The prompts every brand should be tracking

Most brands track a handful of vanity prompts: their own name, their tagline, maybe a feature. Those tell you nothing. The prompts that move revenue are the ones a buyer types when they have a problem and no loyalty yet.

A vanity prompt is one the buyer would only type if they already know you — "what is AnswerPeek" tells you how a model describes you, which is worth checking once, but it never wins you a new customer. A revenue prompt is one a stranger types on the way to a decision. The difference sounds obvious and is ignored constantly, because vanity prompts flatter and revenue prompts sting.

Five patterns that decide the sale

• Best X for Y: "best project management tool for agencies". The shortlist prompt. This is where you are named or you are not.
• X vs Y: "Notion vs Coda". The model picks a side, or hedges. Either way you want to be in the comparison.
• Alternatives to X: "alternatives to HubSpot". If a competitor owns the category, this is how you get into the room.
• How much does X cost: pricing prompts pull structured data and reviews. Missing pricing schema means missing the answer.
• Is X good for Z: use-case prompts where a model decides if you fit a specific buyer.

Each pattern fails differently, which is why you track them separately. Losing a "best X for Y" prompt usually means the roundups the engine reads do not include you. Losing an "X vs Y" prompt usually means your comparison story is weaker or staler than the competitor's. Losing a pricing prompt is almost always missing or malformed structured data on your own pricing page. The pattern tells you where to look before you have read a single answer.

How many prompts is enough

Ten to thirty, tracked consistently, beats two hundred tracked once. The goal is a panel: the same prompts, the same engines, run on a schedule, so that a change in the answers means something changed in the world rather than in your methodology. Past thirty prompts, reading the answers properly stops happening, and unread answers are the same as untracked ones.

Phrasings are not interchangeable

"Best CRM for startups", "top CRMs for a small startup" and "which CRM should an early-stage startup use" look identical to a human and can produce three different shortlists. Models are sensitive to phrasing in ways rank trackers never had to worry about. For your most important two or three prompts, track a couple of phrasing variants; where the shortlists diverge, read the cited sources for each variant — the difference is usually one roundup that one phrasing surfaces and the other does not.

How often to re-run

Weekly is the honest default. Answers drift constantly: engines re-retrieve, sources update, and model versions change without announcement. Monthly is enough for slow categories; daily is noise for almost everyone. What matters more than frequency is keeping the panel fixed, because a moving prompt set cannot show a trend.

What to log from each answer

• Named or not — the binary that everything else hangs off.
• Position in the shortlist, when there is one; being the first name is not the same as being the fifth.
• The exact sentence about you. Wrong descriptions are their own category of problem.
• Which sources the engine cited, because those pages are your target list for outreach.
• Which competitors were named, so you can compute share of voice across the panel over time.

Build the list in an afternoon

Read your own sales calls and support tickets. The questions buyers ask a human are the questions they now ask a model. Turn each into the patterns above, run them, and you have a tracking list that maps to real demand instead of guesses.

Questions, answered

How many prompts should a brand track?

Ten to thirty. Enough to cover the five buying patterns — best-for, versus, alternatives, pricing, and use-case prompts — across your main segments, but few enough that you actually read every answer. A small panel tracked weekly beats a huge list run once.

How often do AI answers change?

Constantly. Engines re-retrieve live sources, the sources themselves update, and model versions change without notice. Week to week you should expect some churn in shortlists, which is exactly why one-off checks are misleading and a fixed panel on a schedule is the only trustworthy setup.

Should I track branded prompts at all?

Track one or two, like "what is [brand]" and "is [brand] legit", to catch wrong or stale descriptions. But the bulk of the panel should be unbranded buying prompts, because those are the ones that decide whether new buyers ever hear your name.

Do I need to track every AI engine?

No. Start with ChatGPT and Perplexity — the reach leader and the most transparent citer. Expand only when you are consistently acting on what those two tell you. An untracked engine is a smaller loss than an unread answer.