Chapter five of ten
The prompts you track are the backbone of every visibility number you report. Build the set around themes your buyers care about, then keep it stable enough to show a trend.
Buyers ask Which prompts should we track?
By Maeva Cifuentes · Updated October 2026
Several related prompts get grouped and tracked as one score.
Each buyer asks the same thing in their own words.
The short answer
Pick commercial, category-level prompt themes that would automatically generate a response with a list of you and your competitors. Write a handful of prompts for each theme, varied by buyer, competitor or feature. Keep prompts that include your brand name in a separate set that checks accuracy. Read the trend for each theme and treat any single prompt as one sample.
A visibility score is a share of a prompt set. Change the set and the score changes with it, even when nothing about your brand has moved. So the choice of prompts is the first measurement decision in an AEO program.
Mention rate and citation rate are both counted against it. Chapter 8 covers those metrics.
The themes where you are absent point to the pages worth building or fixing in chapter 6.
The sources engines cite for your prompts become the sites to earn a place on in chapter 7.
Ask who chose the prompts behind any score you are shown and how many there are. Ask whether the set changed between the two dates being compared.
Write down a reason for every prompt. Get the set approved before you record a baseline, since later edits reset the comparison.
Prompts run longer and more personal than search keywords. People describe their company, their constraints and the task they want done.
words in the average ChatGPT prompt with search switched on, up from 4.7 a year earlier
Semrush clickstream study, April 2026, comparing January and February of 2025 and 2026of ChatGPT prompts matched zero keywords in a database of 27 billion
Semrush, April 2026, for much of October 2024 to February 2026words in the median observed prompt in June 2026, up from 16 in July 2025, across 387 prompts
Seer Interactive, June 2026, from seven observation studiesThe figure of 23 words that circulates widely comes from an older Semrush study published in February 2025. That figure covered prompts sent with search switched off. Semrush's April 2026 update puts those prompts at 13.5 words, down from 24.9 a year earlier.
Zero AI platforms publish the prompts their users type. Every volume figure in a tracking tool is modeled from a browser panel, from keyword volume, clickstream data or from a mix of them all. Ahrefs states that its numbers are built from Google keyword volume without user prompt data (August 2026). Scrunch describes prompt volume as directional and fragile in its own help centre (September 2026).
At Flying Cat, sometimes we'll use search volume to help support our decisions, but there are many other proxies we can use. Sales calls, support tickets and GSC impression data are stronger evidence that buyers ask a question.
A theme is a group of prompts that share one buyer intent, such as finding the best product in a category or replacing a named competitor. We choose themes first and write prompts second. Tracking every long variation of one question raises the tool bill and adds little signal.
Pick one representative prompt for every way a buyer might phrase the same need. "Best project management tool for a 20-person agency?" and "Give me a short list of project management tools for small agencies" ask for the same thing, so you track one of them. A new prompt earns its own slot only when the answer would change, for example when the buyer adds a must-have integration or a much bigger team.
SparkToro and Gumshoe ran 12 prompts 2,961 times in January 2026. The same list of brands came back in fewer than 1 in 100 runs. In a follow-up run of 142 headphone prompts, the leading brands still appeared in 55 to 77% of 994 answers, so a share measured across many prompts and runs is a fair metric. Peec, a tracking vendor, tested 1,754 prompts in June 2026 and found that rewording a prompt rarely changed which brands appeared until the meaning itself drifted.
| Who | Suggested size | How the set is built |
|---|---|---|
| Scrunch, a vendor | 12 to 15 prompts per topic cluster | Topic clusters built from SEO and paid keywords, then pruned after two weeks of data |
| Ahrefs, a vendor | A few dozen prompts | A narrow set of topics, expanded as you enter new areas |
| Kevin Indig, Growth Memo | 15 to 30 per buyer segment | Brand, category and problem prompts, run several times |
| Aleyda Solis | 30 to 50 to start | 100 to 250 for SaaS with several personas and competitors |
Sources are Scrunch (2026), Ahrefs (2026), Growth Memo (February and June 2026) and Aleyda Solis (June 2026). Solis notes that her ranges are starting points.
We build 5 to 8 prompts per theme across 6 to 10 themes. That gives 30 to 80 prompts for one product in one market. A theme with three prompts or fewer is too thin to report on, so we merge it into a neighbour.
Ask for results by theme. A single rolled-up number hides where you are strong and where you are absent.
Tie each theme to a page type you want to win, such as comparison pages or use case pages. The theme then tells you what to build.
Start from questions buyers have already asked you. A prompt invented at a desk measures what the team imagines buyers ask.
