AI visibility · Method · updated 31 October 2026

How we measure whether ChatGPT names a business

The short answer: we ask ChatGPT the questions a business's customers ask, through OpenAI's official API, with web search switched on and the location set to the business's city. We keep every answer in full with the links it cited, then count: was the business named, was its website cited, who was named instead, and which pages the answers relied on.

The setup

• Assistant. ChatGPT's model through OpenAI's API (gpt-5.4 in October 2026). A subscription keeps the same model setting every week, so the weeks are compared like with like.
• Web search: always on. Every question must search the web before answering, the way ChatGPT does when someone asks for a local business.
• Location: your city. The approximate location is set to the city (and region and country), so the answers are the local ones.
• Questions. The free snapshot asks eight questions, each once. A subscription asks twelve questions, each five times, and the same set every week.
• What we keep. For every answer: the time, the model OpenAI reports, the location, the full text, every cited link and the usage. Nothing is summarised away before it is counted.

The questions

The free snapshot builds its eight questions from fixed templates, filled in with your type of business, your main services and your city. Most use the plain word customers type for your trade (mechanic, hotel, dentist), not the way you describe yourself: a niche label that only you use would make the result look better than it is. For an auto repair shop in Denver they would read like this:

1. best mechanic in Denver
2. who is the best mechanic in Denver?
3. top rated mechanic in Denver
4. mechanic in Denver with good reviews
5. I need a mechanic in Denver, who do you recommend?
6. best auto repair shop in Denver
7. who offers brake repair in Denver?
8. I need an oil change in Denver, who do you recommend?

Before anything is asked, you confirm what the business is and what it offers, so the questions are about your trade and not a guess. A subscription asks twelve questions built the same way, and keeps the same set every week so the weeks can be compared.

What we count

• Named: the answer names your business in its text.
• Cited: the answer links to your website.
• Cited but not named: the answer used your page and recommended someone else. We always report this separately, because it is a different problem from not being found.
• Named instead: every business the answers recommended, with how often.
• Sources: every page the answers linked to, sorted by type: the businesses' own websites, Google business profiles, directories and review sites, articles and forums.

Why each question is asked more than once

The same question can get a different answer a minute later. One answer is one observation, not a position. That is why the free snapshot is labelled a sample, and why a subscription asks each question five times: "named in 3 of 5" says more than "named once".

Small numbers stay small. With 15 answers, a change of one or two answers is noise, and we say so instead of calling it a result.

What a measurement can't tell you

• What each person sees. Real ChatGPT users have their own history, settings and memory. An API sample is repeatable; it is not anyone's personal session.
• Why ChatGPT chose. Nobody outside OpenAI knows exactly how it picks. We report what the answers said and cited, and we label our ideas about the reasons as ideas.
• Other models. Results depend on the model. In our first test in Tulsa, a healthcare directory was cited in 58% of a smaller model's answers and in 21% of gpt-5.4's. We report the model with every number and never mix two models in one comparison.
• Other assistants. We measure ChatGPT. Google's AI answers and Perplexity are not measured yet.

From numbers to a plan

• Every claim is traceable. Each number and each statement in a report comes from a stored answer, a stored page, or your written confirmation.
• Every page claim is checked live. Before we say something is missing from your website, we read the page as delivered and as rendered in a browser, on the day the report goes out.
• Every action is labelled: a confirmed gap, an opportunity, or a hypothesis. At most three, in order.
• Your facts, confirmed by you. Services, prices, hours, credentials and anything regulated (clinical claims, for example) are only written as true once you confirm them.
• Work and results are kept apart. What was changed on your pages is one list; what the answers did afterwards is another. One never promises the other.

Try it

The free snapshot runs this method for your business with eight questions, on screen in about a minute. Monitor ($149 a month) adds the five-times sample, a full report with up to three things to do first and a new measurement every week. For an example of what the numbers look like, read the Tulsa case study, or the guide Why doesn't ChatGPT recommend my business?

ChatGPT is a trademark of OpenAI; Talsara is not affiliated with OpenAI. Questions about the method: hello@talsara.com.