Prompting for Market Research — Getting a Useful Answer, Not a Confident One

A model answers by producing what usually follows a question like yours. When the question is broad, what usually follows is a broad answer: a paragraph on the business, a paragraph on the sector, a nod to growth drivers, a nod to risks. It is a shape, filled with the most typical content that fits the shape. Nothing is being retrieved from a document, because you did not give it one.

The trouble is that this output is indistinguishable from real extraction. It has the same confident register, the same tidy structure, sometimes the same specific-sounding numbers. There is no visual difference between a figure read off page 118 of a filing and a figure the model produced because a number of that magnitude is what usually appears in that sentence position.

The deeper problem is that a vague prompt cannot be graded. If you ask for "an analysis" and receive an analysis, on what basis do you reject it? You have no criterion. Precision in the question is not politeness towards the machine — it is what gives you the standing to say the answer is wrong.

So the first move in every research prompt is to narrow until the question has exactly one correct answer, or a small set of them. Which company. Which document. Which period. Which metric. Once a question is that narrow, a wrong answer becomes visible, and visible wrong answers are the only kind you can defend against.

Every prompt that reliably produces usable research has the same four components. They are not stylistic preferences. Each one closes a specific route by which invention gets into the output, and removing any one of them reopens that route immediately.

Scope tells the model exactly what it is being asked about, so the answer is about one thing rather than an average of many. Source supplies the text the answer must come from, so it is reading rather than recalling. Structure fixes the shape of the reply, which forces specificity and, more usefully, makes gaps visible. Permission tells the model that reporting an absence is a correct answer rather than a failure.

The order matters as much as the presence. Scope and structure go above the pasted source. The source sits in the middle, clearly fenced. Then the permission rule and a one-line restatement of the question go underneath, so the last thing in view before the model starts writing is what you actually asked.

Once you have written this once, you never write it again. It becomes a template with four or five blanks in it. The work in a research session shifts from composing prompts to selecting which section of a document to paste, which is where the work belongs.

Scope has four dimensions and dropping any of them lets the answer wander. The entity: the full registered name, not a ticker or a nickname, because several listed Indian companies share a first word and a model will happily blend two of them. The document: an annual report, a quarterly result, a concall transcript, an exchange filing. The period: the financial year or the specific quarter. And the metric: the exact line item you want, in the words the document itself uses.

That last point is worth dwelling on. Financial statements use precise terms — revenue from operations, other income, finance costs, profit before exceptional items. Asking about "profit" when the statement reports four different profit lines forces the model to pick one for you, silently. Ask for the line by the name it carries in the document and the ambiguity disappears.

Scope also decides what a good answer looks like before you see one. If you asked for segment revenue for FY24 from the pasted note, you know in advance that a correct reply is a short list of segments with figures and a note reference. Anything longer than that is padding, and anything without the note reference is unsourced. You are grading against a standard you set, not reacting to whatever arrived.

The test for whether your scope is tight enough is simple. Read your prompt back and ask: could two different but reasonable answers both satisfy this? If yes, narrow it further before you press enter.

A model’s training has a cutoff date, after which it has seen nothing. It also does not hold a clean, indexed copy of the documents it trained on. What it has is a compressed statistical impression of enormous quantities of text. Asking it to recall a specific figure from a specific Indian company’s filing is asking it to reconstruct a detail from an impression, and reconstruction produces plausible detail rather than correct detail.

Pasting the source changes the task from recall to reading. This is the single largest quality difference available to you, and it is bigger than any refinement of wording. A mediocre prompt over a pasted document beats an elegant prompt over the model’s memory, every time and by a wide margin.

Fence the source clearly. Put a marker line before it and after it, and tell the model explicitly that the answer must come from between those markers and nowhere else. Without the fence, the model blends what you pasted with what it half-remembers, and the join is invisible in the output.

Where the document is genuinely too long to paste, split it and paste one section per prompt rather than accepting a summary built on recall. A partial answer from a real source is worth more than a complete answer from an imagined one, because the partial answer tells you honestly what it does not cover.

