How to Build a Prompt Set for AI Visibility Tracking

Prompt tracking means running a fixed, documented set of questions through AI engines on a schedule and recording whether your brand is mentioned, cited and recommended in the answers. The prompt set is the foundation: it should mirror the questions real buyers ask at each stage of their journey, in their own words, locations and languages. Build it from your keyword data, sales calls, support tickets and People Also Ask, split it into branded and unbranded prompts, and keep a stable core so trends stay comparable.

This guide covers how to design that set. If you are still picking software, read how to choose an AI visibility tool first; the prompt allowance and run frequency you buy will shape everything below.

Why the prompt set decides your AI visibility numbers

Every metric an AI visibility report shows, from mention rate to AI share of voice, is calculated over the prompts you chose. Pick prompts your buyers never ask and you will measure an imaginary market. Pick only prompts that contain your brand name and you will look dominant while competitors win every unbranded recommendation.

Prompts also differ from keywords in a way that matters for sampling. In the SparkToro and Gumshoe study of AI recommendations, 142 people asked to write a prompt for the same headphone need produced prompts with a semantic similarity of only 0.081, according to Search Engine Journal’s report. Real users phrase the same need in very different ways, so a prompt set needs several phrasings of each important need rather than one “head term”.

Prompt types by funnel stage

Organize prompts by the job the buyer is trying to do. Aleyda Solis’s guide to building a representative prompt library (updated June 2026) segments prompts by journey stage, product line, persona, market and business priority. A simpler working version uses five stages:

The shares are a starting point, not a benchmark. A brand chasing new demand should weight discovery prompts most heavily; a brand worried about reputation should add more validation prompts.

Where to find prompt research inputs

Good prompt research starts from language your audience already uses. These five sources cover most of it:

1. Keyword data. Take your highest-value non-branded keywords and long-tail Search Console queries, then rewrite them as the full questions a person would type into a chatbot. “CRM small business” becomes “What is the best CRM for a five-person sales team that uses Gmail?”
2. Sales calls and CRM notes. Prospects state their constraints out loud: budget, team size, integrations, compliance. Those constraints become the “for Y” part of your discovery prompts. Anonymize anything personal before it goes into a tool.
3. Support tickets and live chat. These feed the post-purchase and validation stages, and they reveal the competitor names customers mention.
4. People Also Ask. PAA questions are already phrased as questions and tied to search demand. Some trackers map prompts to PAA and keyword volume for that reason.
5. Communities and reviews. Reddit threads, forums and review sites show how people phrase comparisons and complaints.

Once you have a long list, check which phrasings the engines expand into sub-searches. Our explainer on query fan-out shows how one prompt becomes several searches, which helps you spot topics your prompts are missing.

Branded vs unbranded prompts

Unbranded prompts (“best accounting software for freelancers”) measure whether you are discovered. Branded prompts (“is [brand] good for freelancers?”) measure whether engines describe you correctly. Competitor prompts (“[competitor] alternatives”) measure positioning. You need all three, but keep them in separate segments and never blend them into one headline number.

A practical split for most brands is about 70% unbranded, 20% competitor and alternatives, and 10% branded. If a single branded prompt is enough to show the engine knows you, more of them just inflates your mention rate.

Locations and languages

AI answers can change with location and language. Google’s Liz Reid has said AI Mode uses Google’s ranking systems to ground responses by location, so the same question asked from Chicago and from Manchester can return different businesses. Peec AI’s analysis of 20 million ChatGPT fan-outs found that 43% of fan-out steps for non-English prompts ran in English, as reported by Search Engine Journal in February 2026. In other words, a Spanish or German prompt may still be answered partly from English sources.

So treat each market as its own prompt set: written natively in the local language (not machine-translated), run from that country, and reported separately. For local businesses, set the city on prompts where location changes the answer, such as “best dentist near me” or “emergency plumber open now”.

How many prompts to track

Start smaller than you think. A useful rule of thumb from Solis’s guide is 30 to 60 prompts for a single product in one market, 100 to 250 for SaaS or services with several personas and competitors, and 30 to 100 per additional priority market. These are starting points she explicitly labels as non-benchmarks.

Budget the repeats too. Because answers vary from run to run, each prompt needs several runs per measurement period before its mention rate means much. Here is the arithmetic for a mid-sized set:

answers per month = prompts × engines × runs per week × 4

With 60 prompts, 4 engines and 3 runs a week, that is 60 × 4 × 3 × 4 = 2,880 answers a month. If your plan only allows 1,000, cut prompts before you cut repeats: 20 well-chosen prompts run three times are more trustworthy than 60 run once. Our guide on AI visibility metrics explains which figures need the most observations.

An example prompt set for a B2B SaaS brand

Below is a starter set for a fictional project management tool, “Taskline”, selling to agencies in the US and UK. Each prompt carries an ID and tags so you can filter reports later.

Notice that most prompts include a constraint (team size, integration, client approvals). Constraints make answers more specific and closer to how buyers ask, and they show you which attributes engines associate with your brand.

How to track AI visibility prompts over time

A prompt set is never finished, but it must not change so often that trends break. Use three layers:

• Core set (about 70%). Fixed prompts you never edit. Changing a prompt’s wording starts a new time series, so retire it and add a new ID instead.
• Rotating set (about 20%). Seasonal topics, new features, campaign themes. Review monthly.
• Experiment set (about 10%). New phrasings or markets you are testing. Promote the useful ones to the core set each quarter.

Review the full library every six to twelve months, and immediately after a pricing change, product launch or rebrand. When you add a competitor, add it to the existing prompts’ brand list rather than writing new prompts for it.

AI prompt tracking tools handle these layers differently. Asheard, for example, lets you set a location per prompt and compare competitors across ten AI engines, which suits a tagged, multi-market set like the one above; see the AI visibility tracking page for details. Whatever tool you use, export the prompt list with its IDs and tags so the set lives outside any one vendor.

Common prompt set mistakes

• Writing prompts like keywords. “CRM pricing” is a search query, not a chatbot question. Write full sentences with context.
• Over-weighting branded prompts. They inflate mention rates and hide discovery gaps.
• One phrasing per need. Given how differently people phrase the same need, add two or three variants for your most important topics.
• Machine-translating prompts for other markets. Local buyers use local terms, brands and constraints.
• Editing prompts in place. You lose comparability. Retire and replace instead.

Frequently asked questions

What is prompt tracking in AI search?

Prompt tracking is the practice of running a fixed set of questions through AI engines such as ChatGPT, Gemini and Perplexity on a schedule, then recording whether your brand is mentioned, cited and recommended. It is the AI equivalent of keyword rank tracking, except each prompt needs several runs because answers vary from one run to the next.

How many prompts should I track for AI visibility?

A single product in one market can start with 30 to 60 prompts. SaaS or service brands with several personas and competitors often need 100 to 250, plus 30 to 100 for each extra priority market. If your plan limits answers, keep fewer prompts and run each one several times rather than tracking many prompts once.

Should AI tracking prompts include my brand name?

Only a minority of them. Prompts that name your brand check whether engines describe you accurately, but they almost always mention you, which inflates your numbers. Most of the set should be unbranded discovery and comparison prompts, because those show whether engines recommend you to buyers who have not heard of you yet.

How often should I update my prompt set?

Keep a stable core of prompts that never changes so trends stay comparable. Review a smaller rotating group monthly for seasonal or campaign topics, test new phrasings in an experiment group, and review the whole library every six to twelve months or after a major product, pricing or brand change.