How to Check What ChatGPT Knows About Your Business

Set up a fair check

Use Temporary Chat so an old conversation about your company doesn't carry into the result. OpenAI says Temporary Chats don't use existing memories or create new ones. Turn on Search, or confirm that the response shows web sources, because current hours, services and local recommendations depend on up-to-date public information.

State the city or service area in every local question. ChatGPT Search can estimate a general location from an internet connection, and precise device location is optional, but neither is a good substitute for naming the market you mean. Keep the same language and location wording in both runs.

Do not begin by telling ChatGPT what the correct answer should be. A prompt such as ‘We are the leading emergency plumber in town, do you know us?’ tests whether the system will follow your framing. It does not test what a customer can find from the public record.

• Use two separate Temporary Chats
• Select Search and open the cited sources
• Keep the business name, city and customer need unchanged
• Save both full answers, including the date

Ask three questions about the public facts

Start with questions where the company is named. These show whether ChatGPT can distinguish the business from companies with similar names and whether it can recover the facts a customer might rely on before calling or driving over.

Replace the bracketed fields before sending each prompt. For a service-area business, ask for the area served rather than pushing ChatGPT to invent a storefront address.

• ‘What kind of business is [business name] in [city or region]?’
• ‘What services does [business name] publicly say it offers, and what area does it serve?’
• ‘What hours, contact details and booking method can you confirm for [business name]?’

Ask three questions about customer fit

Recognition is the easy part. A customer usually asks about a need, not whether an AI system has heard of the company. Use one job that produces real revenue and add a constraint that changes the choice, such as location, timing, accessibility, property type or a required credential.

The first question below doesn't name the business. That matters. It shows whether the company enters the answer without being handed to ChatGPT. The next two questions test whether the stated reason is supported rather than merely plausible-sounding.

• ‘Which [type of provider] in [city or region] would you consider for [specific job and constraint]?’
• ‘Would [business name] fit that same need? Give the reasons and the public sources behind them.’
• ‘What would a customer still need to verify before contacting [business name] for that job?’

Read four different results separately

Do not compress the answers into ‘ChatGPT knows us’ or ‘ChatGPT does not know us.’ A business can pass one part and fail another. Read each run against four separate questions, then write down the exact sentence or source that supports your judgment.

Open every citation. OpenAI says search responses may show inline citations and a Sources panel, but a source link is not proof that every sentence on the page was used correctly. Confirm that the cited page is current and actually supports the claim beside it.

• Identity: Did it find the right company in the right place, without blending in a namesake?
• Accuracy: Are the hours, contact details, services and service area current?
• Fit: Does it connect the business to the specific job and constraint a customer named?
• Evidence: Do the visible sources support the reasons, or is the answer filling gaps with assumptions?

Repeat the questions without shopping for a win

Two runs are a screening check, not a visibility benchmark. They are useful because obvious instability should stop you from treating one answer as a verdict. Save both, even if one recommends the business and the other leaves it out. Regenerating until you get a flattering answer only hides the problem you were trying to measure.

A 2026 pre-registered study tested 2,208 search-grounded responses from ChatGPT, Claude, Gemini and Perplexity across 96 local food-and-drink questions in two Bali markets. Repeated questions and meaning-preserving rewrites produced changing venue sets, and the four systems often disagreed. The study was about restaurants, cafes and bars in those markets, so it cannot supply a universal appearance rate for a roofer or lawyer. It does show why one saved screenshot is thin evidence.

If both runs agree, you have a useful observation under the recorded conditions. If they conflict, mark the result as mixed. A larger audit can add more prompts, runs, platforms and dates, but the first honest move is simply to preserve what happened.

Choose one next action from the result

Fix the first material failure a customer could feel. Wrong hours or an old address deserve attention before a missing recommendation. A confused service description matters before a crawler setting that is already correct. Keep the action tied to the evidence on the screen.

If the public facts are accurate and both runs still disagree, leave the accurate records alone. Document the mixed result and repeat the same test later. There is no useful prize for changing a correct website every time an answer moves.

• Wrong fact: trace it to the cited page or customer-facing profile and correct that record
• Wrong company: make the official name, location and business category unambiguous on the website and major profiles
• Wrong fit: state the real service, service area and customer constraints on the page closest to that decision
• Weak support: add or repair proof a customer can verify, such as a license record, named professional, completed work or association profile
• Accurate but absent: compare the businesses that recur, then look for a real fit or evidence gap before creating new content