Why is the first question of a discovery call your best prompt?

Because a spoken first question is the only artefact in prompt research that carries a person, a date and an uncontaminated vocabulary at the same time. A search query has the date and a partial identity but arrives pre-compressed into search grammar; a community thread has the words but not the person you sell to. The first question on a call is the sentence a buyer used while still describing the problem in their own terms, which is exactly the register people type into an assistant: long, conversational, and phrased as a question rather than as keywords.

The question form is also the form that most reliably produces an AI answer. A study of 55,393 trending queries collected between 13 March and 21 April 2026 measured Google AI Overview activation at 64.7% for question-phrased queries against 9.5% for non-question queries.¹ Call questions arrive already in that shape, which is why they need so little processing: a human being did the work of turning a keyword into a prompt without knowing they were doing it.

There is a third reason, and it is the one that changes what you do next. The same measurement found that 29.8% of the domains an AI Overview cites do not appear anywhere on the corresponding first page of results.¹ A vendor study of one assistant’s most-cited pages points the same way, at a number nobody outside that vendor can check.² Either way, a research method that starts from keyword data cannot see that population. Calls can, because they start from the person rather than from the index.

Apply five edits and no others, keeping the verbatim original in the next column so any later reader can check what you changed. Remove identifiers: names, companies, invoice numbers, anything that identifies the speaker. Drop disfluencies: the “um”, the false start, the repeated word. Resolve pronouns to the noun the speaker had just used, not to the noun you would have used. Split compound questions into one entry each, because a two-part question produces an answer you cannot score. Keep everything else, including their word for your category when it is wrong, because that word is what a real person would type.

What you must not do is the sixth edit everyone wants to make: translating the question into your product vocabulary. “Can I do this without getting my account banned?” becomes “What are the account safety controls?”, and the entry stops measuring anything. The rewritten version retrieves your own documentation because official sources are the largest single category of cited source, at 34.22% to 46.35% depending on platform, in a study of 602 controlled prompts producing 21,143 search-layer citations.⁴ The original retrieves whatever the open web says about bans in your category, which is the thing you actually wanted to know.

Store both columns permanently. Six months later, when the panel produces a number somebody disputes, the verbatim column is the only evidence that the prompts are real questions rather than your team’s idea of them, and an unauditable panel gets discounted by exactly the person you built it to convince.

Saturation is a borrowed concept with a real measurement behind it, and the borrowing is what makes the stopping rule defensible rather than arbitrary. The best-known experiment on the question, a 2006 study in Field Methods analysing sixty in-depth interviews, reported that “saturation occurred within the first twelve interviews, although basic elements for metathemes were present as early as six interviews”.⁵ That study was about a different subject in a different decade, so treat it as a prior rather than a target: the right order of magnitude is a dozen, not a hundred, and the stopping rule is an observed flattening rather than a fixed number.

Plot the curve yourself: three spreadsheet columns, being new distinct questions per call, cumulative distinct questions, and calls processed. Two consecutive batches of five that each add fewer than one new question means the source is exhausted for now. That is also when the useful output changes, because the counts of how many buyers used each phrasing become the more valuable half of the dataset.