Citino teamTrack the questions that would change a purchase, worded the way your buyers ask them. Everything AI search analytics reports comes from that list, so a weak list gives you precise numbers about the wrong things.
Start from buying decisions, not keywords
A keyword list tells you what people type into a search box. A prompt set should tell you what a buyer asks an assistant while deciding what to buy. The two overlap, but a question to an assistant often carries the buyer’s situation with it: “best WMS for a 3PL with 12 clients on Shopify”, not “WMS software”.
For each prompt, ask one question: if an assistant answered this without naming us, could we lose a deal? If the answer is no, leave it out. Questions that mention your category but would never lead to a purchase (definitions, history, salary questions) add answers to your report without adding meaning.
Where to find the questions buyers really ask
The best sources are the places where buyers already write or say their questions in their own words:
- Sales calls and demo notes. The first question on a call is often the one the buyer just asked an assistant. Note the exact phrasing.
- Support tickets and onboarding questions. These show the facts buyers check before and after buying: limits, integrations, migration, price.
- Search Console queries. The Performance report groups your Google traffic by the query people typed. Long, question-shaped queries are good candidates to rephrase as prompts.
- Your own site search, forums and community threads. These show how people describe the problem before they know which product category solves it.
- Lost-deal notes. The reason a buyer chose someone else often becomes a comparison prompt.
Collect more than you need, then cut.
The kinds of question to cover
A useful set covers the whole path from problem to purchase. A set made only of “best X” questions will tell you how you do at the top of the funnel and nothing about the questions that close deals.
| Kind | Example (illustrative) | What it tells you |
|---|---|---|
| Category | “best warehouse management software for a 3PL” | Whether you make the shortlist |
| Problem | “replace spreadsheets for 3PL client billing” | Whether you are named before the buyer knows your category |
| Comparison | “Crateline vs Northport WMS for cold storage” | How you are described against a named rival |
| Alternatives | “alternatives to Palletworks for small warehouses” | Whether you are named when a buyer is leaving a competitor |
| Fact check | “which WMS handles multi-client billing natively?” | Whether engines know a claim you need them to know |
| Implementation | “how long does a WMS implementation take?” | Whether you appear once the buyer is planning |
The brand names in the table are fictional. Replace them with your real competitors: the brands that already appear in answers to your category questions are the ones to compare against.
How to word a prompt
- Use the buyer’s words, not yours. If buyers say “client billing”, do not write “multi-tenant invoicing”.
- Add the details buyers add. Company size, industry, location and the tools they already use change which brands an answer names. Include them where buyers would.
- One question per prompt. “Best WMS and how much does it cost?” mixes two measurements.
- Leave your brand out of discovery prompts. Category, problem and implementation questions should not name you.
- Fix the wording once it is live. Research on large language models has found them highly sensitive to small, meaning-preserving changes in prompt formatting (Sclar et al.). That study tested open models on benchmark tasks, not search assistants, but the lesson carries over: a reworded prompt is a different question, so it should start a new series.
Keep branded prompts in their own group
A question that names you, such as “is Shipmatic good for ecommerce fulfilment?”, will almost always get an answer that names you. Mixed into the main set, branded prompts push visibility up without telling you anything about whether buyers find you.
Track them anyway, in a separate group. They are the best way to check what engines say about you: your price, your integrations, your limits. If an answer gets a fact wrong, that is a content gap you can fix.
How many prompts you need
For one brand, a set of 25 to 150 prompts covers most buying decisions. Two things decide where you land:
- The topics you want to report on. If leadership will ask how you do on billing questions, you need several billing prompts, not one. A single prompt is too easily swung by one answer, and its result cannot be told from noise.
- The margin of error you can live with. Answers to the same prompt tend to move together, so adding prompts narrows the margin far faster than running the same prompts more often. The reasoning, with a worked example, is in how to measure AI visibility.
Each prompt is asked on every engine you track, so a prompt is a question, not an answer: 50 prompts on four engines, asked daily, is 6,000 answers a month (worked example).
Group prompts so the numbers mean something
Tag every prompt with a topic and a buying stage before you collect any data. Groups let you see that you lead on fulfilment questions but are absent from billing ones, which a single overall number hides.
Groups also make fixes measurable. To tell whether a page change worked, you measure the prompts it targets, chosen in advance, against the same prompts for your competitors. That is only possible if the prompts were grouped before the fix shipped. See how to tell whether an AI search fix worked.
Mistakes that skew a prompt set
| Mistake | What it does to the numbers | Fix |
|---|---|---|
| Only “best X” questions | Ignores the questions that close deals | Add problem, comparison and fact-check prompts |
| Branded prompts mixed in | Inflates visibility | Track them as their own group |
| Rewording prompts in place | Splices two questions into one trend | Retire the prompt and start a new one |
| One prompt per topic | Topic results swing on single answers | Track several prompts for each topic you report on |
| Prompts chosen after seeing results | Picks the questions you already win | Fix the set, and each fix’s target prompts, in advance |
| Generic wording | Measures a buyer nobody is | Add the details real buyers include |
Keep the set stable, and review it on a schedule
A prompt set is a measuring instrument. Change it constantly and you cannot compare one month with the next. Review it on a fixed schedule, such as once a quarter: add questions buyers have started asking, retire ones that no longer matter, and keep everything else unchanged. Retiring a prompt should leave its history in place, not delete it.
Assistants also turn one question into several searches. Google, for example, describes AI Overviews and AI Mode issuing multiple related searches across subtopics to build one response. You do not need a prompt for every sub-search, but if your prompts cover the questions buyers ask, the subtopics those searches touch are usually the same ones your content needs to answer.
Where Citino helps
When you sign up, Citino suggests the questions your buyers ask and the competitors named alongside you, and you can add your own. Prompt wording is fixed once a prompt goes live: you can retire a prompt and start a new one, and the chart shows the break. Every prompt is asked daily on every engine in your plan, and each result carries a margin of error, as set out in the methodology. See how tracking works on the Measure page.
Choose prompts the way you would choose what to measure in an experiment: in advance, in the buyer’s words, and then leave them alone.