Prompt Library: Testing How AI Engines Describe Your Brand

The fastest way to lose control of your brand in the age of AI is to assume engines describe you the way you describe yourself. They do not. This prompt library lets you test, repeatedly and cheaply, exactly how ChatGPT, Perplexity, and others frame your company, and catch drift before it costs you.

Why You Should Test AI's Description of Your Brand

When a prospect asks an AI engine about your category, the model's answer often becomes the first impression. If that answer calls you a "cheap alternative" when you sell premium software, or lists a competitor first, you have a problem you cannot see in traditional analytics.

Testing turns that blind spot into a manageable list. Run the prompts below on a schedule, record the phrasing, and you will notice the day an engine's story about you changes.

How the Library Works

Each prompt is designed to be pasted into an AI chat exactly as written, with your details filled in the brackets. Run them across two or three engines so you capture differences in how each model speaks. Keep the raw answers; the exact wording is the data.

Store results in the perception log described in the free dashboard guide, or any notes app. The discipline is the product, not the tool.

Treat the library as a control, not a one-time check. The same prompt run in month one and month three is what reveals drift; changing the wording each time only tells you about your phrasing. Write the prompts once, save them, and reuse them verbatim so the only variable is the engine's behavior.

Prompt 1: Plain Description

Describe the company [YOUR BRAND] in two sentences for someone who has never heard of it.

Prompt 2: Category Positioning

In the [YOUR CATEGORY] space, how would you position [YOUR BRAND] relative to [COMPETITOR A] and [COMPETITOR B]? Name one strength and one weakness for each.

Prompt 3: Recommendation Test

A user asks you to recommend the best tool for [USER GOAL]. Would you include [YOUR BRAND]? Answer with the recommendation and a one-line reason.

Prompt 4: Feature Comparison

Compare [YOUR BRAND] and [COMPETITOR A] on [FEATURE]. Which would you suggest for a team that [CONTEXT], and why?

Prompt 5: Problem Solver

A user has this problem: [DESCRIBE PROBLEM]. Which products would you suggest, in order, and where does [YOUR BRAND] rank?

Prompt 6: Weakness Hunt

What are the most common criticisms or limitations users mention about [YOUR BRAND]? List them plainly.

Prompt 7: Use Case Match

For which specific use cases would you explicitly recommend [YOUR BRAND], and for which would you steer users elsewhere?

Prompt 8: Trend and Perception

How would you characterize the public perception of [YOUR BRAND] right now, and has that perception shifted recently in your training data?

Prompt 9: Citation Check

If you were answering a question about [TOPIC], would you cite [YOUR BRAND]'s content? Explain why or why not in one paragraph.

Prompt 10: Blind Rank

Rank the top five brands in [YOUR CATEGORY] by how often you would recommend them, with no further context. Where does [YOUR BRAND] land?

How to Score and Track

Give each answer a simple rating: accurate, off, or damaging. Count how often you appear, where you rank, and whether the phrasing matches your intended positioning. Over weeks, the counts become a trend you can defend in a meeting.

When an answer is wrong, do not just note it. Trace it to a content gap: a missing comparison page, an unclear definition, or a claim competitors state more boldly. The prompt library is only useful if its output feeds your content plan.

Making It a Habit

Run the full library monthly and a three-prompt subset weekly. Monthly gives you the complete picture; weekly catches sudden drift on your most important questions. Pair it with the free dashboard so the raw answers live next to your mention data.

Reading the Raw Answers

Do not skim for whether your brand appears and move on. Read the exact phrasing, because "X is a low-cost option" and "X is the market leader" are both appearances with opposite value. Note adjectives, ranking order, and which competitor is named first; those details are the signal.

Pay attention to what the engine volunteers without prompting. If a recommendation test returns a competitor with no mention of you, that silence is data. It tells you the model has not connected your brand to that query, which is often a content coverage problem rather than a reputation one.

A Simple Scoring Table

Turn the raw answers into a trackable score with three columns: appears, rank, and tone. "Appears" is yes or no. "Rank" is the order in which you were named, since first mention carries the most weight. "Tone" is accurate, off, or damaging.

Over a few cycles, count how often each rating shows up per query. A query that moves from "off" to "accurate" is proof your content work is working. A query stuck on "damaging" is your next priority. The table is deliberately crude; the trend is what matters, not decimal precision.

When to Escalate

A single off answer is noise. A pattern across engines and cycles is a problem worth a meeting. Escalate when the same wrong description appears in two or more engines, when a competitor consistently outranks you on a high-intent query, or when tone turns damaging on a topic tied to revenue.

Escalation should produce a content task, not just anxiety. Assign the gap to an owner, publish the fix, and re-test the same prompts in the next cycle to confirm the description moved. Closing the loop is what separates testing from complaining.

Building Your Own Prompts

The ten prompts above are a starting set, not a ceiling. As you learn how engines talk about you, write your own that target the exact moments that matter, such as a pricing question or a security question a buyer might ask. The formula is constant: state the scenario, name your brand and competitor, and ask for a concrete, ranked answer you can score.

Keep new prompts in the same library so old and new results stay comparable. A prompt you change every week measures nothing; a prompt you keep measures movement.

Frequently Asked Questions

How many engines should I test against?

Test at least two with different underlying models, such as one OpenAI-based engine and one that uses a different source mix. That spread shows you whether a problem is model-specific or universal.

What do I do when an engine describes us wrongly?

Trace the wrong answer to a content gap, then close it. Publish a clearer comparison, definition, or claim, and re-test the same prompt in a later cycle to confirm the description moved.

How is this different from normal brand monitoring?

Normal monitoring watches what people write. This watches what models say when asked, which is the answer a growing share of buyers now hear first. Both matter, but they measure different things.