Writing tracking prompts for AI visibility: five question types and the rules

A tracking prompt is the instrument. Change the wording and you have changed what is being measured, so the choice of prompts decides whether last month's number and this month's number can be compared at all. A prompt set that is reworded between rounds cannot be compared with the earlier rounds.

This post sets out the five prompt types the Dopbot Index uses, shows the real wording on its category pages, reports how each type behaved in the October 2026 edition, explains what changes in November, and ends with rules you can apply to your own set. One caution first: we wrote these prompts, and we did not observe shoppers or their searches. They are a consistent yardstick, not a record of how people phrase things.

What did each question type return in October 2026?

Every category has five prompts, one per type. In the October 2026 edition the fifth type was an alternative to a named brand; from November 2026, in the 40 updated US categories, it becomes a buying-advice question that names no brand. The October edition asked each prompt once per engine: OpenAI's API with web search (gpt-6-luna) and Perplexity's Agent API with the Sonar model (perplexity/sonar). These are API answers, not what the apps show. The table shows the October answers, OpenAI and Perplexity together, because this measure is not split by engine.

Table: October 2026 edition, all 160 categories, OpenAI and Perplexity answers together. Brands are those an answer recommends. The buying-advice type starts in November, so it has no October row.

Three things stand out. The plain "best" question was the one most likely to recommend no brand at all, 9.1% of its 318 answers, and the data cannot say why. Budget questions almost never recommended none, and 93.4% of their answers recommended one to three brands. The alternative question stands apart on the mean and on the share of answers with 1 to 3 brands, and its story has its own section below.

What does the real wording look like?

These are the prompts for water bottles, which is one of the 40 US categories updated from November. Its category page shows the October wording until the November update is published, and the November wording after that. The October column is the wording as asked in October 2026. The pages of the frozen US, Mexico and Canada categories keep the October wording.

Table: one category's five prompts, with the October 2026 wording as published in October 2026 and the wording set for the November 2026 update. From November, the 40 US categories still updated monthly use the November column; the other 60 US categories and all Mexico and Canada categories keep October's.

The pattern repeats elsewhere. For mattresses, the October premium prompt asked about a luxury mattress worth spending $2,500 or more on; the November one asks about one worth paying more for. The October fifth was an alternative to a named mattress brand; the November fifth is a back-pain scenario: "My lower back aches most mornings and I mostly sleep on my back. Which mattresses would you recommend for me?" The budget prompt, best queen mattress under $800, stays the same.

Markets use their own wording. Mexico's prompts are in Spanish with peso prices, for example "mejor colchón matrimonial por menos de $6,000 pesos" on the Mexican mattress page. Canada's use Canadian dollars, as in "best queen mattress under $1,000 CAD" on the Canadian mattress page.

How do you write the buying-advice question?

It describes a situation and asks for a recommendation, with no brand and no retailer. A situation has a person, a need and a constraint:

• "I'm training a new puppy that will grow fast. Which dog harnesses adjust well as the puppy gets bigger and stay comfortable?"
• "I wake up with a stiff neck and I sleep on my stomach. Which pillows would you recommend for stomach sleepers?"
• "I'm learning to cook and want one good knife I can sharpen at home. Which chef's knives would you recommend for a beginner?"

These come from the index's definitions of its active categories. The point of the type is to ask in the plain language of a buying situation, and to leave the choice of brand entirely to the assistant.

What changes in November, and why?

Four things, all recorded in the methodology.

1. The fifth prompt changes from an alternative to a named brand to a buying-advice question that names no brand.
2. The premium prompt changes from a wording that mostly carried a price floor to one that asks what is worth paying more for, with no floor.
3. Each prompt is asked 4 times per engine, which makes 40 answers per category instead of 10, for the 40 US categories still updated.
4. One carry-on luggage prompt is replaced. It cited the size limits of a particular airline.

The methodology's reason for the first change is that a category with 10 answers cannot support a rank order, and that the alternative prompt named a brand in every category, which skewed each category's counts. The data behind that reason is next.

What did the alternative-to-a-brand question teach us?

It named a brand and asked for alternatives, so it measured something different from the other four; the data cannot separate the two. Those answers recommended 5.0 brands each, against 2.3 for the other four types together (1,271 answers, excluding the alternative question). The brand named in the question was itself recommended in only 21 of the 296 answers where we could match it, or 7.1%.

If your tracking prompt is "alternatives to [your brand]", you are measuring which brands an assistant offers instead of yours. In October the brand named in the index's question was itself recommended in only 21 of 296 answers, which is what such a question asks for, so it cannot be your visibility number. It can be a useful study of your competitors. The full account is in asking AI for alternatives to a brand.

What rules make a neutral tracking prompt?

1. Name no brand, retailer or third party's rule. Your own brand's name turns the prompt into a test of the name. A named airline, store or standard ties the prompt to that thing's rules. For November, the index replaced a carry-on luggage prompt that cited the size limits of a particular airline.
2. Ask one thing. One intent per prompt, so that a result has one meaning.
3. Freeze the wording. The methodology says a revised prompt is stored as a new one and the old one is kept, so a ranking always lists the prompts that produced it. Do the same: version every prompt.
4. Keep budget and premium as separate prompts, and put prices in local currency. Budget and premium answers shared only 8.2% of their brands per category on average in October (shared brands divided by brands in either, averaged over categories with equal weight; usually 2 answers each, 155 of 160 categories). That is observed overlap, not a measured effect of the price word, so test budget and premium prompts side by side in your own set. A budget prompt needs a ceiling that fits the category; from November, the index's premium prompts carry no price figure.
5. Take wording from your own first-party language. Your site search terms, support questions and product reviews show how customers describe a need. We have not observed shoppers, so we cannot supply that vocabulary for you.
6. Pilot each prompt before you add it. Ask it a few times. If the answers recommend no brand, decide whether you want that prompt, since 9.1% of best-overall answers recommended none.
7. Keep each market's language. A Spanish-language market needs Spanish prompts, written for that market, not translated from an English set.

How many prompts do you need?

The index uses five per category. A small brand can start with five prompts per product line and repeat each four times per assistant, which is 20 answers per assistant, as set out in how to measure your brand's AI visibility. Add prompts for a product line before you add more repeats of the same prompt. Repeating one prompt tells you how that prompt varies. More prompts tell you more about the topic. The reasoning is in how many answers you need before trusting a change.

Frequently asked questions

Should I ever put my brand name in a tracking prompt?

Only as a separate, labelled study, such as "Is [brand] a good choice for [use]?". It shows how an assistant describes you when asked. It does not show whether you are recommended unprompted, so keep it out of your visibility number.

Can I reuse the index's prompts?

Yes for categories that match. They are public on each category page. Replace the category noun, keep the structure, and note that you changed it, because your series then differs from the index's.