What sources ChatGPT uses to recommend local businesses.
Nobody outside OpenAI can see the full retrieval pipeline. Quite a lot is observable from the outside anyway, and the difference between the two is the whole point of this page.
What you can see for yourself.
It searches the live web when the question is local
Ask ChatGPT for a plumber in a named town with search enabled and the reply shows a search step and a list of cited pages. The citations are visible in the interface. Open them and you will find what it read: a mix of the business's own site, directories, review platforms and local press. This is observable in any session and does not depend on anyone's claim about the model.
It reads the page text, not the page as you see it
The retrieval client fetches HTML and reads the text in it. Content that appears only after a script runs in a browser, or only inside an image, is not in what it reads. You can check what it sees by viewing your own page source, or by running the free audit, which reports whether the key facts are present in the served text.
It is blocked when you block it
OpenAI publishes the names of its crawlers and the rule files they honor. A firewall rule or a robots.txt line that denies them denies retrieval. This is one of the checks in the audit because it is the single most common reason a business with a good website is absent from answers: the answer engine was never allowed in.
It prefers agreement across sources
This one is inferred, not documented, so it is labeled as such. Practitioners report that when the same business name, address and service list appear on the business site, on Google, on Yelp and in a local news item, the business is named more reliably than when those sources disagree. Our own sample has not yet confirmed it, and no published mechanism for it exists.
How we are measuring this, and what we have so far.
The source classes below are a working list, not a finding. To say which sources ChatGPT actually cites for local questions, and how often, you need recorded answers. This is how Fowned records them.
| Field | What is recorded |
|---|---|
| Question | The exact wording asked, frozen before collection |
| Town and trade | The named place and service category in the question |
| Engine and mode | The assistant, with search on or off, and the date of the answer |
| Cited URLs | Every source link the answer shows, copied as shown |
| Source class | Each cited URL assigned to one class from the table below |
| Businesses named | Every business the answer recommends, whether or not it is linked |
- Collection. The same frozen questions are asked across several towns and trades, each answer is saved in full, and nothing is scored until the set is complete.
- Reporting. Every share is published with its count and denominator, for example 12 of 40 answers, never as a bare percentage.
- Current state: insufficient data. Fowned has published zero recorded answers so far. Until a complete set exists, this page states no share, ranking or frequency for any source class.
Where the citations tend to point.
| Source class | What you control | What to do |
|---|---|---|
| Your own website | Everything | Serve the facts as text; keep name, address, phone, hours and service area on a page a crawler can fetch |
| Google Business Profile | Most of it | Complete every field; keep categories and hours current; it is often the first corroborating source |
| Review platforms | Responses, not reviews | Claim the listing; make the name and address match your site exactly |
| Directories and aggregators | Your listing | Fix contradictions before adding new listings; one wrong phone number outweighs ten new directories |
| Local press and community sites | Little | Earn a mention by being mentionable; do not buy one |
| Wikipedia and knowledge bases | Almost none | Ignore unless you are genuinely notable; a deleted entry is worse than none |
What is sold as known but is not.
- A ranking. There is no position to hold. Two people asking the same question in the same hour can get different businesses, and the sample size in most agency reports is one.
- A submission form. There is no list to be added to. Anyone offering to submit your business to ChatGPT is describing a directory listing with new packaging.
- A guaranteed recommendation. The assistant selects from what it retrieved. You can make yourself retrievable and consistent. You cannot make it choose you.
- A causal lift number. Without a sampled baseline before a change and the same sample after, any percentage improvement is a story. Measurement of answer presence is what Citability does; this site does the repair work it points to.
The other guides in this set.
- Local SEO vs GEO vs AEO for local businesses Three acronyms, one plain comparison table, and which one to fund first.
- GEO for local businesses: does yours need it? Three questions that decide it, what the work actually is, and how to prove it worked.
- AI visibility audit for local businesses: what we check and why Every check in the free audit, with the reason it exists.
- How to choose an AI SEO or GEO agency for a local business Nine questions to ask, and the answers that should end the call.
- How to track leads from ChatGPT and AI search without guessing What the referrer can prove, what it cannot, and what to record instead.
See which of these sources can reach you today.
The free audit checks whether AI retrieval crawlers are permitted, whether your facts are in the served text and whether your identity is machine-readable.