Blog · 25 September 2026 · By HeardOf

How to choose the prompts you track: buying questions, not keywords

What prompts should you track for AI visibility?

The questions a buyer asks an assistant while choosing a product in your category, in the words that buyer would type: which tool is best for a team like theirs, which is best at one job, how three named products compare, what to use as the business grows. That is the short answer to what prompts to track for AI visibility, and it is what the 25 prompts of our research run of 4 September 2026 were built from; how to get your SaaS cited by ChatGPT and Perplexity prints all 25 verbatim, five per category, and says they "were written the way a buyer types them, not the way a marketer would."

A keyword names a topic; a buying question names a situation. Peec's setup guide, read on 25 September 2026, puts the contrast in two lines: the keyword "Best CRM software" against the prompt "What CRM would work best for a sales team of 10 people?". This post covers which questions go on the list. Running them — the sheet, the scoring, and where one answer per prompt stops being enough — is in how to track your brand in ChatGPT yourself. How many prompts to track, and whether a prompt that names a brand inflates that brand's score, are separate questions this post does not answer.

What kinds of question did our 25 prompts cover?

Five kinds, one of each in every one of the five categories, so each kind appears five times. The table gives two of the five categories as examples; the other three follow the same five slots.

Every one of the 25 asks for a product choice. None asks what a category is or how to do a task, none asks for alternatives to a named product, and none carries a price limit; they run 7 to 11 words each. Our audit uses a different mix: per our post on which engines we check, as updated 23 September 2026, its prompts come in "a fixed mix — 15 'best X for Y' questions, 10 alternatives, 10 head-to-head comparisons, 5 educational — generated from your category and the competitors you name", our own rules as read on 19 September 2026. The 25 are the research run's list, then, not the audit's, and the audit includes two kinds the 25 left out, alternatives and educational questions.

Why not just track your keywords?

Because, in the one study we read that looked at it, people writing a prompt for the same need did not reduce it to a few keywords. Rand Fishkin of SparkToro, with Patrick O'Donnell of Gumshoe, an AI-tracking startup, asked survey volunteers to write their own prompt for one need, headphones for a travelling family member, and wrote: "Across 142 responses, there were barely two prompts that, if you squinted, looked similar to me at all." The average semantic similarity between pairs of those prompts was 0.081, and the post's reading was that "People don't reduce their search intent to the fewest, most logical set of 2-5 keywords, they get creative, and weird, and highly specific." The study was published on 28 January 2026 per the page's structured data, which the page shows as 27 January, and is historical.

The same study found that the situation mattered more than the wording. Across 994 AI responses to those 142 differently worded prompts, Bose, Sony, Sennheiser and Apple appeared 55-77% of the time, and the post concludes that "the AI tools had captured the core, underlying intent across those 142 prompts". Asked about headphones for gaming, podcasting or noise suppression instead, Fishkin writes that he "got very different answers". Its conclusion is that visibility percent across many prompts, "whether written by humans or generated synthetically", is "likely to be" a decent proxy for how brands show up. The study makes that claim about many prompts, "dozens to hundreds of prompts run multiple times", not about any single wording; how many a list needs is a separate question this post does not answer.

Vendors' guides, read on 25 September 2026, say much the same. Profound's turns "best corporate card small business" into "What's the best corporate credit card for small teams?". Athena's says "Think buyer language, not internal language" and that "What's the best platform for tracking AI search visibility" beats "AthenaHQ competitors". Peec's guide says "Exact phrasing doesn't matter much." Semrush calls its prompt research "keyword research for the AI era".

Should the prompts you track name a brand?

Some can, if they are counted apart from the rest. In our run of 4 September 2026, 20 of the 25 prompts named no vendor; the five that did were the head-to-heads, and our post on getting cited says of them that "Those three get a mention rate that is partly handed to them by the question." How large that effect is belongs to a different post; the practical rule is to label the prompts that name brands so they can be read on their own.

The vendors' setup pages treat it as a setting. Athena, read on 25 September 2026, says "Non-Branded prompts are generic category queries where you want to show up without being named. You need both", and that non-branded ones show "whether you're showing up when buyers haven't decided yet". Peec's prompt discovery lets users set the share of branded prompts, with "20% branded" as the default. Conductor's setup has a tab where users choose branded prompts, non-branded prompts or both. Evertune's FAQ names three prompt types, one of them "aided, brand name in the prompt".

What do AI visibility tools suggest when you ask them for prompts?

A starting list, built from different inputs depending on the vendor. On the fifteen vendors' pages we read on 25 September 2026, the inputs named include your website or one of its pages (Peec, Athena, AmICited, Otterly); your SEO keywords (Otterly); your brand with its industry and competitors (Otterly, HubSpot, Surfer, and Ahrefs Brand Radar, from the brands in a saved report); topics you define, with the intents and personas you select (Profound, Conductor); a persona you write (Scrunch); localized topic suggestions in five languages (SE Ranking); or a panel or database of prompts: Evertune's "panel of 150M+ real user conversations", Semrush's "prompt database of 317M+ AI queries", Writesonic's "Millions of real prompts" from users of its Chatsonic product, on a page last updated 18 June 2025 and historical, and Goodie's "real customer queries across all major AI platforms". Surfer says it suggests only when a project is created. We opened no trial, so this is what each page says, not what the feature returns; the pages are listed with this post's sources, and the other four of the nineteen vendors we follow, Rankscale, Bluefish AI, BrightEdge and Similarweb, are not among them.