Ask for a paragraph and you will get a paragraph, and a paragraph can absorb an enormous amount of vagueness without looking vague. Transitions cover missing links. Qualifiers cover missing figures. The prose reads smoothly precisely because the gaps have been smoothed over.

Ask for a table with fixed columns and the gaps become holes you can see. A column headed "figure" that is empty is obviously empty. A column headed "page reference" with nothing in it tells you instantly that the claim beside it is unsourced. Structure does not make the model more accurate — it makes the model’s inaccuracy visible to you, which is the part you actually need.

Fixed headings do the same work for qualitative material. If you always ask for the same five headings from every concall transcript, then a heading that comes back thin is a signal about the transcript rather than about your prompt. Consistency of structure is what turns a series of one-off answers into something you can compare across quarters and across companies.

Add one more heading than you think you need: a section for what is missing. Ask explicitly for the things you would normally expect in a document of this type that are absent from this one. An empty "missing" section is either good news or a lazy answer, and you will learn quickly which one you are looking at.

This is the highest-leverage sentence in any research prompt, and it is the one people leave out. Written into the prompt, it says: if the answer is not in the text I gave you, reply "not stated in the source" rather than producing something.

The reason it matters so much sits in how these systems are built. A model is optimised to produce a helpful, complete-looking response. Across the enormous body of text it learned from, a confident answer is overwhelmingly what follows a question, and an admission of ignorance is rare. Later training stages reward responses people rate as useful, and a blank is rarely rated useful. The result is a strong default pull towards answering — "I do not know" behaves like a failure state to be avoided rather than a legitimate output.

You are not fixing that tendency. You are overriding it locally, for this prompt, by redefining what counts as success. When the prompt states that reporting an absence is a correct answer, the absence stops being a failure and becomes one of the things you asked for. That is a change in the task, not a change in the model.

Say it in the imperative and be specific about the format. "If it is not in the source, write: not stated in the source." Then add the two clauses that close the side doors: do not estimate it, and do not supply it from memory. Without those, you sometimes get a helpful approximation instead of an honest gap, which is the exact failure you were trying to prevent.

Add one column to any extraction and the whole exercise changes character: the exact sentence the claim came from. Not a paraphrase, not a summary — the verbatim string, plus the page or note reference where it sits.

This does two jobs at once. It gives you something to verify against, so checking a claim means finding one quoted sentence in the document rather than re-reading a section. And it acts as a filter at the moment of generation. A claim that has no source sentence has nowhere to put its quote, which makes an unsupported row conspicuous instead of invisible.

It is not a guarantee. A quote can itself be fabricated, and a fabricated quote from a document you have in front of you is trivially caught — you search for the string and it is not there. That thirty-second search is the whole verification, and it is only possible because you asked for the quote.

Concall transcripts are where this pays off most. The prepared remarks are written to be reassuring and are usually the least informative part of the call. The analyst question-and-answer section is where the specifics live: someone asks about a margin decline directly, and management either answers it or visibly does not. Quoting the exchange preserves that distinction, which a summary would erase.

When an answer is not quite right, the instinct is to rewrite the whole prompt and try again from scratch. That is usually the wrong move. The source is already in the thread, the structure is already established, and a follow-up question inherits both. Rewriting throws away context you paid for.

Narrow instead. If the extraction is too broad, ask for one row of it in more detail. If a figure lacks its comparative, ask for the comparative from the same source. If a claim has no quote, ask for the quote and watch what happens — a request for a quote that cannot be met is how an invented claim announces itself.

Reserve the fresh start for two situations. When you are switching to a different document, because a new source in an old thread competes with the one already there. And when a wrong answer has gone uncorrected for several turns, because by then it has settled into the transcript as established context and follow-ups will build on it rather than question it.

The most valuable follow-up in market research is the comparison. One quarter’s extraction tells you the state of things. Two quarters in the same format tell you the direction, and direction is what a single snapshot can never give you. Because both extractions used an identical structure, the comparison is mechanical rather than interpretive — which is exactly the kind of work to hand to a machine.