AI SEARCH (GEO)

Test how an assistant currently describes your brand

Use when you want to see what an AI assistant says about you before changing anything.

brand-description-test.md

Download .md

Answer the following as you normally would for a user, using only what you already know plus any sources you can retrieve. Do not consult material I paste.

Brand: {{BRAND_NAME}}
Sector: {{SECTOR}}

Answer each of these separately and label them:
1. What is {{BRAND_NAME}}? Answer in 50 words.
2. Who is it for, and who is it not for?
3. What does it cost, as far as you can tell?
4. Who are its main alternatives, and how does it differ from each?
5. What would you warn a prospective customer about?
6. Which sources are you drawing on for each of the above? Name them.

Then add a final section headed CONFIDENCE. For each of the six answers, state high, medium or low confidence and say exactly which part you are unsure about.

Constraints:
- If you do not know something, say "I do not know" rather than producing a plausible-sounding answer.
- Do not soften negative perceptions to be polite. Report them as you would to the user.
- Flag anything you suspect may be out of date and say roughly how old you think it is.

Fill in before running

Replace each placeholder with your own detail. The more specific you are, the less the model invents.

{{BRAND_NAME}}
{{SECTOR}}

Getting a better result

Run the same prompt in a fresh session on each assistant you care about and diff the answers.
Question 6 is the important one; it tells you which pages to fix.
Repeat monthly and keep the outputs, so you can see drift rather than guess at it.

Questions about this prompt

When should I run this instead of auditing my own pages?

Before you change anything. It is a baseline measurement, not a fix. Auditing your own pages tells you what you published. This tells you what an assistant says when nobody is watching, which is often assembled from a review site or an old profile rather than from your site at all.

What do I need in front of me?

Almost nothing, and that is deliberate. The prompt instructs the model to ignore material you paste, so supply only {{BRAND_NAME}} and the sector. What you do need is a clean session with no earlier discussion of you in it, because anything above the prompt contaminates the answer you get.

What comes back, and which answer matters most?

Six labelled answers covering what you are, who you are for, price, alternatives and warnings, then a confidence rating for each. Read question six first. The sources it names are the pages actually feeding the description, and they are frequently not the pages you have been spending your time editing.

What is the mistake that costs me?

Treating a single run as a verdict. Answers vary between sessions and between assistants, so one unflattering paragraph proves little on its own. Run it on each assistant you care about, keep the outputs dated, and compare. Fixing the wording in answer one while ignoring answer six wastes the exercise.