When somebody asks ChatGPT for a good bakery in Brooklyn, the model searches the web mid-answer and reads whatever it finds. For a single-location business it almost always finds the Google profile, the map data behind it, and the services that copy that data. If your profile holds only a name, an address and a phone number, the model knows exactly three things about you and none of them answers the customer’s question.
What a model pulls out of a profile
Not every field carries the same weight. Below is the order you can infer from what models actually quote in their answers.
| Field | What the model needs it for | Weight |
|---|---|---|
| Primary category | Decides which questions you are considered for at all | very high |
| Reviews and your replies | The only text about you that you did not write | very high |
| Opening hours | Half of all questions are some form of "what is open now" | high |
| Services and products | Turns a broad category into a quotable specific | high |
| Questions and answers | Ready-made sentences answering customer questions | medium, badly underused |
| Business description | Context, though models rarely quote it directly | medium |
| Photos | Indirectly, by making the profile look alive | low |
Five mistakes we see most often
1. A broad category instead of a specific one
Restaurant instead of pizzeria. Contractor instead of roofer. The primary category acts as a filter: asked about pizza, a model considers pizzerias, not every restaurant in town. Pick the narrowest category that is still true and add the rest as secondary.
2. A business name with the town and trade bolted on
This one is a trap few people know about. A name like "Bella Pizza Brooklyn Best" looks clever and causes two problems. It breaks Google’s rules, which require the name on your sign. And it corrupts measurement: every answer about Brooklyn contains the word Brooklyn, so a shallow tool counts a hit even though the model never named the business.
3. Hours that disagree between services
Google says you close at six, Apple Maps says five, Facebook says you are shut on Sunday when you are not. The model sees a contradiction and usually skips the business rather than guess. This is one of the three most common causes of absence we find.
4. Two profiles for the same business
They appear after a move, or because somebody once added the place by hand. Reviews split across two profiles, details drift apart, and the model cannot tell which is current. Duplicate profiles need to be reported and merged.
5. An empty questions section
The most underused field in the whole profile. You can ask a question yourself and answer it as the owner. It is the only place where you produce a ready-made question-and-answer pair, which is exactly the shape a model wants when it is assembling a reply.
What to put in the questions section
Start with the five questions you answer on the phone most often. For a food business they usually look like this:
- Can I book a table, and how?
- Is there parking, and is it free?
- Do you have gluten-free or vegan options?
- How late are you open on Friday and Saturday?
- Do you take cards?
Answer in a full sentence, not one word. "Yes, there is free parking for eight cars behind the building" is something a model can quote. A bare "Yes" cannot be used in any answer at all.
Matching details, the dull part that matters most
Name, address and phone must look identical everywhere. Not similar, identical. The places to check:
- Google Business Profile
- Apple Business Connect, which feeds Maps on iPhone
- Bing Places, the source behind Copilot
- Facebook, if you have a page
- Vertical sites: booking platforms, trade directories, review sites in your industry
- Your own site footer and structured data
The usual mismatches are Street versus St, a phone number with and without the country code, and a legal suffix such as LLC present in one place and dropped everywhere else. Pick one form and carry it across.
How to tell whether it worked
Profile work does not show up overnight. Crawlers have to read the corrected data first. A realistic timeline:
| When | What you see |
|---|---|
| 1 to 3 days | Changes are approved and live on the profile itself |
| 1 to 2 weeks | Data spreads into maps and the services that copy them |
| 4 to 8 weeks | Models start quoting the corrected information |
To see a difference you need a starting point. Measure your share of answers before the fixes and after, with the same method and the same questions. The method is in the post on generative engine optimization, and if you would rather not do it by hand, the free scan takes two minutes.
The profile is not the whole job
A complete profile gives the model material, but it does not remove the other obstacles. If your site blocks assistant crawlers or carries no structured data, the model is still guessing. The next two steps are covered separately: LocalBusiness structured data and AI crawlers in robots.txt.