears in the citations. Empty or half-finished listings get retrieved too — they just say less about you.
• Build reviews on the platforms the answers cite. Review platforms are heavily retrieved for “best X in Y” questions. Volume, recency, and rating all live on pages the model reads.
• Pitch the listicles and local press that keep showing up. A spot on a “best of [city]” page that ChatGPT already cites is worth more than most link-building.
• Make your own site machine-readable. Clear service and location pages, LocalBusiness schema, and consistent name, address, and phone. An llms.txt file is an emerging convention some AI crawlers read — cheap to add, not a magic bullet.
• Fix inconsistencies. If your phone number differs between two directories, the model is reading conflicting evidence about you.

To be plain about limits: nobody can guarantee an AI mention, and anyone promising one is guessing. This work improves the odds by fixing the inputs AI systems observably rely on — and it is measurable month over month.

What to do this week

1. Ask ChatGPT, with search on, three phrasings of “best [your category] in [your city].” Save every cited domain.
2. Repeat the same three questions on two more days. Tally which domains appear more than once.
3. Check whether your business exists on the top five recurring domains. Claim and complete anything missing or thin.
4. Compare your name, address, and phone across those listings and your site. Fix every mismatch.
5. Fill out your Google Business Profile completely — categories, services, hours, photos, service area.
6. Ask five recent customers for a review on the platform your citations mention most.
7. Add LocalBusiness schema to your homepage if it is not there yet.

If a customer question like “why does my competitor keep getting named instead of me” is on your mind, the mechanism behind that is covered in how AI assistants choose businesses to recommend.

Keep both channels honest

AI-referred customers are real: 37% of consumers now start searches with AI instead of Google. But so is the old channel — roughly 84% of US consumers still use Google daily for local business searches, and only about 24% prefer AI chat for local queries. The right read is not “abandon Google for AI.” It is that the same source presence now feeds both, because Google’s own results are increasingly AI-composed. The domains ChatGPT cites and the profiles Google reads overlap heavily. Fix the sources once and both channels see it.

If you would rather read your market’s actual citation list than guess at it, start with a free AI visibility scan — three questions across two models plus Google, results emailed in about 30 seconds.