Audience and Market Research Synthesis

Audience research is the highest-return, lowest-risk thing you can point AI at, and it is the thing operators reach for least, because it produces nothing you can publish. There is no asset at the end, just a sharper read. That is exactly why it pays: the read is what makes everything downstream non-generic. A model handed customer language, real objections, and the specific pain your buyers keep naming does not regress to the bland center the way an empty prompt does. The research phase is where you acquire the proprietary context, and the generation phase inherits it. Skip the research and you are asking the model to invent your market from its averaged prior, which is the definition of slop produced faster.

The trap sits inside that same convenience. Ask a model to "summarize what customers think about onboarding" and it will hand you a clean, confident paragraph. Some of it is genuinely synthesized from the reviews you pasted. Some of it is the model filling gaps with the statistical average of everything ever written about onboarding — plausible, well-shaped, and not actually about your customers. The two are indistinguishable on the page. Both read as a finding. The whole discipline of this module is refusing to let them stay indistinguishable.

There is a measured reason the averaged version is so seductive. A 2023 HBS/BCG study of 758 consultants found that giving them AI raised average output quality by roughly 40% while narrowing the spread of ideas across the group — the work got better and more samey at the same time. That is the failure in miniature. The model lifts you to a competent middle and quietly removes the edges, and the edges are where a real market insight lives. A "trend" the model surfaces with no specific source behind it is usually that middle wearing the costume of a discovery: the consensus take, the thing every competitor's model also produces, dressed up as something you found. Ship it into a campaign and you have spent budget broadcasting the average of your category back at it.

So synthesis is not the act of asking the model to summarize. It is the act of forcing provenance. Every line in the read gets tagged: is this grounded in a signal I actually supplied, or is it the model's inference? A grounded claim cites the review or the call it came from. An inferred claim is a hypothesis — possibly true, not yet load-bearing. You can still act on hypotheses, but you act on them as bets to test, not facts to build on. The verified read is the subset that survived that sort.

Only once you have that verified read does angle generation earn its keep. Here the move inverts: you want quantity, not a single tidy answer. Ask for one campaign angle and the model gives you the obvious one — the angle your three nearest competitors are also running, because it is the center of the same distribution. Ask for eight or ten angles spanning different pains and audiences and you force it out toward the edges, where the non-obvious ones live. Then you select. Generation is cheap and convergent; selection is where your judgment does the work, keeping the angle that ties to a verified customer pain and discarding the ones that sound clever but trace back to nothing.

This whole approach assumes your signals are real signals. Feed the model fifteen five-star reviews you cherry-picked and the verified read faithfully synthesizes a market that does not exist — every claim is grounded, every claim traces to a quote, and the read is still wrong because the input was skewed. Verification confirms that a claim came from your sources; it cannot confirm your sources represent the market. Thin or biased input produces a confidently-sourced fiction, which is worse than an obvious guess because it passes the provenance check.

The verification step also has a real cost, and it is not always worth paying. For a quick gut-check before a low-stakes social post, tagging every line as grounded-or-inferred is overhead you do not need. The discipline earns out when a read is about to drive real spend, a positioning decision, or a launch — when shipping the averaged prior would be expensive. Match the rigor to the stakes rather than running the full sort on every casual question.

And the failure mode this module exists to prevent does not announce itself. There is no flag on the model's output marking which sentences are synthesized and which are invented. A fabricated "customers increasingly want X" looks exactly like a real one — same confidence, same clean phrasing. You catch it only by going to trace it and finding no source underneath, which is why provenance has to be a step you run, not a vibe you trust. The model will never volunteer that it made something up.