How to check what ChatGPT says about your business

To see what ChatGPT says about your business, run a fixed list of prompts in a clean session. Record every answer with the sources it cites, and repeat the same list every month. One question in your own account tells you very little, because ChatGPT tailors answers to your history, your settings and whether it searched the web.

Why one question in your own account is misleading

ChatGPT answers from two places. The first is what the model learned during training, which can be months or years old. The second is live web search: the assistant pulls current pages and shows them as links or source chips under the answer. If an answer has no links, assume it came from training data.

Your account adds four more factors that a stranger's session does not have:

Factor	Where you find it	Effect on your test
Memory	Settings, Personalization, Memory	The model remembers that you own the business and leans toward naming it
Custom instructions	Settings, Personalization	Your stated location, job or preferences shape the answer
Chat history	Sidebar	Earlier chats about your company can be referenced
Search toggle	Tools menu in the message box	Decides whether you see training data or current web pages

Hypothetical example: a Jumeirah clinic owner asks "best dental clinic in Jumeirah" from an account that has discussed the clinic for months, and the clinic comes first. A tourist with a fresh phone asks the same question and gets three other clinics. Both answers are honest. Only the second one tells the owner what customers see.

A Temporary Chat does not use memory and is not saved to history. That makes it the closest you can get to a stranger without a second device.

Set up a clean test

Before you type a single prompt, fix the conditions and write them down:

Use a logged-out browser window, or a Temporary Chat in an account with memory turned off.
Note the model shown in the interface and the date.
Run each prompt twice: once with search enabled, once without. With search on, ChatGPT usually shows sources; without it, you see what the model "remembers".
Use neutral phrasing a customer would use. Do not mention that you own a business.
If you serve Russian or Arabic speakers, run the same prompts in those languages as separate rows.

A VPN set to another country changes local results, so test from the UAE if your customers are here.

The prompt set

Write 15 to 20 prompts and keep them identical month to month. Mix five types, because each type shows a different weakness.

Prompt type	Hypothetical example: dental clinic in Jumeirah	What it tells you
Brand	"What do you know about [Clinic name] in Dubai?"	Whether the model knows you exist and gets facts right
Category + area	"Best dental clinic in Jumeirah for implants"	Whether you appear when the customer does not know your name
Comparison	"[Clinic name] vs [competitor]: which is better for veneers?"	How you are described next to rivals
Problem	"My tooth broke on Friday night in Dubai, where can I go?"	Whether you show up for urgent, situational needs
Practical	"How much does teeth whitening cost at [Clinic name]?"	Whether prices, hours, location and booking info are current

Hypothetical example: a car rental company in Business Bay would swap these for "rent a G-Class for a weekend in Dubai", "car rental that delivers to DXB terminal 3" and "[Company] deposit and insurance terms".

Keep the category prompts in the words your customers use. If people ask for "shisha lounge" and your site says "hookah bar", that mismatch alone is worth a row in the log.

What to record in the log

A simple sheet with one row per prompt per run is enough. Columns:

Date, model, language, search on or off
Prompt text
Mentioned: yes or no
Position: first, second, third, later
How you are described, copied word for word
Factual errors: wrong address, old phone, closed branch, invented service
Competitors named
Sources cited, with URLs

The word-for-word description column matters most. It shows which phrases the model associates with you. If ChatGPT calls a family clinic "a cosmetic surgery centre", that phrase came from somewhere, and you need to find where.

Open every cited source

When search is on, ChatGPT lists the pages it used. Open every one. You will usually find three groups:

Your own pages. Check that the page it picked is the one you would choose. If the model cites a three-year-old blog post instead of your services page, your site structure is sending it to the wrong place.
Third-party listings: directories, review sites, map profiles, booking platforms. Wrong hours or an old phone number here travel straight into the answer.
Articles and roundups: "top 10 clinics in Dubai" style pages. If competitors appear in these and you do not, that explains a large part of the gap. Why AI recommends your competitor goes through the usual reasons.

Write down which sources appear for category prompts but never mention you. That list becomes your outreach and listings plan.

Check the other assistants your customers use

ChatGPT is one of several assistants. Run a shorter version of the same set, five to eight prompts, in:

Assistant	Where its local data tends to come from	Your first check
Google AI Overviews and Gemini	Google Search index and Google Business Profile	Profile categories, hours, photos, reviews
Microsoft Copilot	Bing index and Bing Places	Whether your site is indexed in Bing Webmaster Tools
Perplexity	Live web search with visible citations	Which pages it cites for your category
ChatGPT	Training data plus web search	Covered above

If Google's AI Overview gets your hours wrong, the fix is almost always the Google Business Profile, covered in the local SEO checklist.

