How many AI-search prompts to track: cover decisions, not wording variants

Count the decisions represented

A prompt allowance is not a sampling plan. Twenty-five questions that cover distinct buying decisions can be more useful than a hundred paraphrases of the same broad recommendation. The right size depends on the products, locations, audiences and constraints the business can actually serve.

Start with observed customer questions where available. Keep their source labels: a recorded conversation, a search query, a sales question and a generated test prompt are different types of evidence. A generated question can fill a test gap without becoming proof of demand.

Group prompts into families

Map each family to a real decision: location fit, price or contract conditions, compatibility, comparison, or a relevant use case. Keep related follow-up turns when they change the answer. A budget constraint added after the first response may matter more than another paraphrase of the opener.

For each family, record the offer it concerns and the page or independent source that could answer it. Remove tests that the business cannot satisfy. A competitor appearing on an irrelevant question is not automatically a winnable opportunity.

Use a workload calculation, not a universal benchmark

For example, 25 prompt families tested across two engines with four repeats require 200 answer captures per measurement wave. Moving to 100 families under the same design requires 800. This is planning arithmetic, not a recommendation that every brand should track either number.

The earlier 10, 100 and 1,000-prompt ladder described service-planning heuristics. It did not establish that those counts are representative samples. Use your coverage gaps and available review capacity to choose a size.

Separate coverage from repeat precision

More distinct families improve coverage only when they add meaningful requirements. More repeats help assess variation within a chosen family. Neither automatically fixes bias in how the original questions were selected.

Read the repeat-cadence guide before treating one answer per prompt as a stable visibility score. A larger list of near-duplicates can make a chart look precise while over-weighting one topic.

Maintain a stable core and a labelled exploration set

Keep a core set for comparisons over time. Put newly discovered requirements in a separate exploration set until you decide whether to add them. Record additions and removals so a score does not change silently because the denominator changed.

Review the core against actual business changes. A new location or product can justify new families. A wording suggestion alone usually needs less attention than a missing buyer requirement.

Aiso's methodology explains the distinction between sampled conversations and designed tests. Use both when they serve the decision, without merging them into an unsupported market-share estimate.