AI Market Research: Useful, and Confidently Wrong

Ask AI to research your market and you will get a document in about ninety seconds. Headings, a market size, competitor profiles, trends, an opportunity section. It will look like something a consultant charged four thousand dollars for. Some of it will be genuinely useful. Some of the numbers will not survive being checked. Knowing which parts fall into which category is the whole skill here.

What it is genuinely good at

Three things, and notice what they have in common.

Structuring the question. You get a sensible framework for what to investigate instead of a blank page, which is where most small-business research dies.

Summarising material you supply. A pile of customer emails, reviews, or survey responses becomes a set of themes. This is excellent and badly underused.

Listing what you have not thought to ask. Often the most valuable output, and the one nobody requests.

None of those require the model to know a fact about your market. They require it to organise and summarise, which is exactly where it is strongest.

Where it goes wrong

Ask	Reliability
Summarise these 200 customer emails	High
What should I be asking about this market?	High
How big is this market?	Low, verify every figure
What is happening in my city or region?	Lowest, the material is thinnest

The bottom row catches people out most. The more specific your geography, the less underlying material exists, and the more the model fills the gap with something plausible. A confident paragraph about your regional market is exactly the output to distrust.

Change the method for competitors

Do not ask what it knows about your competitors. That produces a blend of outdated facts and reasonable-sounding invention, and the two are indistinguishable in the output.

Give it the raw material instead. Their website copy. Their pricing page. Their job postings, which are surprisingly revealing about where a company is investing. Their recent reviews. Then ask it to compare, find patterns, and tell you what their positioning implies about who they are chasing.

That shifts the work from recall to summarisation, and everything in the answer is grounded in something you can point at. It is also the same principle behind why retrieval matters more than conversational polish: what the model can see beats what it can remember.

Open the sources

When a report cites a figure, click through. Not all of them, but every number you plan to act on.

You will find three kinds of problem. Sources that do not say what was claimed. Sources that are real but five years old, presented as current. And occasionally a citation that does not exist at all. This takes a few minutes and it catches almost everything, which makes it the highest-return habit in this entire post. Adding a follow-up question about what the model is least confident about, as in asking AI what it is likely to get wrong, will usually point you at the weakest claims before you start.

The best use, and almost nobody does it

You are sitting on years of research nobody has read. Customer emails. Enquiry forms. Reviews. Complaints. Notes on deals you lost.

Nobody has ever read that as a set, because it would take a week and there was never a week. Feed it in and ask what patterns appear, what people repeatedly ask for, and what words they use to describe the problem they came to you with.

That last one is worth the exercise on its own. The language your customers use is rarely the language you use, and closing that gap improves your website, your quotes, and your ads immediately. It is also the raw material for being found when people ask an assistant instead of a search engine, which we covered in optimising for AI search.

What it does not replace

Talking to ten actual customers. That remains the highest-value research available to a small business, and it is the thing people skip most reliably, because it is uncomfortable and cannot be done from a desk in an afternoon.

AI makes the preparation better. It will give you the questions worth asking and summarise what you heard afterwards. It cannot have the conversation, and the conversation is where the surprising things come from. For most businesses the honest comparison is not AI research against paid research. It is AI research against the nothing that was happening before, and against that bar it is an easy win.

Frequently Asked Questions

What is AI market research good at?

Three things, reliably. Structuring the question, so you get a sensible framework for what to investigate rather than staring at a blank page. Summarising material you supply, such as a pile of customer emails, reviews, or survey responses, which is genuinely excellent and badly underused. And generating the list of things you have not thought to ask, which is often the most valuable output of all. Notice that none of those require it to know a fact about your market. They require it to organise and summarise, which is where it is strongest.

Where does it go wrong?

Numbers, sources, and anything local. Ask for the size of a market and you will often get a specific figure with a plausible attribution that does not survive checking, because a confident number is easier to produce than an accurate one. Regional detail is worse still: the more specific your geography, the thinner the underlying material and the more it fills gaps. Treat every figure and every cited source as unverified until you have opened it yourself, which takes a few minutes and catches most of the problem.

Can I use it for competitor analysis?

Yes, with a change of method. Do not ask it to tell you about your competitors from memory, because that produces a mix of outdated facts and reasonable-sounding invention. Instead give it the raw material: their website copy, their pricing page, their job postings, their recent reviews. Then ask it to compare, find patterns, and tell you what their positioning implies. That plays to summarisation rather than recall, and the output is grounded in things you can point at.

Is it good enough to replace paid research?

For most small businesses the honest comparison is not against paid research, it is against doing nothing, which is what usually happened. Against that bar it is a clear improvement. It does not replace a proper study when a decision is large enough to justify one, and it is a poor substitute for talking to ten actual customers, which remains the highest-value research available to a small business and the one people most reliably skip because it is uncomfortable.

What is the single best use?

Turning material you already own into something you can act on. Most businesses are sitting on years of customer emails, enquiry forms, reviews, complaints, and lost-deal notes that nobody has ever read as a set. Feed that in and ask what patterns appear, what people repeatedly ask for, and what language they use to describe the problem. That is real research on real data about your actual market, and it costs an afternoon rather than a consulting engagement.