How to write a great prompt for LLM tracking

A prompt is the unit of measurement for AI visibility. Write them well and your data becomes a strategy. Write them lazily and you'll track noise. Here is the long version of how we think about it.

A well-formed prompt is audience plus intent plus constraint plus market.

The prompt is the unit of measurement

In classic SEO, the keyword was the atomic unit. You picked one, you ranked for it, you measured the click. In LLM tracking, the equivalent atomic unit is the prompt, the actual sentence a buyer types into ChatGPT, Claude, Perplexity or Gemini.

The shift looks small. It is not. A keyword is one or two words optimized for a search engine. A prompt is a full sentence written by a human who already expects a complete answer. The grammar is different, the intent is denser, and the model uses every word you give it to decide who to mention.

That means the quality of your prompt set decides the quality of your visibility data. A bad prompt set will tell you that you are dominating a category nobody is searching for. A good one will show you, week over week, exactly where you are winning and losing the conversations that drive revenue.

The four parts of a great tracking prompt

Every prompt that produces useful data has the same four ingredients. Miss one and the answers drift into noise.

• Audience: who is asking. A CMO and an SEO manager phrase the same question differently. Bake the persona into the wording.
• Intent: what they want. Compare, choose, learn, troubleshoot. The verb at the front of the prompt shapes the entire answer.
• Constraint: what narrows it. Budget, region, integration, company size. Constraints are how a generic question becomes a buying question.
• Market: where it applies. Geography, language, vertical. The same prompt in Danish vs. English will surface different brands.

What separates a vague prompt from a useful one

The fastest way to feel the difference is to compare. Each pair below is the same underlying business question, one written lazily, one written with intent.

• Too vague: "best CRM". Tracking-grade: "best CRM for a 20-person B2B SaaS startup that needs HubSpot-style automation but at a lower price point".
• Too vague: "AI SEO tool". Tracking-grade: "AI visibility platform that tracks brand mentions in ChatGPT and Perplexity for European martech companies".
• Too vague: "alternative to Ahrefs". Tracking-grade: "Ahrefs alternative for a small in-house SEO team that also needs LLM citation tracking".

Build a prompt set that mirrors the funnel

A single prompt is a data point. A funnel of prompts is a strategy. We recommend building your tracking set across four stages, and roughly weighting it the way real buyer journeys do.

• Awareness, about 40% of the set: "what is AI visibility and why does it matter for B2B marketing".
• Consideration, about 30%: "how do I track whether my brand is mentioned in ChatGPT answers".
• Comparison, about 20%: "Serpier vs Profound vs Peec for European B2B SaaS".
• Decision, about 10%: "best AI visibility platform with native Perplexity tracking and EU data residency".

The weighting is a starting point. A long-cycle enterprise product can shift more toward consideration and comparison. A self-serve product can pull harder on decision.

The 8-point quality check

Before you commit a prompt to your tracked set, run it through this list. If it fails on more than two, rewrite it.

• It sounds like a sentence a human would actually type, not a keyword.
• It contains at least one constraint (budget, region, integration, size, vertical).
• It has a clear verb of intent (compare, choose, find, troubleshoot, evaluate).
• It is specific enough that two competitors would not both legitimately be the only answer.
• It is generic enough that more than one of your customers would actually ask it.
• It is written in the language your buyer uses, not the language your product team uses.
• It does not name your brand. Tracking prompts measure unbiased recall.
• It will still be relevant in six months. Avoid prompts tied to one news cycle.

Five mistakes that quietly ruin a prompt set

• Stuffing brand names. If every prompt mentions you, you are measuring search for yourself, not discovery.
• Tracking too few prompts. Below about 200 prompts the data is too noisy to spot weekly movement with confidence.
• Tracking too many prompts. Above about 2,000, signal drowns in maintenance. Most teams should live between 300 and 800.
• One language, multiple markets. English prompts will not measure your German visibility. Translate, do not duplicate.
• Set it and forget it. Buyer language drifts. Audit your prompt set every quarter. Retire dead ones, add new intent.
• Mixing intent levels in one report. Always segment results by funnel stage. A 90% awareness score and 10% decision score mean very different things.

A repeatable way to build your first 300 prompts

You do not need a platform to start. You need an afternoon and a spreadsheet. Here is the exact sequence we walk new customers through on day one.

• List your top 10 buyer questions from sales calls. Real language from real calls beats anything you brainstorm at a whiteboard.
• Expand each into 5 variants across funnel stages. Same underlying need, different awareness levels. That gives you 50 high-signal prompts to start.
• Add 50 competitor-comparison prompts. Pairwise: you vs. each direct competitor. These are the highest-leverage prompts you will ever track.
• Add 100 category-defining prompts that never mention any brand. These measure pure category share-of-voice.
• Add 100 long-tail use-case prompts. Specific industries, specific integrations, specific job titles.
• Run them across 3 models. Repeat weekly. ChatGPT, Perplexity, Gemini at minimum. Same prompts, same day, same time.

A prompt set is a living artifact. Treat it like product code, not like a one-time research project.

The teams who win in the LLM era are not the ones with the most prompts. They are the ones whose prompts most faithfully reflect what their customers are actually asking, this quarter, in their actual language. Build that, maintain it, and the visibility data will earn its place in every executive meeting you walk into.