# How to choose the prompts to track in ChatGPT, Gemini, and Perplexity > Track prompts in topic clusters that cover your category, your differentiators, and your buyers' jobs, in all four forms people type. Never track your own brand name; that is a vanity metric. Around forty-five candidates and twenty-five tracked is a good start. - Published: 2026-09-05 - Author: Selfwin team - Canonical: https://selfwin.co/blog/how-to-choose-prompts-to-track - Tags: ai-visibility, prompts, measurement ## Key takeaways - Prompts should be organized under five to seven topics that each cover a distinct reason a buyer would want you. - Every topic needs all four forms: keyword fragments, questions, imperative asks, and constraint-laden long-tail prompts. - At least half of your prompts should hinge on a real differentiator, because those are the answers you can actually win. - Never track prompts that contain your own brand name. You already win those, and they hide the losses that matter. The prompt list is the instrument. Choose it badly and every number downstream is wrong, however pretty the chart. Here is how to build a set that measures something real. ## Start from what the product actually is Before writing a single prompt, write down the brief: the category in the buyer's words, the platforms you run on, the three to eight concrete capabilities that set you apart, who buys, what they hire you for, how you charge, and who you get compared to. Every prompt should trace back to a line in that brief. If it cannot, it is a generic prompt and it will measure the category, not you. Selfwin builds this brief by reading your website, then asks you to correct it before it writes anything. The correction step matters: models are confident and websites are vague. ## Organize prompts into topics A flat list of thirty prompts is hard to reason about. Five to seven topics are not. A good topic set covers: 1. the category head term, for example "AI iOS app builders"; 2. two or three topics built on your differentiators and platforms, such as "native Swift code export" or "App Store submission"; 3. one topic on the main audience's job to be done, in their words; 4. optionally, pricing or alternatives. Topics must not overlap in meaning. If two topics would attract the same prompt, merge them. ## Write every topic in four forms People do not type one way. Inside each topic, include: - **Keyword fragments.** `ai app builder ios`. Short, lowercase, no punctuation. - **Questions.** "What is the best AI app builder for iOS?" - **Imperative asks.** "Recommend an app builder that exports Swift code." - **Constraints.** "I need to build an iPhone app without coding, publish it to the App Store, and keep the source. What should I use?" Engines answer each form differently, and the constraint form is where the winnable prompts live, because the answer names two or three brands instead of ten. ## Weight toward differentiators At least half of the set should hinge on something specific to you. Head-term prompts are worth tracking because they show the category, but they are also the hardest to win and the slowest to move. Differentiator prompts move first, and a win there is proof that the engines have learned what you are for. ## Leave your own brand out Prompts containing your brand name are answered from your own website. You will win them, the number will look great, and it will tell you nothing. Track them separately if you want a reputation read, but keep them out of the visibility score. Competitor names are different. A few comparison prompts, under a fifth of the set, show where buyers weigh you against others and which alternatives the engines reach for. ## Deduplicate ruthlessly "ai app builder" and "ai app builders" are the same prompt to an engine. Near-duplicates inflate your prompt count and double-weight one question. Selfwin drops candidates whose normalized tokens overlap another prompt by more than eighty percent, both within a batch and against everything you already track. ## Add search volume when you can Volume is a rough proxy, because AI prompts are not search queries, but a keyword-sized prompt with real monthly searches is a better bet than one nobody types. Where a volume source is configured, Selfwin attaches it to the shorter prompts so you can prefer the ones people ask. ## A reasonable starting size Generate around forty-five candidates, review them, and track about twenty-five. Spread the selection across topics so no cluster is skipped, prefer differentiator and long-tail prompts, then commercial intent, then volume. Revisit monthly: retire prompts that never vary and add prompts for topics where you have started to appear. ## FAQ ### How many prompts should I track for AI visibility? Start with twenty to twenty-five tracked prompts across five to seven topics. That is enough for a trend in two weeks without drowning in noise, and you can add more once you see which topics move. ### Should I track prompts that mention my brand? No. Prompts like 'is Acme good' are answered with your own website, so you win them by default. They inflate your visibility number and hide the category questions where you are absent. ### Should prompts name competitors? A few should. Comparison and alternative prompts show where buyers weigh you against others. Keep them under a fifth of the set so competitor names do not dominate what you measure. ### What is a long-tail prompt in AI search? A long-tail prompt adds constraints a real buyer has: a platform, a budget, a team size, an integration. Engines answer them with fewer brands, so they are the easiest to win and the most honest signal.