How to Do Market Research With AI Without Making It Up

Ask a language model for the size of the UK pet grooming market and it will give you a number. It will be plausible, specific, and delivered with complete confidence. It may also be entirely fabricated.

This is not a flaw that better prompting solves. It is a consequence of how the models work: they produce likely-sounding text, and a specific figure is more likely-sounding than an admission of ignorance. The failure is silent by design.

Which is a problem, because AI is genuinely good at the parts of market research that are slow — reading widely, spotting patterns, structuring findings. The task is separating those from the parts where it will confidently invent.

The dividing line

There is a clean test, and it holds up better than any prompt technique.

Can the model verify this, or is it recalling it?

A model asked to summarise ten competitor pricing pages it has just been shown is doing verifiable work — the source is in front of it. A model asked what those competitors charge, from memory, is producing a guess shaped like an answer.

That last row deserves emphasis. Models fabricate citations readily — a real-sounding journal, a real-sounding author, a report that does not exist. If a model hands you a source, the source is a hypothesis until you open it.

Five research jobs, and how to run each one

1. Competitor pricing and positioning

Don't ask what competitors charge. Give the model the pages.

Collect the pricing and homepage copy of eight to twelve competitors, paste them in, and ask for a comparison table: tiers, prices, what's included at each level, and what language each uses to describe who it's for.

The output is genuinely useful and genuinely grounded — every cell traces to text you supplied. What emerges is usually the positioning gaps: the segment nobody addresses, the feature everyone charges extra for, and the price band that is conspicuously empty.

Time: roughly an hour of collection, minutes of analysis. Manually it's most of a day.

2. Customer language

The most under-used research source is text your customers have already written: reviews, support tickets, sales call notes, Reddit threads, comments on competitor posts.

Gather several hundred and ask the model to identify recurring complaints, the words people use for the problem, and the moment they decided to look for a solution.

This is pattern-finding across a corpus — exactly what these systems do well, and it is the difference between writing "streamline your workflow" and writing the phrase your customers actually typed.

3. Market sizing

The most dangerous request, and the one most people make first.

Sizing from memory is unreliable, but sizing is not impossible — it just has to be built rather than recalled. Use published figures you can source yourself — national statistics offices, industry associations, company filings — then have the model do the arithmetic and hold the logic.

A defensible bottom-up structure:

Businesses in the category × share matching your customer profile × realistic annual spend

Where:

• Businesses in the category comes from a national business register.

• Share matching your customer profile is your estimate and should be clearly labelled as such.

• Realistic annual spend comes from your own pricing or published benchmarks.

Every input is traceable. Every assumption is visible and arguable. That is what makes it usable in a plan or a pitch — not the size of the number, but the fact that a reader can dispute a specific line of it.

A number you cannot source is not research. It is a placeholder that will eventually be quoted back to you.

4. Finding competitors you don't know about

Direct competitors are easy; you already know them. The ones that matter are the alternatives your customers choose instead of anyone in your category — spreadsheets, an agency, a freelancer, or doing nothing.

Two grounded approaches:

• Search the phrases your customers would use — the problem, not the product category — and record everything that ranks. Ranking is evidence of demand.

• Ask ChatGPT, Claude, and Perplexity what they would recommend for that problem, and record what they name. Increasingly this is the shortlist your buyer sees.

Both produce a list from live sources rather than recall.

5. Trend and demand validation

Do not ask a model whether a market is growing. Ask it to interpret data you provide.

Google Trends exports, search volume from a keyword tool, or your own analytics are all fine inputs. The model reads the shape and explains it; the shape came from somewhere real.

Verification, briefly

Three rules cover most of the risk.

Every number gets a source before it leaves the document. Not a source the model named — one you opened.

Ask for the reasoning, then check the reasoning rather than the answer. A bad chain is much easier to spot than a bad number. "Where does this figure come from and what would make it wrong?" is the most useful follow-up available.

Give the model permission to have no answer. Adding "if you don't have a reliable source for this, say so" to a prompt measurably increases the rate at which it declines — the default behaviour is to answer regardless, because that's what the training rewards.

A workable sequence

For a small business or a solo founder, a full research pass is about a day.

Every step has a named source. Nothing in the finished document rests on the model's memory.

What good output looks like

A market research document you can defend has three properties, and they're easy to check.

Every figure has a source you could click. Every assumption is labelled as one. And the conclusions follow from the evidence rather than arriving alongside it.

A document that reads smoothly but fails all three is the standard output of asking a language model to "do market research for my business." It looks like the real thing, which is precisely the risk — plausible research is more dangerous than obviously bad research, because it survives review.

Questions