Prompt Research

Prompt research is the step that decides what gets measured. Prompt tracking runs a defined set of questions on a schedule and records what the answers say about a brand; prompt research is the work of assembling that set in the first place. The distinction sounds procedural and is not, because a tracking set is a claim about what matters. Every prompt in it will be run repeatedly, will consume budget, and will contribute to whatever headline citation rate the dashboard reports. A set assembled carelessly does not produce a slightly worse number — it produces a number that is measuring something other than what you meant, and reports it with the same confidence.

The failure this discipline exists to prevent is the flattering set. It is very easy to fill a tracking set with questions containing your own brand name, because those are the questions you can think of and the ones you win. An engine asked about a brand it has read about will usually describe that brand, so a set built this way tends to return a high citation rate that is close to meaningless: it measures whether the engine knows you exist, which you already knew, rather than whether it recommends you to someone who does not. The prompts that carry information are the ones a buyer would type before they have heard of you — category questions, "best tool for" questions, and comparison or alternatives questions, where the engine is most likely to volunteer a list and most likely to leave you off it.

Real search demand is the most defensible starting point, because it is evidence rather than imagination. Search Console holds the questions people already use to reach you, and a meaningful subset of them are phrased as questions rather than as keyword fragments. Filtering for that shape — an interrogative opener as the first word, or an explicit question mark — surfaces candidates whose wording already resembles what someone would put to an assistant. It is worth being precise about what this filter is and is not: it is an inference from the wording of a query, not a detection of where the query came from. Search Console does not report whether a search originated from a conversational interface, so any tool claiming to show your "AI queries" from that data is describing a proxy. The proxy is a good one, and it is still a proxy, and knowing which one you are holding matters when you present the number to someone else.

Word count is a tempting second signal and a poor one. Long queries feel conversational, but a seven-word query is just as often ordinary long-tail phrasing — "best project management software for small teams" is not a prompt, it is a keyword. More importantly, a length threshold cannot be explained to the person reading the report: there is no answer to "why is six words not AI-style but seven is" that survives contact with a client. Signals that select a tracking set should be ones you would be willing to defend out loud, which in practice means a small number of explainable rules rather than an opaque score.

The second source of candidates is your own content, approached from the opposite direction. Rather than asking what people search, ask what your existing pages already answer well, then check whether engines credit you for those answers. A page that states a definition cleanly and early is a page an engine can quote, and the question it answers is a prompt worth tracking precisely because you have a plausible claim to it. This is a different kind of candidate from a demand-derived one: demand tells you what is worth winning, content tells you what is winnable now, and a healthy set contains both. The gap between them — questions with demand that none of your pages answers cleanly — is the content backlog, and it usually turns out to be the most valuable output of the whole exercise.

Prioritisation matters because tracking sets are bounded in practice. Every scheduled prompt is a repeated cost, and a set large enough to cover everything is a set nobody reads. Ranking candidates by the demand behind them is the usual first cut; excluding malformed ones is the necessary second. Queries carrying search operators, leading digits, or excessive length will be answered by an engine, but the answer will not mean anything, and they quietly depress a citation rate by sitting in the denominator returning nothing forever. Filtering those before they enter the set is not tidiness — it protects the honesty of the headline figure.

TriRank builds prompt research into the surface where the tracking set is edited rather than treating it as a separate exercise. Where a site has Search Console connected, candidate queries are drawn from real impressions, sorted by the demand behind them, filtered to exclude what is already tracked and what is malformed, and offered as suggestions next to the tracking set itself. The keywords view separates queries whose wording reads as a question from the rest, and labels them as exactly that — question-shaped demand — rather than dressing them up as AI queries, because a proxy presented as a measurement is a number nobody can defend. The free audit runs a representative prompt against the engines with no account and no tracking set at all, which is often the fastest way to see whether this line of work is worth starting on a given domain.