AI Prompts for Market Research in Competitive Categories

Every market figure ever published is the answer to a question of the form: if you count these products, sold to these buyers, in these countries, how much money is that. The counting is arithmetic. The three lists in front of it are judgements, and they are what decides whether the number is ten billion or three hundred.

Those judgements almost never travel with the figure. A number lifted into a deck arrives without the definition that produced it, gets compared with a second number produced under a different definition, and the comparison looks sound because both are expressed in dollars. This is the single most common defect in market research and it has nothing to do with how the research was done.

Which is why the first prompt here is about definitions

If you take one habit from this page, take this one: before asking what a market is worth, ask each of your sources what it thinks the market is. The answers will differ, the differences will explain most of the numerical spread, and you will have learned something that no single figure could have told you.

Three of these four are files. The second is a decision, and it is the one that most changes what comes back.

The fourth input is what separates a market read from a literature review. Published material is available to every competitor you have, and a document assembled entirely from it is a document any of them could assemble too. Your survey responses, your support tickets and your lost deals are the only part nobody else can reproduce.

The first is the one to run if you only run one. The second and third exist because of the two failures that make market research quietly wrong: sources that define the category differently, and sources that agree because they are copying each other.

None of them asks for a recommendation. A market read that arrives with a strategy attached has skipped the step where you decide whether the evidence supports one, and the strategy is the easiest part to write once the three lists exist.

We asked a model how big the CRM software market is, took the answer, and then opened two research firms’ published pages for the same named market in the same year.

A factor of two and a half, one market name, one year, and both firms are behaving reasonably. Read the right-hand column and the gap stops being mysterious: one report is counting a service category and the other is counting most of a go-to-market stack. Neither is wrong. They are not measuring the same thing.

Why the confident single figure is the dangerous one

The model’s own answer was not absurd. It sat in the region people quote, expressed with the usual growth rate, and would have passed unchallenged in most rooms. That is precisely what makes it worse than an obvious error: false precision survives review, and an obvious error does not. Nobody asks the follow-up question about a figure that sounds about right.

The second thing that goes wrong in a market read is subtler than a boundary dispute. You gather twelve documents, a claim appears in seven of them, and seven feels like weight. It usually is not. Trade coverage cites analyst summaries, vendor pages cite trade coverage, and a claim that originated once acquires the appearance of consensus by moving.

This matters more than it sounds, because the claims that circulate hardest are the ones most useful to whoever first published them. A figure that makes a category look large gets repeated by everyone selling into it. Tracing the chain is unglamorous and it routinely removes the most quotable line in the folder.

What survives the check is usually narrower and better

A read that has been through this comes out shorter. Two or three claims stand with named first-hand sources, a handful become supported conclusions you can defend, and a surprising amount moves to the unsupported list. That third list is not a failure of the research. It is the part that tells you which of your plans currently rests on nothing, and it is the only output here that changes a decision.

Everything above describes one sitting with one folder. The folder ages from the moment you close it, and market material ages unevenly: a category definition holds for years, a vendor list holds for months, and a funding or acquisition claim can be wrong the week after it was written.

That unevenness is what makes a market read hard to maintain. Nothing in the document marks which claims decay fast, so a read that was accurate in March is still being quoted in October with no way to tell which parts have expired. The figures look exactly as authoritative as they did when somebody checked them.

No amount of watching produces a market size. The denominator is not a thing that happens in public, and the honest version of that figure stays a range with its definition attached, which is a better artefact than the confident number it replaces.

Flares watches the companies in your category and dates every change it finds, so the fast-decaying half of a market read gets refreshed by what actually happened rather than by a calendar reminder. The boundary, the segments and the judgement about what the evidence supports stay yours.

So the read that survives is the one built the way this page argues for: a boundary at the top, three lists underneath, and a source and a date beside every figure. That structure is what lets somebody refresh the third of it that decayed without re-reading the rest.