Prompt baskets: why hand-picked lists lie

Somebody has to write the prompt list. Usually it goes like this. Someone opens a blank document on a Thursday, the team brainstorms twenty questions their buyers might ask an assistant, and that list becomes the denominator of every AI-visibility number the company reports for the next year.

Nobody was being lazy. The problem is authorship. A list written by the team being measured drifts toward what that team hopes is true, and every number downstream inherits the drift without ever mentioning it.

Hand-picked lists flatter their makers

Ask people to write prompts about their category and they write prompts about their brand.

The list fills up with questions a loyal customer would ask. The replay then finds the brand everywhere it looked, and the dashboard reports magnificent visibility on demand that nobody was contesting. That is the you-typed-our-name problem, and it is the cheerful half of the failure.

The other half is quieter and costs more. A hand-picked list misses the questions real buyers actually ask, so the gaps that matter never get measured at all. A list cannot find what it never thought to ask.

There is a third failure underneath both. Hand-picked lists freeze. Buyer language moves every quarter, new competitors reframe the category, and a static list keeps measuring the questions of the year somebody wrote it. Nobody ever schedules the meeting to admit their list aged out.

Baskets built from demand you did not invent

Quattr generates prompt baskets from evidence instead. Your prompt basket is the set of questions replayed in the engines to measure you, and every one of them comes from demand you did not invent.

The recipe also decides what branded means. Branded prompts get tracked separately, never silently blended into the competitive read, because demand that already contains your name was never contested. Non-branded is where visibility is actually won and lost.

Then your taxonomy does the organizing. Prompts group by intent in your own business language, so a coverage gap reads as a category you under-serve rather than as a random pile of misses.

Four first-party sources feed it, and each carries something the others cannot. Intents ground the category questions. Logs and referrals show what the assistants already send you. Sales language captures demand that has not reached a search box yet, which is usually where next quarter's queries are hiding.

• The search queries and intents your site already earns.
• The questions showing up in your log files.
• The AI referrals that actually landed on your pages.
• The words buyers use in sales conversations, before those words become queries.

Sampling and replay make it a metric

Your prompt basket is sampled rather than exhaustive, sized to your tracker tier, and replayed on a schedule.

Stability is the whole point. The same questions asked the same way, week after week, is what turns a pile of answers into a trend line you can act on. Movement only means something against a constant denominator.

Where an answer carries no citations at all, the measurement falls back to Google rather than inventing a signal to fill the gap. And when a basket needs to grow, it grows from the same first-party sources, never from a fresh brainstorm.

A stable sample you can replay beats an exhaustive crawl you can run once. That is the trade, and the honest thing is to say it out loud rather than bury it in a footnote.

Good teams audit the denominator first

Here is the pattern we keep seeing. Before an AI-visibility number goes near a leadership deck, somebody audits where it came from. Non-business domains out of the competitor set. Prompts re-validated against real buyer personas. Definitions and proof demanded for any claim the number implies.

That instinct is correct, and this recipe is its systematic form. A number that survives the who-chose-these-prompts question is genuinely rare in this industry, which is exactly why it is worth owning.

Vendors can feel that scrutiny arriving, and it is healthy for everyone. The ones with a defensible recipe publish it. The ones without one publish a bigger number.

See also: how citations, mentions and share of voice get measured →

Where the prompt basket becomes work

A prompt basket is not only a measuring device. It feeds AI-answer visibility, one of the eight kinds of work the Quattr Method runs, and its output turns into a queue. Prompts where you never appear become content briefs. Pages that already earn citations join the protect list.

The example below shows that read running across five AI segments in a single ask, with the per-engine cuts first.

If you take one habit from this article, take the audit question. Who chose these prompts, and from what evidence? Any vendor, this one included, should be able to answer that in one sentence.

See also: the eight kinds of work, and which question each one moves →

The workflow that does this: Five AI segments, one ask →