How to Check If ChatGPT Recommends Your Brand

Before you can improve how AI assistants talk about your brand, you need an honest baseline. Here is a method you can run by hand in an afternoon, what to record, and how to turn it into something repeatable instead of a one-off screenshot.

First, why "do you know my brand" does not work

The instinct is to open ChatGPT and ask if it knows you. Do not. Two things make that test lie to you. Buyers never phrase it that way, so you are measuring a prompt no real customer uses. And assistants lean agreeable: ask leadingly and you will often get a yes that evaporates the moment the question is neutral. You want the unflattering, realistic version.

1. Write prompts the way buyers do

Real buyers ask about the problem, not about you. Build a set around three shapes:

• Category questions: "What is the best [category] tool for [specific use case]?"
• Competitor comparisons: "Alternatives to [a well-known competitor]?" or "[Competitor] vs what else should I consider?"
• Job-to-be-done: "I need to [outcome]. What should I use?"

Aim for 10 to 20 prompts covering your top use cases and your main rivals. Write them once and freeze them. This fixed set is the instrument; changing it later makes results incomparable.

2. Run each prompt across multiple assistants

Run the full set on ChatGPT, Claude, Gemini, and (if relevant to your buyers) Perplexity. They do not agree. It is normal to be named reliably by one and entirely absent from another, which is exactly why a one-model check is misleading. Use a fresh chat per prompt so prior turns do not contaminate the answer.

3. Record four things per answer

For every response, capture:

1. Mentioned? Were you named at all (yes/no).
2. Position. First name listed, mid-list, or an afterthought.
3. Sentiment. Recommended, neutrally listed, or hedged/negative.
4. Competitor set. Which other brands appeared, and where you sat relative to them.

The competitor column is the one people skip and the one that teaches you the most.

4. Score it

Your headline number is unbranded mention rate: the share of prompts, per assistant, where you were named. Compute it per model and overall. Resist the urge to grade it against an imaginary benchmark. The only comparisons that mean anything are against the competitors the same prompts surfaced, and against your own number next time. A 0% baseline is common and is not a failure; it is the starting line.

5. Read the misses, not just the score

Where you were absent, write down what the model recommended instead and, where it explains itself, why. Patterns surface fast: a competitor owns the comparison article the model leans on, a directory everyone else is in and you are not, or a capability you have that the model simply does not know about because nothing credible says so. The misses are the to-do list.

6. Make it repeatable

Model outputs drift with every retrain and, for retrieval-backed assistants, every re-crawl. A single audit ages within weeks. Re-run the identical prompt set on a schedule (monthly is a reasonable floor) so you can tell whether a change you made actually moved the number, rather than guessing.

Mistakes that quietly ruin the test

• Leading prompts. Mentioning your own brand in the question. You are testing recall, not reading comprehension.
• A drifting prompt set. Editing prompts between runs. Now you cannot attribute changes to your work.
• One model only. Declaring victory on ChatGPT while Gemini never names you.
• Scoring once. Treating a photograph as a trend.

Doing it at scale

The manual version above is a genuinely good first pass, and you should run it at least once by hand so you trust the method. Past that point it becomes tedious and easy to do inconsistently, which is the problem GEO tooling exists to solve: a fixed prompt set, scheduled multi-model runs, mention and competitor tracking, and a prioritized list of the gaps worth closing first. Whether you automate it or not, the discipline is the same: fixed prompts, multiple models, honest scoring, repeated on a schedule.