How to choose the prompts to track in AI search
Benjamin Libor10 min read
The right prompts to track are the questions your buyers actually ask an assistant on the way to a purchase: the problem, the category, comparisons, alternatives, pricing and the how-to questions your product answers. Write them as full sentences with context, keep branded questions in a separate group, and fix a set of 20 to 50 for at least a quarter so the numbers stay comparable. Here is the method, with a worked list of 12 prompts for a made-up scheduling tool.
In short
- Build the list from the stages of the buying journey, not from your keyword list; a keyword is not a question.
- Write each prompt as a person would type it into ChatGPT: a full sentence, a role or a constraint, and the context that makes the answer specific.
- Keep prompts that name your brand in a separate group, so they do not inflate the headline number.
- Fix a core set of 20 to 50 prompts for at least a quarter, add a few rotating ones, and review the set once a quarter; a list that changes every week produces numbers nobody can compare.
Start from the buying journey, not from keywords
People type keywords into Google, but they ask assistants questions, and the question carries a stage of the buying journey with it. Someone asking "why do my sales reps double-book meetings" is at a different point from someone asking "is Fernwick worth it for a five-person team". Track only the second kind and you miss everyone upstream.
Write two or three questions for each stage:
- Problem. The buyer has a pain and no category name yet. "How do I stop clients from no-showing on sales calls?"
- Category. They have learned there is a kind of product for this. "What is a scheduling tool and do I need one?"
- Best for. They want a shortlist with a constraint. "What is the best scheduling tool for a recruiting agency?"
- Comparison. Two or three names, head to head. "Fernwick vs Calendly for a sales team of ten, which is better?"
- Alternatives. They are unhappy with something. "What are good alternatives to Calendly that support round-robin booking?"
- Pricing and worth. "How much does Fernwick cost and is the paid plan worth it?"
- How-to. A task where your product is the answer. "How do I let prospects book a meeting with whichever rep is free?"
The last stage is the one most teams skip, and often where the cheapest wins are: an assistant answering "how do I set up round-robin booking" has to recommend tooling.
Write prompts the way people type them
"scheduling software b2b" is a Google query; "I run a 12-person agency and we book client calls by email, what scheduling tool should we use?" is a ChatGPT prompt. The second produces a more specific answer, and your visibility on it tells you something real.
Three things make a prompt realistic:
- Context. Company size, industry, the tool they use today, what is going wrong.
- A role or a constraint. "as a head of sales", "for a team on Google Workspace", "under 10 euros a seat", "that works in German".
- A plain question at the end. The assistant answers what is asked; make the ask clear.
One sentence of context and one question is enough. A useful test: type each candidate into ChatGPT and Perplexity once. If the answer is a list of products in your category, the prompt works. If it is a general essay about productivity, the prompt is too vague and you will never be named.
Keep branded prompts in their own group
Prompts that contain your brand name ("what does Fernwick do", "Fernwick pricing") almost always name you. That is not visibility; that is the assistant reading your website back. Mixed into the main set, they inflate the average and hide the problem.
Rule: a prompt that names the brand is a branded prompt. Track it, but score it separately, and never let it count toward the headline visibility number.
Branded prompts are still worth tracking, for a different reason: they show what the assistant says about you when asked directly, which is where wrong facts show up (an old price, a feature you removed). That is a job for finding and fixing wrong facts, not a visibility job. In Echo, the Prompts page keeps the two groups apart for exactly this reason.
Comparison prompts that name you and a competitor ("Fernwick vs Calendly") count as branded too, because you are named by construction; on them, position and sentiment are the whole story.
Cover every segment, market and language
A set that only reflects one buyer tells you about one buyer. List the segments you sell to (say, agencies, in-house sales teams and clinics for Fernwick), the countries and the languages, and give each segment three or four prompts across the journey.
Language matters more than most teams expect. Assistants answer in the language of the question and draw on sources in that language, so your visibility in German prompts can be very different from English. If a quarter of your revenue comes from France, a quarter of your prompts should be in French, written by someone who speaks it, not translated word for word. Markets with the same language still differ, because the review sites, forums and regulations do; write the country into the prompt when the country changes the answer.
Where to find the questions
Do not invent prompts in a meeting. Four sources give you real ones:
| Source | What to pull | How |
|---|---|---|
| Search Console | Queries with question words (how, what, which, best, vs, alternative) | Filter the queries report; rewrite each as a full sentence |
| Sales calls | What the prospect asked before they booked, and what they compared you to | Ask the reps for the five questions they hear most |
| Support tickets and chat | How-to questions where the answer is a feature | Sort by frequency; each recurring one is a how-to prompt |
| Forums and reviews | How people describe the problem in their own words | Read the threads about your category on Reddit and the review sites; copy the phrasing |
The forum source matters because AI answers lean on Reddit and review sites for exactly these questions. If buyers on Reddit write "something like Calendly but for teams that share a calendar", that is a prompt, nearly verbatim.
