How to Write a Market Research Summary Using AI

What a Real Summary Owes the Reader

A market research summary is a decision document, not a compression exercise. Its job is to let someone who has read none of your source material make a call — invest, wait, or walk — and understand what would change that call. A shorter version of the research is not the same thing.

Most professionals get handed the wrong template here. The instinct is to condense: fewer words, same shape as the original. But your reader does not need the report abridged. They need the decision extracted, which is a different operation entirely — closer to the discipline behind a good executive summary written with ChatGPT than to a book report.

The difference is not writing quality. It is that the second version commits to something, and every commitment is traceable back to a source you can produce if challenged.

Ask yourself what your reader will do within an hour of finishing it. If the honest answer is “nothing, but they will feel informed”, the summary has failed regardless of how good it looks. A brief that changes nobody’s behaviour is a status update wearing better clothes.

The Four-Block Brief, Explained

The Four-Block Brief is a market research summary built as four decision blocks — Market Size, Customer Segments, Competitor Position, and Viability Verdict — where every line carries a source tag, and anything the source material does not support gets flagged rather than filled in.

The blocks run in that order for a reason. A verdict written before the evidence is an opinion looking for support, and the model will happily supply that support. Ordering the work forces the conclusion to arrive last.

Most of what people get wrong about how to write a market research summary using AI is a sequencing problem rather than a prompting one. They ask for the whole brief in one request, the model produces a verdict in paragraph two, and everything after that quietly bends to justify it.

How to Write a Market Research Summary Using AI

The workflow is four moves: prepare the source material, set the rules before you ask anything, run one block at a time, then verify. Skipping step two is the single most common reason these summaries come back full of confident nonsense.

Sort your files into the four blocks before you open a chat. Industry reports feed Market Size. Survey exports and interview notes feed Segments. Competitor pages and pricing screenshots feed Competitor Position. Anything that fits none of them probably is not needed.

Then strip what should not leave the building: customer names, account IDs, internal revenue figures, anything a colleague would wince at. If you are working across several long documents, chunking beats cramming — the approach in our guide to synthesising multiple documents into one brief applies directly here.

This is the part almost everyone skips. Before you ask a single question, you tell the model what it may and may not do with your files. These are constraints, not politeness, and they work because they give the model something concrete to do when it hits a gap — the mechanic behind negative constraints in prompting.

You are acting as a market research analyst preparing a brief for a leadership team.

Rules for this entire conversation, applied to every response:

1. Use ONLY information contained in the files and text I provide. Do not use your own training knowledge about this market, these companies, or this industry.
2. Every factual claim must end with a source tag in square brackets naming the file and the section or page it came from.
3. If I ask for something the source material does not contain, write [NOT IN SOURCE] and move on. Do not estimate, infer a range, or offer a typical industry figure.
4. Quote customer language exactly. Never paraphrase a quote and never merge two quotes into one.
5. No marketing adjectives. Every sentence must contain a fact, a number, or a quote.

Confirm you understand these five rules, then wait. Do not summarise anything yet.

Wait for the confirmation before uploading anything. It sounds fussy. It changes the output more than any other single thing in this workflow.

Ask for all four blocks at once and you get four shallow blocks. Ask for one, review it, then move on. Market Size goes first because it is the block most likely to attract invented numbers, and you want to catch that pattern early rather than after three more prompts have built on it.

Using only the attached files and the five rules above, write the Market Size block.

Give me, in this order:
- Total market size, with the figure exactly as stated in the source and a source tag
- Growth rate or trajectory, with a source tag
- The time period the figures cover
- Any figure the source implies but does not state, listed separately under "Inferred, not stated"
- Anything I asked for that is missing, listed under "[NOT IN SOURCE]"

Maximum 120 words plus the lists. No introduction, no conclusion.

Open each source file and check every tagged figure against the page it claims to come from. This takes ten minutes and is the difference between a brief and a liability. Our walkthrough on fact-checking AI research before you publish it covers the triage: check every number, every quote, and every claim about a named competitor. Everything else can pass.

Prompts for the Other Three Blocks

Each of these assumes the five rules are already set and the relevant files are attached. Run them in sequence, reviewing between each.

Using only the attached responses and the five rules above, write the Customer Segments block.

Identify no more than four segments that actually appear in this data. For each one:
- A short label describing who they are
- The problem they describe, in their words, with two exact quotes
- How many responses fall into this segment, as a count not a percentage
- One thing this segment says that contradicts the others

If fewer than four distinct segments appear in the data, return only the ones that do. Do not create a segment to fill the list.