Technical checks behind a bad result

Some bad answers come from settings on your site. Ask your developer, or check yourself:

robots.txt. OpenAI documents separate crawlers: OAI-SearchBot for search results and GPTBot for model training. If your robots.txt or a security plugin blocks OAI-SearchBot, your pages are less likely to appear as cited sources in ChatGPT search. Decide on GPTBot separately; blocking training does not have to mean blocking search.
Content in the HTML. Open your page, view the source, and search for your phone number and main service name. If they only appear after JavaScript runs, some crawlers will see an empty page.
Bing indexing. Copilot runs on Bing, and other assistants use it too. Submit your sitemap in Bing Webmaster Tools and check how many pages are indexed.
Google indexing. Use Google Search Console to confirm your key pages are indexed and not marked as duplicates.
llms.txt. A short file at the root of your site that lists your main pages in plain text. No assistant promises to read it, and it costs little to add. Two of the sites I built, Moon Mars and Lucent Rental, ship with llms.txt and a sitemap. The llms.txt guide explains what to put in it.

When the answer is wrong or you are missing

You cannot edit ChatGPT's answer directly. You change what it reads, then retest.

Wrong facts. Find the cited page with the error. Correct your own site first, then the directory or listing. Make the same facts (name, address, phone, hours, services) match everywhere.
Missing from category prompts. Create a page on your site that answers that exact question: "Emergency dentist in Jumeirah, open Fridays" beats a generic "Services" page. Put the answer in the first two sentences, since assistants pull from the opening lines most often.
Described as something you are not. Rewrite the first paragraph of your homepage and your Google Business Profile description so they state what you do in plain words.
Competitor named in your place. Look at the roundup articles that list them. Some accept corrections or new entries.

You can also use the thumbs-down button in ChatGPT and explain the error. Treat it as a minor signal; the page fixes do most of the work. The broader mechanics are in the pillar piece on how AI assistants choose who to recommend.

What not to do

Testing from your own account with memory on and calling the result "visibility".
Changing the prompt wording every month. You lose the ability to compare.
Asking ChatGPT "Is my business good?" and treating flattery as data.
Paying anyone who promises that ChatGPT will recommend you. Nobody controls that, and a contractor who says otherwise is selling a claim.
Stuffing pages with "best clinic in Dubai" phrasing. Assistants quote specific, checkable statements: services, areas, hours, prices, licences.
Publishing fake reviews or invented awards to look stronger in answers. In the UAE this creates legal risk under advertising and consumer rules; check with a lawyer before any claim you cannot document.
Blocking all AI crawlers by default because a plugin suggested it, then wondering why you never appear as a source.

How often to repeat the test

Run the full set once a month, on roughly the same date. Run it again two to four weeks after any major change: a new website, a new branch, a rebrand, a new Google Business Profile category. Changes on your site can take weeks to reach assistants, and training data can lag much longer, so compare runs over three months before you judge a fix.

If you hand this to a contractor, ask them to deliver the same sheet each month with the prompt set unchanged, plus a list of the pages they edited in between. When I build websites, I handle the site side of this work: page structure, llms.txt, sitemap and crawler settings.

Checklist

Logged-out window or Temporary Chat, memory off.
Model name and date written down.
15 to 20 prompts across brand, category, comparison, problem and practical types.
Each prompt run with search on and off.
Separate rows for Russian or Arabic if you serve those customers.
Log: mentioned, position, exact description, errors, competitors, sources.
Every cited source opened and sorted: your pages, listings, roundups.
Short prompt set repeated in Google AI Overviews, Gemini, Copilot and Perplexity.
robots.txt checked for OAI-SearchBot and GPTBot.
Key facts present in raw HTML.
Sitemap submitted to Bing Webmaster Tools and Google Search Console.
Name, address, phone and hours identical across site, Google Business Profile and directories.
One page per important customer question, with the answer in the first two sentences.
Same prompt set rerun next month, results compared row by row.

Questions and answers

Can I ask ChatGPT once and trust the answer?

No. The same prompt can return different businesses on different days, in different accounts and with search on or off. Treat one answer as one data point and look at the pattern across many prompts.

Is it possible to remove a wrong statement about my business from ChatGPT?

There is no form for editing an answer. You fix the pages ChatGPT reads: your site, your Google Business Profile, directories and press mentions. Then you rerun the test to see whether the answer follows.

Do I need a paid tool to track AI visibility?

For a small business, a spreadsheet and an hour a month cover the basics. Paid trackers help when you need hundreds of prompts across several markets and languages.

Can anyone guarantee that ChatGPT will recommend my business?

No. Nobody controls how a language model ranks answers. A contractor can make your information clear, consistent and easy to cite, and show you a before-and-after log of your prompt set.