How many prompts, and how often to change them
Aim for a core set of 20 to 50 prompts, fixed for at least a quarter. Fewer than 20 and a single answer flipping moves your visibility by several points. More than 50 and nobody reads the list, and the long tail hides the prompts that matter.
The number matters less than the stability. Visibility is a share: the fraction of answers that name you. Every prompt you add or remove changes the denominator, and last month's number stops being comparable with this month's. A list tuned every week gives you a chart of your editing, not the market.
Keep a second, small group of rotating prompts, five to ten, for things you want to look at this month: a new feature, a competitor's launch, a market you are testing. Score them separately, and promote one into the core set only at the quarterly review.
A team starting from zero: 25 core prompts, unbranded, across the seven stages and the main segments; 8 branded, including the two or three comparisons that matter; 5 rotating. Under 40 prompts, enough to produce a share of voice number that moves for reasons, not noise.
What a bad prompt looks like
Most weak sets have the same faults:
- Too vague. "What is the best software for teams?" No assistant will name a scheduling tool here. If the answer cannot plausibly contain your category, cut the prompt.
- A keyword, not a question. "scheduling tool pricing" is a search term. Nobody asks it that way, and the answer is less specific than for "how much should I expect to pay for a scheduling tool for 20 people".
- Too branded. A set where half the prompts contain your name gives you a flattering number and tells you nothing about new buyers.
- Written for the product, not the buyer. "What scheduling tool has two-way calendar sync, round-robin and a Zapier integration?" is your feature list. Buyers describe their problem.
- All one phrasing. If every prompt starts "What is the best…", you are measuring one kind of answer. Mix in how-to, comparison and problem questions.
A worked list: 12 prompts for Fernwick
Fernwick is a made-up scheduling tool for small sales teams and agencies. A starter set, with the stage in brackets:
- Our sales team of eight keeps double-booking demos and we schedule by email. What is the fix? (problem)
- What is a scheduling tool and is it worth paying for one at a small company? (category)
- What is the best scheduling tool for a recruiting agency that books interviews for clients? (best for)
- Which scheduling tools let prospects book with whichever rep is free, round-robin style? (best for, feature-led)
- I am a head of sales at a 40-person SaaS company on Google Workspace. Which booking tool should we standardise on? (best for, role)
- Fernwick vs Calendly for a sales team of ten, which is better and why? (comparison, branded)
- What are good alternatives to Calendly for teams that need shared availability? (alternatives)
- How much does a scheduling tool cost per user, and what do the paid plans usually add? (pricing)
- Is Fernwick worth it, or is the free tier of a bigger tool enough for a five-person team? (worth it, branded)
- How do I let clients book a 30-minute call only on days when two of us are free? (how-to)
- Was ist das beste Terminbuchungstool für eine kleine Agentur in Deutschland? (best for, German market)
- Does Fernwick work with HubSpot and does it log meetings to the deal automatically? (branded, integration)
Nine are unbranded, three count as branded (two direct, one comparison), and one is in German because Fernwick sells in Germany. A real set adds three or four per segment and a couple per market, which gets you to the 25 to 40 range.
Review the set every quarter
Put a quarterly review in the calendar, one hour, with this agenda:
- Which prompts never produce a list of products? Cut them; they are too vague to win.
- Which prompts have you won every week for three months? Move them to a "won" group you check monthly.
- What changed in the market? A competitor launched, a feature shipped, a new segment opened. Add a prompt for each, into the rotating group first.
- What do the sources say now? Re-run the Search Console filter and ask the reps again; copy the new phrasings.
- Promote and retire. Move rotating prompts that proved useful into the core set, retire the ones that did not, and write the change down with the date.
Make the changes in one batch, on one day, note it on your reports, and leave the set alone for another quarter. The AI visibility routine for a two-person marketing team shows where this review fits in the year.
Questions people ask
How many prompts should I track for AI visibility?
Twenty to fifty core prompts is the practical range for a small team. Under twenty, one answer changing moves your visibility by several points and the weekly chart turns into noise. Over fifty, the long tail dilutes the prompts that matter. Add five to ten rotating prompts for things you want to watch this quarter.
Should I include my brand name in the prompts?
Yes, but in a separate group that does not count toward the headline number. Prompts with your name in them almost always name you, which tells you nothing about buyers who do not know you yet. They are still worth tracking to catch wrong facts and to see how you are described.
How often should I change the prompts I track?
Change the core set once a quarter, in one batch, and write the change down. Visibility is a share of answers, so every edit changes the denominator and breaks comparison with earlier weeks.
Written by Benjamin Libor, founder, echo.
Something wrong or out of date? Write to hello@echo-aeo.com and we'll fix it.