Counts rather than percentages is a deliberate choice. Percentages invite rounding, rounding invites estimation, and estimation is exactly what you are trying to keep out. If you are working from survey exports, the prep steps in our guide to analysing survey results with ChatGPT will save you a pass.

Using only the attached competitor material and the five rules above, write the Competitor Position block as a table.

Columns: Competitor | Who they target | What they charge | What they do not offer | Source tag

One row per competitor. Leave a cell as [NOT IN SOURCE] rather than filling it from general knowledge — this includes pricing, which changes often and is frequently wrong in training data.

After the table, list any gap in the market that is visible from these rows alone. If none is visible, say so.

That last instruction matters more than it looks. Asked to find a gap, a model will always find one. Given permission to report that no gap is visible, it sometimes takes it — and that answer is worth more than a manufactured opportunity, particularly if you are building on the same source material used for competitor research with ChatGPT Deep Research.

Using ONLY the three blocks we produced above and nothing else, write the Viability Verdict block.

Structure:
1. The verdict in one sentence: proceed, wait, or do not proceed
2. Three reasons, each pointing to a specific line from blocks 1 to 3
3. The two biggest risks, each with the evidence that suggests it
4. What would have to change for the verdict to flip
5. What we would need to find out next, and where that data would come from

If blocks 1 to 3 do not contain enough to support a verdict, say that instead of producing one.

Point five is the one leadership actually reads twice. It turns your summary from a full stop into a next step, and it is where a structured SWOT analysis built with ChatGPT often earns its place as a companion document.

Why AI Invents Market Numbers

Language models complete patterns. Asked for a market size when the source material contains none, the model does not experience a gap — it produces the kind of sentence that usually follows that question, populated with a figure that looks plausible for that industry. It is not lying. It is doing the only thing it does, applied to a question it should have refused. The mechanics are worth understanding in full, and our explainer on what AI hallucination actually is covers them.

The fix is not a better adjective in your request. It is giving the model somewhere to put a gap.

“Do not make anything up” is an instruction with no fallback behaviour attached. “Write [NOT IN SOURCE]” is an instruction the model can follow. The same principle drives every element of the Role, Goal, Context, Format structure we use across the site — be specific about the output, not just the intention.

What This Workflow Cannot Do

The Four-Block Brief makes AI a fast, disciplined synthesiser. It does not make it an analyst, and a summary that reads well is not the same as a judgement you can stand behind in a room.

There is also a quieter limit worth naming. AI is very good at making thin research look substantial. Four well-structured blocks built on two weak sources will read more convincingly than a messy handwritten analysis built on twelve good ones — and the polish actively hides the problem. If your source material is thin, the honest move is to say so in the brief, not to let the formatting cover for it.

Before You Paste Anything Sensitive

Market research files are among the most sensitive documents most professionals handle: unreleased strategy, customer names, pricing nobody outside the company has seen. The rule that matters is not free versus paid. It is consumer account versus business account — a distinction almost every guide on this topic gets wrong.

On consumer plans, conversations may be used to improve the models unless you change a setting yourself, and that setting sits in the same place whether you pay nothing or twenty dollars a month. OpenAI documents the toggle and its scope in its data controls documentation, and Google sets out equivalent handling for its own assistant in Google’s Gemini Apps privacy hub. Business, enterprise, and education accounts are the ones where no-training is the default rather than a checkbox.

Check the setting before your first upload, not after. Opting out applies going forward only — it does nothing about material already sent. If your company has an approved workspace, use it, and read our fuller treatment of whether ChatGPT is safe to use at work before you make that call alone.

When the Output Goes Wrong

Most failures in this workflow come from four causes, and all four have a fix that takes under a minute.

The context problem deserves a note of its own, because it drives the tool choice more than any feature comparison will. Free tiers carry a far smaller working context than paid ones — enough for pasted text and short extracts, not enough for a sixty-page report and three hundred survey responses in one sitting.

Run this once end to end on a project you have already finished, where you know what the right answer turned out to be. You will see immediately which blocks your source material genuinely supports and which ones it never did — and that is a more useful thing to learn on an old project than a live one.

• Save the five rules somewhere you can paste them in two seconds
• Keep the four block prompts together as one reusable set
• Keep the source files until the decision they informed has been made

Done properly, how to write a market research summary using AI stops being a question about tools at all. It becomes a question about whether you can point at any line in your brief and say exactly where it came from — which is the only test that has ever really mattered.