How to Check What AI Says About Your Brand: A Prompt Set You Can Run Today

You can audit your own AI reputation in about two hours with no software. Write between forty and eighty questions the way your buyers actually ask them, run them across six engines, and store the full answer text with the date and the engine. That archive is the whole method, because nothing is retained on the engine side and an answer you did not save cannot be compared to anything later.

Most brands have never read what an assistant says about them. The ones that have usually asked a single flattering question, got a reasonable answer, and stopped. That is not a measurement. This is the prompt set we build for clients, written out so you can run it yourself.

The rule that makes the results usable

Ask the question your buyer asks, not the question your marketing wishes they asked. Nobody types your tagline. They type the thing they are worried about.

Run every prompt in a fresh session with no memory of previous chats, because a personalized session tells you about your own account rather than about a stranger. In ChatGPT use a temporary chat. Elsewhere use a private window while signed out.

The six question types

1. The named question

What the engine says when your brand is the subject.

• What is [brand]?
• Is [brand] a reliable [category]?
• Who owns [brand] and where are they based?
• What do customers complain about with [brand]?
• Has [brand] had any legal or regulatory problems?

The last two are uncomfortable and they are the point. Your buyer asks them.

2. The unnamed recommendation

Where you are not mentioned in the prompt at all. This is the one that reveals whether you exist commercially.

• Best [category] in [city or country]
• Who should I hire for [problem]?
• Recommend a [category] for a [company size] business
• What is the most trusted [category] for [use case]?

Ask about your category question and you get a different source mix than the named question. Studies of open recommendation queries put community forums far higher in the citation set than first-party pages, which is the opposite of what the named question produces.

3. The comparison

• [Brand] vs [your real competitor]
• Alternatives to [brand]
• Is [competitor] better than [brand] for [use case]?

Note which competitor the engine reaches for. It is often not the one your sales team names.

4. The trust question

• Is [brand] a scam?
• Is it safe to use [brand]?
• What are the risks of working with [brand]?

Skip this set and you will miss the answers that cost the most. In regulated categories it is the first thing a buyer asks.

5. The person

• Who is [founder or executive name]?
• What is [name]’s background?
• Is [name] connected to [company]?

Individuals fail differently from companies. The usual failure is a merged identity with a namesake rather than an invented allegation.

6. The language and market variants

Ask your top five prompts again in the other language your market uses. The source pool is thinner in most languages other than English, which means fewer sources carry more weight and a single bad thread can dominate. If you sell in Arabic, French or any language where the published record is smaller, this set will look different from the English one.

The six engines to run it on

They disagree, and the disagreement is the diagnosis rather than noise.

• ChatGPT. Run each prompt twice, once with search on and once without. The gap between the two answers tells you whether the problem is retrieval or training memory.
• Perplexity. Retrieves for nearly every query and numbers its citations, so it hands you the source list for free. Start your tracing here.
• Google AI Overviews. Run the query in a signed-out window. Search Console does not separate AI Overview impressions from ordinary ones, so this has to be done by hand.
• Google AI Mode and Gemini. Note which sources it grounds on. If a competitor appears and you outrank them in the blue links, your problem is structure rather than authority.
• Claude. Watch whether it searches at all. Answering from memory without searching is a signal that your entity is weak.
• Microsoft Copilot. Follows Bing, which is a different ranking set from Google. A page sitting second on Google can be first here.

What to record

One row per run, with these columns: date, engine, exact prompt, search on or off, full answer text, cited sources if shown, and a verdict of accurate, stale, wrong, or absent.

Store the full text rather than a summary or a screenshot. In three months the argument will be about a specific sentence, and a paraphrase will not settle it. Our clients get every run stored automatically in the client portal so a run from three months ago can be reopened and read.

How to read what comes back

Sort every problem answer into one of four buckets, because each has a different fix.

• Stale. True once, not true now. The most common and the most fixable.
• Merged. Correct facts about the wrong entity. An identity problem rather than a content problem.
• Invented. A specific detail with no source anywhere. The fix is to close the gap that invited it.
• Absent. Not wrong about you, just does not mention you. Commercially this is often the most expensive, and it never triggers an alert because there is nothing to alert on.

How often to repeat it

Monthly is enough for most brands and matches how fast the engines move. Perplexity shifts within weeks once a better source is published. Anything held in training memory waits for a model release.

Keep the prompt set frozen between runs. If you rewrite the questions you cannot compare the answers, and comparison is the only thing that turns this from an anecdote into a measurement.

Frequently asked questions

Do I need a tool for this?

No. A spreadsheet and two hours cover the first baseline. Tools save time on repetition and on running the set across engines and languages at once, which matters from the second month onward.

Why does the answer change every time I ask?

Generation is not deterministic, and retrieval reaches different pages on different days. Judge the substance rather than the wording, and run each prompt more than once before deciding an answer is representative.

Should I use my own account or a clean session?

A clean session. Your own account has memory, history and location attached, which tells you what the engine says to you rather than what it says to a stranger who is about to buy.

What if the answers are fine?

Save them anyway and date them. A baseline taken while things are good is what lets you prove, later, that something changed and roughly when.

If you would rather not run it yourself

Our AI reputation management engagement starts with exactly this: an agreed prompt set, six engines, every answer stored with its date, and every problem answer traced back to the source that produced it. If you want the wider picture of how AI engines pick sources, our GEO service covers the visibility side of the same problem.

Related guides

• What AI Overviews Cite in Arabic vs English: A 16-Prompt Audit
• Do Trustpilot and G2 Reviews Affect What ChatGPT Says About You?
• Google’s Search Generative Experience (SGE) and AI Overviews: What Egyptian Brands Need to Know in 2026
• GEO: Generative Engine Optimization – Your Guide to Ranking on ChatGPT