Which prompts to track for AI visibility

Start from questions your buyers already asked

Your prompt list should sound like a transcript of buyer conversations. A prompt nobody has ever asked in some form is trivia with a label on it.

The raw material is already inside the company. Call recordings where a prospect asks how you compare. Support tickets phrasing the same complaint three different ways. The search box on your own site, and the threads where your category gets argued about. People write AI prompts the way they would ask a friend, longer and more sceptical than a search query: “what are the downsides of switching CRM mid-year, and will we lose our dashboards”.

Write prompts in the buyer's phrasing. “Best enterprise CRM platform 2026” is how a page title is written, not how a person asks for help.

Six shapes of prompt worth tracking

Category, use-case, comparison, alternatives, branded and objection-shaped prompts cover what buyers ask before they know you, while they shortlist, and after they have your name in mind.

• Category. “What is the best help desk software for a 20-person team”. Answers whether AI puts you in the set at all.
• Use-case. “How to cut churn emails without a data team”. Shows you doing the specific job, which is where most buying starts.
• Comparison. “Zendesk vs Freshdesk for a small support team”. Reveals who else makes the list and in what order.
• Alternatives. “Cheaper alternatives to Intercom”. A switcher's prompt. Missing here means missing from a conversation you will never hear.
• Branded. “Is Northwind CRM worth it”. What AI tells someone who already knows your name.
• Objection-shaped. “What goes wrong when you switch CRM mid-year”. Surfaces the doubts AI repeats back to buyers on your behalf.

Six shapes is a frame, not a quota. If your category has one obvious vendor and nobody asks comparison prompts, drop that shape rather than padding it. What should not change is that every prompt traces back to something a person asked.

How many prompts to start with

Start with 15 to 25 prompts spread across the funnel. Below 15 each metric rests on too few answers to move, and past 25 the list becomes a job of its own, which is when people quietly stop maintaining it.

Weight the set toward the stages before anyone knows you. That is where a mention changes a shortlist, and it is the harder state to reach. A set made mostly of branded prompts measures a question your marketing already answers.

The setup wizard in advisible suggests prompts when a workspace is new. Treat them as a first draft: they describe the category in generic language, and the edit that matters is rewriting them in your buyers' words.

Keep branded prompts on their own line

Branded prompts mostly measure whether AI has read your site, so report them separately. Stack them into the main set and they will carry your Visibility while the prompts that reach new buyers sit flat.

This is the most common way a prompt set goes wrong. A team lists 25 prompts, 12 of them start with the brand name, the dashboard shows a comfortable Visibility figure, and everyone stops looking. The branded prompts were never in doubt: the engines have read the site, the docs, the review pages. They answer a different question, namely what AI tells someone who already knows you, and that question is worth its own row in the report.

One detail sharpens the distortion: in advisible, prompts you mark as more important count for more in Visibility. Giving branded prompts extra weight on top of their numbers double-counts the tilt, so keep their weight plain and read them as their own segment.

A decision rule for adding a prompt

Add a prompt when a buyer has asked something like it and the answer would change what you do. Both conditions, or it stays out of the list.

1. A person asked something like it. Quoting a call or a ticket beats inventing a clever variant nobody types.
2. The answer would change an action. If a bad result would not move a roadmap, a page or a phone call, the prompt is decoration.
3. It is not a near-duplicate. Two prompts that always produce the same answer are one prompt with extra maintenance.
4. You can name its funnel stage. If you cannot say which buyer moment it represents, you will not know what to do with the result.

Prune the list each quarter: drop prompts nobody acted on, add the questions sales heard last month. Keep the set stable inside a reporting window, because changing the sample mid-month makes the trend line a comparison of two different experiments. For the same reason, a fresh prompt has no number worth reading yet, which is the subject of how many answers a number needs.