Local SEO · October 21, 2026 · 8 min read
How to Get ChatGPT to Recommend Your Local Business
There is no switch. ChatGPT assembles local answers from sources it can reach and verify at question time, so consistent facts, plain pages and an unblocked site are what get you named.
By FluxWriter
Get ChatGPT to recommend my business is a search that assumes a switch exists somewhere, and there isn't one. What exists is a pipeline: the assistant assembles local candidates from sources it can reach and check at the moment someone asks, then writes a short answer naming two or three of them. Most businesses left out are left out for dull reasons — hours that disagree across three sites, a homepage that never says which towns you cover, a crawler rule somebody switched on two years ago. This walks through where those candidates come from, what an assistant can verify about you, and the short list a local business can finish this quarter.
The Answer Is Assembled At Question Time, Not Recalled
Two things can happen when someone asks an assistant for a local recommendation. It can answer from the model's trained memory, frozen in the past and weighted towards businesses a lot of text was written about. Or it can search the live web, read a handful of pages, and summarise them. For local questions the second one matters, because local facts move faster than any training run.
That changes what you are competing for. There is no list of ten positions. The assistant issues its own queries, retrieves a small set of pages, and names businesses it can describe from them. You compete to be inside that set. Being on the second page of anything is the same as being absent. The assistant will also rarely admit it does not know a plumber in your town; it will name someone.
The index behind that search is not one thing you can court. OpenAI runs its own crawlers and has also drawn on third-party search results, and those arrangements have changed before. So stop optimising for one index: be findable, consistent and quotable on the open web, and you are covered whichever plumbing is in place next year.
Assistants Favour Businesses They Can Describe Without Contradiction
An assistant has to produce a confident sentence. "Open until 8pm, and they do emergency callouts" is only writable if the sources agree. Where they disagree, one listing saying 5pm, your site 6pm, a directory from 2019 saying 9pm, the safer move is a business whose facts are clean. Assume that is roughly what happens: an inference about behaviour rather than a documented rule, but it costs nothing to act on.
So the first job is reconciliation, and it is boring. Legal name against trading name. The suite number on some listings and not others. Phone numbers, which is where call tracking quietly does damage: a tracking number that appears on your site and nowhere else reads as a different business. Opening hours, including the holiday hours nobody updated. Your primary category, which should be the words customers use, not your accountant's. The towns you serve, written out as words on a page, not implied by a map pin.
Do this once across your listing, site footer, contact page and the four or five directories that still show up. An afternoon covers most of it. The point is not the directories. It is that an assistant reading five sources finds five that agree.
A Site Without Plain Answers Gives Retrieval Nothing To Quote
Retrieval works on passages, not on reputations. A page whose first two hundred words are about passion and community gives an assistant nothing to lift. "Serving the area with pride since 1998" cannot be turned into "they do Sunday callouts on the north side, usually around $180". The second sentence is the one that gets you named.
Write pages that answer in the first two sentences: what you do, where, who for, roughly what it costs, how fast you turn up. Put the detail underneath for the human who wants it. It is the page you would write for a customer in a hurry, and it happens to be the shape retrieval can extract.
One page per question beats one page carrying everything; your homepage cannot answer nine different searches at once. Start from the phone: write down every distinct question you were asked this week. If that comes to twelve, and you publish two pages a week, the list is covered in six weeks. A forty-question list at the same rate takes twenty weeks, which is most of a year's plan pulled off a notepad.
Prices are the hard one. "Contact us for a quote" is not quotable, and "how much does this cost near me" is among the most common questions people put to an assistant. A published range, with the conditions that move it, makes you answerable. Whether to publish one is a commercial decision, and plenty of trades decide not to.
