Prompt Set Design for Visibility Tracking

A prompt set is the list of questions you ask AI answer engines, such as ChatGPT, Perplexity and Google AI Mode, every time you measure your brand. A prompt is simply the text a buyer types into the engine. The set stays the same from one round to the next, so the numbers can be compared over time.

The set is the measuring tool. Mention rate, citation rate, position and share of voice are all counted over it, so a badly built set gives clean-looking numbers that mean little. This page shows how to design the set. The parent guide, how to track brand mentions in ChatGPT and Perplexity, covers the full tracking method, and the AI visibility metrics that matter explains how to read the results.

The questions you pick decide most of the answer. In Leadee's tracker round of 2026-10-25, the seven websites of Netbase JSC, the company behind Leadee, were named in 62 of 84 US answers to questions that contained their brand name (73.8%, 95% interval 63.5-82.0%). For questions without the name, the count was 20 of 274 (7.3%, 4.8-11.0%). A set made mostly of branded questions would have reported a rate about ten times higher for the same brands on the same day.

Question shape matters too, though less clearly. Among the unbranded questions, "best" lists named a brand in 9 of 98 runs, comparisons and alternatives in 8 of 98, how-to questions in 3 of 60 and other shapes in 0 of 18. The intervals overlap, so we cannot rank the shapes. We can say that a set built from one shape only shows one slice of how buyers ask.

Write one sentence before you write any prompt: which brand, which product or service line, which market and which buyer. "CMSmart, WooCommerce product designer plugins, US store owners" is a scope. "Our visibility in AI" is not. One set should answer one question about one offer, so you know what to change when the numbers move.

If you sell several things to different buyers, build one set per line rather than one big mixed list. CMSmart (cmsmart.net, part of Netbase JSC) shows why: in the 2026-10-25 round it was named in 24 of 52 US buyer questions overall and in 13 of 40 unbranded ones, all asked about its own plugin niche. Mixing those questions with an unrelated product line would blur both numbers.

Sort candidate questions into shapes and keep a few of each. The common shapes are "best [product] for [need]", "[option A] vs [option B]", "alternatives to [well-known brand]", "how do I [job]" and "who supplies [thing] in [country]". A comparison question that names a well-known market player is fair game: it is how buyers ask, and the answer shows whether you sit next to that brand.

Spread the set across the subtopics of your offer, too. Google's AI features guide says AI Overviews and AI Mode may use a "query fan-out" technique, issuing several related searches across subtopics to build one answer. A set that only asks about price, for example, misses the subtopics the engine itself searches. To collect candidates, follow how to find the prompts your customers ask ChatGPT or try the free ChatGPT Prompt Finder.

A branded question contains your name, such as "is [brand] reliable" or "[brand] pricing". An unbranded question does not. Unbranded questions show whether new buyers hear about you; branded ones show whether the engines describe you correctly. Both are useful, but they answer different questions, so tag every prompt as branded or unbranded and report the two rates on separate lines.

A practical mix is 3 to 5 branded questions and the rest unbranded. Use the branded ones to check facts: price, location, what you sell and who you are not. That last point matters when another company has a similar name. On 2026-10-25 most raw "AdminKit" mentions were about an unrelated website template, not AdminKit.net (part of Netbase JSC). Adding a short qualifier to the prompt, such as the product type, keeps the answers about the right company.

Start with 15 to 30 questions per scope and ask each one at least three times on every engine in every round. AI answers change from run to run, so one run per question is a sample of one. A 95% confidence interval, the range the true rate probably sits in, shows how much a small set can tell you. Leadee uses the Wilson score interval from the NIST/SEMATECH statistics handbook: at an observed 30%, 10 runs give 10.8% to 60.3% and 100 runs give 21.9% to 39.6% (arithmetic examples, not measurements).

Unbranded questions need the repeats most. In the 2026-10-25 round, when a brand was named in at least one of three runs, a single run picked at random missed it in 15 of 36 unbranded runs (41.7%, 95% interval 27.1-57.8%) but in only 7 of 69 branded runs (10.1%, 5.0-19.5%). So 20 unbranded questions times three runs gives 60 answers per engine and market, enough to see a large change. Our study of manual and automated tracking has the full count.

A buyer in the US and a buyer in the UK can get different answers to the same words. Keep each country as its own set, or at least its own line in the report, and never pool markets into one rate. In the 2026-10-25 round, UK, German and Singapore questions together produced only 4 mentions across all seven Netbase JSC brands, while the US questions produced far more. A pooled rate would have hidden both facts.

Set the location in the request, not in the wording. OpenAI's web search guide lets you pass an approximate user location: country as a two-letter ISO code such as US, city and region as free text, and a timezone such as America/Chicago. Keep that setting the same for every run of a set, and write it down next to the set so the next round uses it too.

Use the words buyers use, not your marketing words. Your own search data helps here: the Search Console API returns clicks, impressions and position grouped by query, page and country, so the questions people already type into Google can become prompts. Sales emails, support tickets and site search logs are just as good.

Then clean every prompt with a few rules. Ask one thing per prompt. Do not lead the engine, as in "why is [brand] the best". Keep the length and tone of a real person, not a keyword string. Do not add instructions such as "answer in a table", because a buyer would not. Write down the exact text, including spelling, because small changes in wording can change the answer.

Once the set is built, lock it for the period you report on, such as a month. Give it a version number and a date. If you need to add, drop or reword a question, do it between periods, start a new version and note what changed. A rate from version 2 is only comparable with version 1 for the questions both versions share.

Review the set every quarter. Drop questions that no engine ever answers with any brand, add questions for new products or new buyer worries, and check that the branded share is still small. Keep the set in the same sheet as your results: our ChatGPT prompt tracking sheet has columns for the prompt and its shape, with branded as one of the shapes.

Leadee is an AI-native platform that stands on the shoulders of giants: it asks ChatGPT, Gemini, Perplexity and Google AI Mode through their official APIs and never scrapes a consumer chat app. In the Visibility Tracker, each prompt set is a named list of buyer questions with its market and engines. Leadee can suggest questions for your category, and nothing enters the set until your team accepts it.

The tracker runs the set on a fixed schedule, repeats each question, and reports mention rate, citation rate, position and share of voice with a 95% Wilson interval. It tracks other market players with the same questions and method. Every other ChatGPT guide is on our ChatGPT SEO hub.