Several pages also name kinds of prompt: Peec sorts by three intents with "a 25/50/25 intent split" as the default, Semrush by five, Profound's guide by four types. AmICited says the page you feed it decides the kind: "a product page tends to surface bottom-of-funnel comparison and 'is it worth it' questions, while an educational blog post tends to surface earlier, definitional questions."

Otterly's help article of 17 July 2026 says "AI search engines like ChatGPT and Perplexity do not publish query data." The panels above are data the vendors say they collect, which is a different thing, and Evertune's FAQ, read on 25 September 2026 and printing no date, says "over 80% of AI prompts are phrased uniquely and have never been asked the same way twice". We could not check any of the panels or the 80% figure. Where the vendors say their prompt-volume figures come from, in their own words, is in where ChatGPT search volume comes from.

What did SparkToro's 12 prompts ask?

Each of the twelve asks for a list of recommendations and none asks the tools to compare named products, as the study's results site lists them on 25 September 2026. The study post says its "600 volunteers ran 12 different prompts through each of the 3 tools a combined 2,961 times" in November and December 2025. Three of the twelve are business purchases, each with its situation in the wording: a digital marketing consultant for "My e-commerce business", CRM software that "A dental clinic asked me to recommend", and cloud computing providers "SaaS startups should consider". Three name a maker or platform as the place to look, Volvo, Instagram and TikTok; none names a product to judge. Of the other nine, three ask about local services, three about social media accounts, and three about consumer products: science fiction novels, microphones and a third. Five of the twelve ask for exactly five items.

That third consumer product is one thing we cannot square. The study post quotes a prompt about "the top chef's knives, brand and model, for an amateur home chef with a budget <$300" as one of the twelve; the results site labels its question 4 "best power station for camping", while the answer previews it shows under question 4 are about chef's knives. We report both and do not know which label is right.

The study also left the question of this post open. Among the questions it says others need to answer: "What variety of prompts are needed to get statistically meaningful answers about a sector/intent?" We found no published answer to it on the pages we read.

How do you write your own list?

Start from the buying situations in your category, not from your keyword list. The five slots above are one set: best for a segment, best with a qualifier, a head-to-head, best at one capability, a recommendation for a growing business. Our audit's mix, per the same post, as updated 23 September 2026, our own rules as read on 19 September 2026, includes alternatives and educational questions, and Profound's guide gives "Instacart alternatives" and "Alternatives to Sephora" as examples of an alternatives topic.

- Write each one as a sentence a buyer would type, with the situation in it. Peec's guide lists audience, use case and constraints as the context a prompt carries; SparkToro's 142 prompts show how far real wording varies for one need, which is why the situation is what to get right.

- Label the prompts that name brands, and read them apart. Athena labels each prompt branded or non-branded, and Conductor's setup asks which of the two to generate; our own head-to-heads are the example of why.

- Treat a vendor's suggestions as a draft. AmICited's own guide says "Treat the output as a draft to prune, not a final list to accept wholesale", and asks of each suggestion "would you actually type this into ChatGPT?".

- Fix the wording before the first run. Otterly's help article says "The prompt text itself cannot be edited" and that historical data "won't carry over" to a reworded prompt; Athena's says that once a prompt has response data "its text is locked". A reworded prompt starts a new history.

Then run the list the way how to track your brand in ChatGPT yourself describes, with the same wording every time.

How was this read?

Every page was fetched on 25 September 2026 between 00:58 and 01:01 UTC with curl and a Chrome user-agent and read in full; all answered 200 and none returned a bot challenge. Every quoted span was checked by script against the served text of the page named. The count of 20 of 25, the five slots and the word counts were made on the same day from the 25 prompts as printed; they describe the wording of the prompts and carry no result from the run. SparkToro's page is dated from its structured data, which the page shows a day earlier; historical means published more than 90 days before the read.

What we could not do: check what any vendor's suggestion feature returns for a given category, since no trial was opened; check the size or make-up of any vendor's prompt panel or database, or Evertune's 80% figure; say which of the two labels on the SparkToro results site's question 4 is right; and find a published count of the kinds of question real buyers ask assistants about B2B software. Semrush's knowledge-base article 1503 now answers at a new address ending "prompt-tracking". Peec's, Athena's and Conductor's documentation pages open with notes addressed to AI readers pointing at an index file, and Writesonic's with a note on fetching its documentation index; all were read as data and not followed. Everything this post says about our own audit comes from our post on which engines we check and cannot be checked from here; the rest carries a link or the day it was read.

Where does "ChatGPT search volume" come from? What five vendors say about their prompt data