Six Places A Local Recommendation Can Come From
The layers are worth seeing separately, because you control them to different degrees. The timings below are illustrative rather than anything a vendor publishes.
| Source layer | What it contributes | Your control | Rough time to show |
|---|---|---|---|
| Your own site pages | The sentences that get quoted back | Total | Days to weeks after a recrawl |
| Map and place listing | Name, address, hours, category, area | Total | Days, once verified |
| Reviews on other sites | The specifics an answer repeats | Indirect | Continuous, never on demand |
| Directories, trade bodies | Corroboration that you exist and do this | Mostly | Weeks |
| Local press, forums, suppliers | Mentions that read as independent | None directly | Unpredictable |
| The model's trained memory | What the web said before the cutoff | None | Only when retrained |
Read down the control column. Three of the six are yours outright or close to it, two you can nudge, and the last is out of reach entirely. That last one is what owners worry about most.
Blocking The Wrong Crawler Takes You Off The Answer Surface
One failure mode undoes everything above, and it is invisible from the front of the site: if your server refuses the crawlers that feed the assistant, none of your plain pages exist as far as it is concerned.
OpenAI publishes the user agents it uses, and they are not interchangeable: the crawler that gathers training data is documented separately from the one supporting search. A blanket rule added to robots.txt during the first wave of scraping anxiety may be doing more than whoever added it intended. If someone on your team opted out, the question to ask is what that opt-out was meant to cover, and whether today's rule still matches.
Your CDN is the second place to look. Several hosting and CDN providers added one-click controls for blocking AI crawlers, some on by default. Check what yours is set to, because a rule at the edge overrides anything polite in robots.txt.
Third, rendering. If the text on your service pages only exists after client-side JavaScript runs, a crawler can arrive at what amounts to an empty page. Test it crudely: view the page source, or load the URL with JavaScript off, and check whether your opening two sentences are in the HTML.
Nobody Can Guarantee A Recommendation, And Answers Move Anyway
Two identical questions, asked in two sessions, can name different businesses. There is no position to track, no rank to report, and no fixed shortlist the assistant draws from. Anyone offering guaranteed placement should be asked precisely what is being sold and how delivery would be proven.
What you can do is sample. Open a temporary or logged-out chat so your own history is not shaping the result, ask the four or five questions your customers ask in the words they use, and write down what comes back. Do that monthly, not daily. Absence is one signal. More useful is what it says about you when it does name you: wrong hours, an old address, a service you dropped. Those are correctable, and they hint at which source it leaned on.
Keep the effort proportionate to the evidence. Nobody outside these companies knows what share of local discovery runs through assistants, so spending against a guess is a bad trade. The reassuring part is that almost everything on the list, consistent facts, plain pages, real reviews, an unblocked site, is the work that has always made local search work. You are doing the ordinary job in a format a machine can read.
FAQ
Does ChatGPT read my Google Business Profile?
An answer that searched the web will show you the pages it cited, but that still does not tell you whether your place listing fed it, so treating any single listing as the direct pipeline is a guess. What holds regardless is that your place listing is the cheapest way to make your core facts consistent everywhere they get copied. Keep it accurate because it feeds the wider ecosystem, not because of an assumed integration.
Can I pay to appear in an assistant's recommendations?
Organic recommendations and paid placement are different products, and the paid side has been changing quickly. If an agency offers to place you, ask what specifically is being bought, on which surface, and how you would verify that it appeared. A guarantee nobody can verify is not a guarantee.
How long after I fix things will the answers change?
Site changes depend on when your pages are recrawled, which usually runs from days to a few weeks on an active site. Listing corrections tend to move faster once they are verified. Anything sitting in the model's trained memory only updates when the model itself is retrained, which is not on your schedule.
The Practical Takeaway
This week, do the reconciliation. Pull up your listing, site footer, contact page and the three directories that still rank for your trade, and make name, address, phone and hours identical across all of them, tracking numbers included. Check robots.txt and your CDN's bot settings for a rule nobody remembers adding. Then write down the questions you were actually asked on the phone this week; that list is your publishing plan for the next six weeks. Before any of that, open a temporary chat today and ask the exact question your best customer would type, then note every business it names and every fact it gets wrong about you.
Question pages written by FluxWriter can link to the ones it has already published, up to three per post where the wording matches (not yet on Wix), so six weeks of phone questions starts to read as a connected set rather than a dozen pages standing on their own.