Start with the buying decision
Use an intent_id for the task a buyer is trying to complete and a wording_id for each exact question used to test it. A cluster might cover choosing a customer support tool for a small team. “Which support tool suits a five-person team?” and “What should a small support team use?” could belong together if their decision criteria match.
Similar language alone is insufficient. A question requiring all data to stay in the EU may lead to a different shortlist from one with no location constraint. A question about replacing an existing system may differ from a first purchase because migration is part of the decision. A reviewer should record the purchase task, material constraints, and reason for grouping each question. A similarity score can suggest candidates; a person decides whether the buying decisions are equivalent.
Freeze every question exactly as it will be sent, including punctuation and supplied context. Keep each wording_id attached to that text. If an editor changes the question, preserve the old version and register the new wording for a later round. This lets a reviewer distinguish an answer change from a prompt change.
The same answers can produce two useful totals
Consider this invented example: one platform, one round, and one answer for each wording. The measured outcome is whether the answer names the target brand. These numbers illustrate the calculation; they are not measured business results.
| Intent | Buying task | Wordings tested | Answers naming the brand | Within-intent rate |
|---|---|---|---|---|
| A | Choose a tool for a small team | 10 | 8 | 80% |
| B | Choose a tool under a location constraint | 1 | 0 | 0% |
| C | Replace an existing tool | 1 | 1 | 100% |
Within A, the wording results are A01–A08: brand named and A09–A10: brand absent. Each wording row still keeps its own frozen question and answer reference; the grouped display above is only a summary.
Across the twelve prompts, nine answers name the brand: 9 / 12 = 75%. That total gives every tested wording equal weight. Intent A therefore supplies ten of the twelve observations and strongly influences the result.
Under equal weighting for the three intents, the summary is (80% + 0% + 100%) / 3 = 60%. Each buying task contributes one third, regardless of how many paraphrases it has. The difference comes from the weighting choice, not from a change in the underlying answers.
Show both figures with their labels. The prompt figure describes performance across the tested wording set. The intent figure describes performance across the selected buying tasks under the declared rule. Neither estimates the number of independent buyers or the prevalence of those needs in a market. Equal intent weights are a measurement choice, not evidence that the three tasks occur equally often.
Declare the rule before collecting answers
Write a short specification for each round before running it. For a simple equal-weight design, use the following sequence:
- List the included
intent_idvalues and the frozenwording_idvalues belonging to each one. - Define the outcome separately for each metric, such as brand appearance or a linked citation. Calculate their scores separately.
- Give each intent
1 / number_of_intentsof the summary. Within an intent, give each wording1 / number_of_wordings_in_that_intentof that share. - Record the rule version and the collection conditions, including platform, language, and planned repetitions. State how repetitions produce each wording's result before averaging wordings.
Other within-cluster rules are possible. A team could weight wordings using independently recorded usage frequencies, but it would need a stated source and a rule fixed before collection. Writing more variations for a favored topic is not a frequency estimate. Without that evidence, equal wording weights within each intent are easier to inspect.
This choice sits alongside the denominators used in visibility reports. A weighting specification should refer to that existing calculation policy so readers can follow both parts of the score.
Keep the wording results and the history
Publish a row for every wording inside its intent cluster, with the frozen question, observed outcome, and answer reference. An 80% cluster rate in the example still contains two wordings whose answers did not name the brand. Readers need those rows to see whether a useful expression consistently exposes a gap that the average conceals.
Adding paraphrases in a later round changes the test design. Save the new membership and weighting version with that round. Leave earlier rounds on their original rules; do not redraw their trend using today's cluster assignments. If comparison across a change is necessary, run a separate comparison using shared frozen questions, the same cluster membership, and the same weights in both rounds. Label that comparison separately and keep the original series available.
For ongoing AI visibility monitoring, the practical deliverable is therefore the wording table, the intent summary, the prompt summary, and the versioned rule. When auditing an agency's AI visibility report, ask for those components together. They let a reviewer trace a summary to its answers and see how much influence each buying task received.
Want to see how AI engines describe your brand today?
Get a free growth audit