Pull the questions prospects ask when they want a vendor named for them. Thirty call transcripts is a workable sample.
These carry the buyer's own words for problems and features, which often differ from your marketing terms.
Filter queries of ten words or more with the pattern ^(\S+\s+){9,}\S+$. Long conversational queries hint at how people phrase the need.
Terms like "best", "alternatives" and "vs" that you and your competitors rank for map straight onto prompt themes.
The searches an engine runs for your category, covered in chapter 3, show which angles it checks before answering.
Tracking tools suggest prompts. Treat them as a second input and check each one against the sources above.
Search Console shows queries typed into Google. It carries zero labels for AI Mode, so treat long queries as a hint. The pattern comes from Chris Long, Search Engine Land, February 2026.
These rules come from auditing tracker setups that reported a flattering number and taught the team nothing.
A prompt that names you will mention you nearly every time. Measure visibility only on prompts where the engine has to choose you unprompted. Competitor names are fine.
Prompts such as "What does [brand] do?" show whether engines describe you correctly and which sources they draw on. Report this set on its own, away from the visibility number.
Broad educational questions often get an answer with zero products named. Check the response before you pay to track the prompt.
Vary a prompt on something that changes the answer, such as the buyer, the market, the competitor or the feature. Two prompts that differ only in wording count as one.
Keep two similar prompts only when their results differ by more than a few points over a month.
Auto-generated topics rarely match your strategy. Tag prompts by theme, competitor and priority so you can filter reports later.
Build yours
Fill in the form and the builder writes a starter set sorted by theme, with a separate branded set. The set is built in your browser from the rules in this chapter.
A category and one buyer are enough for a first draft.
The builder writes from patterns, so read every prompt aloud and reword any that a buyer would never say. Run each one in an engine before you track it. Drop those that come back with zero brands named. The CSV columns are prompt, theme, buyer, stage, tags and set.
Engines tailor answers to what they know about the person asking. iPullRank tested this on Google's AI Mode in May 2026. In accounts with Personal Intelligence switched on, the brands under test were mentioned in 66.8% of answers, against 23.9% in a blank control account. Tracking tools usually collect answers from a logged-out session or the API, so they see the general answer.
Exact prompts are close to impossible to match under that much variation. We group the likely variants and set up personas in Scrunch. One prompt runs several times, each time framed for a different buyer such as a company under 100 employees or a global enterprise. A simulated persona is a rougher signal than a person's own chat history. It still shows how a brand performs for each audience.
Use interviews, calls and reviews. Kevin Indig suggests a short card per persona covering the job to be done, constraints, success measure, decision criteria and vocabulary (Growth Memo, February 2026).
Set the country and language of each prompt to the buyer it represents. A German buyer persona runs German prompts from Germany.
Tools bill per prompt per platform. Three personas on forty prompts across four platforms is 480 tracked variants, so add personas to your priority themes first.
To see whether you are improving, you have to compare the same prompts from one month to the next. If you add five new prompts in March, your March score moves because the set changed, even if engines see you exactly the same way. So make changes to the set on a schedule, and write down the date of each change so everyone reading the report knows where the comparison resets.
Read the answers for your priority themes. Flag prompts that stopped surfacing brands or drifted off topic.
Add prompts for new products, competitors or markets. Retire the flagged ones in the same batch.
Record the date and restart the baseline for any theme you edited. Keep a core group of prompts untouched all year.
Aleyda Solis recommends a similar rhythm with a stable core, monthly checks on priority prompts and a full review every 6 to 12 months (June 2026). Scrunch suggests a monthly review with additions each quarter.
Expect a spot check to disagree with the report. Typing a tracked prompt yourself gives one personalized sample. AI visibility tracking only works as a trend.
Sometimes a client or a colleague tracks the same brand in a different tool and gets a different visibility score. Before you debate which number is right, compare the two prompt lists. Different prompts produce different scores, so the gap often comes from the lists themselves, even when the brand's standing is the same in both.
Eli Schwartz argues that tools charge per prompt, which rewards tracking more variations. He prefers a broader share of voice that checks whether you appear wherever your competitors appear (Previsible, January 2026).
Profound, a vendor whose pricing scales with prompt count, tells customers to start at 100 and says they track up to 1,000 (February 2026).
Kevin Indig recommends repeating each prompt several times per platform and reporting a range (Growth Memo, June 2026). Graphite found that 10 runs gave an average error of about 5.6% (April 2026).
Victoria Spall of Browser Media argues that trackers measure simulated prompts and that Search Console, Bing and analytics data deserve the attention (August 2026).
Choose your next stop
We use analytics cookies to see which pages help people. They load only if you accept. Cookie policy