Research you can check, not a report you have to trust

Read what people say, not what a survey made them say

Surveys ask people to describe their behaviour from memory, in the vocabulary the survey supplied. What people write unprompted, in forums, reviews, support tickets and comment threads, is a different and usually more accurate record.

It is also written in the words they actually use, which is directly useful. The phrase a customer types is the phrase that should appear on the page, and it is rarely the phrase the industry uses internally.

• Forum and community threads where people discuss the problem without a vendor present
• Reviews, especially the three-star ones, which explain rather than praise or vent
• Your own support tickets and sales calls, which are the most underused research asset in most businesses
• Search and autocomplete data, which shows the question before anyone has been sold to

Competitor analysis beyond a feature grid

A feature comparison is the least useful competitor output, because features converge and the grid is obsolete in a quarter. The durable questions are about positioning and evidence.

Who do they say they are for. What do they claim, and what can they actually prove. Where does their own audience contradict them in public. Which questions do they refuse to answer, since a competitor's silence is usually a gap worth occupying.

Where AI helps and where it fabricates

AI is genuinely strong at reading volume. Ten thousand comments clustered into themes, a year of tickets summarised by root cause, thirty competitor sites compared on a consistent set of questions.

It is weakest at exactly the thing research reports are judged on, which is confident quantification. A model asked for a market size will produce a plausible figure with no basis whatsoever. The discipline is to distinguish what was read from what was inferred, and to label the difference in the output rather than in a footnote.

A number in a research report without a source is not a finding. It is a guess that has been formatted.

What a usable research output looks like

It states what was actually observed, with sources that can be opened. It separates observation from interpretation. It says what remains unknown and what it would take to know it. And it ends with decisions rather than themes.

The last part is the one usually missing. A report that concludes with insights leaves the reader to do the work again. A report that concludes with a recommendation and the evidence for it is a report that changes something.

What a study costs, and what decides it

Research is bought as a project. It has a question, a method, a deadline and a deliverable, and when the deliverable lands the engagement is finished. Some businesses then buy a standing arrangement to track the same measures quarterly, but that is a second decision taken after the first study has proved it was worth having.

Three things set the figure. How many sources have to be read, since a study covering one market in English is a fraction of the work of one covering three markets in three languages.

Whether the evidence can be gathered from public material or requires talking to people, because interviews are the expensive part and usually the valuable part. And how much of the analysis has to be traceable back to a quote or a page, since traceability is work and is also the only thing that makes a report checkable.

Before quoting, the question has to be written down in a form that could be answered wrongly. Vague briefs produce expensive documents nobody uses. A provider who accepts a brief like understand the market, without narrowing it first, is selling pages.

• How many markets and languages are in scope
• Whether interviews are needed, or public sources are enough
• How much of the output must be traceable to a named source
• Whether this is a one-off study or a measure tracked over time

When research is the wrong purchase

If the decision has already been taken, research commissioned afterwards is expensive reassurance. It is worth saying so before starting rather than producing a document that agrees with whatever was going to happen anyway.

If the category is genuinely new, there may be nothing to read. Talking to a small number of real potential buyers will beat any amount of desk research, and no volume of reading substitutes for the first conversation.

How long a study takes, who reads it before you do, and what you keep

A study runs on a stated clock. The first week is scoping and collection: questions agreed in writing, sources named, the corpus pulled and its size recorded. Analysis and first findings follow inside a fortnight. A research engagement that cannot say what it will have collected by a given date is not a study, it is a subscription.

Nothing reaches you unread. Every quantitative claim waits for a person to trace it back to the row or the post it came from, because a model summarising a corpus will produce a plausible sentence about a pattern that is not in the data, and the only defence is a human being checking the number against the source. Where a figure cannot be traced it comes out of the report, and the gap is stated instead.

Haad, the co-founder who owns growth and client solutions, scopes this work with whoever has to make the decision the research is for. Wobble is answerable for the method rather than for the conclusion being comfortable, and the method is published alongside the output.

You own the corpus, not only the document. The raw collection, the queries used, the dates and the scripts are handed over in your own accounts, so a study can be re-run in a year and compared rather than commissioned again from nothing. That is also the case for running it in-house: if you employ an analyst, buying the collection layer and the method and letting your own team do the reading is often better value than buying a finished report.

Common questions

How is AI market research different from a traditional research report?

It reads far more raw material than a person can, so it draws on what customers wrote unprompted rather than on what a survey prompted them to say. The trade-off is that it must be held to sourcing discipline, because a model asked to quantify something will produce a plausible number with no basis.

Can AI estimate market size?

It can produce a figure, and that figure should not be trusted without a source behind it. Market sizing needs stated inputs and assumptions that a reader can disagree with. A number generated without them is a guess that has been formatted to look like a finding.

What sources are worth reading for customer language?

Community and forum threads where the problem is discussed without a vendor present, reviews and particularly three-star ones which explain rather than praise or vent, your own support tickets and sales calls, and search and autocomplete data which shows the question before anyone has been sold to.

What should a competitor analysis actually cover?

Positioning and evidence rather than features, because features converge and a feature grid is obsolete within a quarter. Who they say they are for, what they claim versus what they can prove, where their own audience contradicts them publicly, and which questions they avoid answering.

How do you know the research is not made up?

Every claim carries a source that can be opened, and observation is separated from interpretation in the output rather than in a footnote. If a finding cannot be traced back to something specific that was read, it should be labelled as an inference, and the reader decides how much weight it carries.

When is desk research the wrong approach?

When the decision has already been made, in which case the research is reassurance rather than input. And when the category is genuinely new, in which case there may be nothing written to read, and a handful of conversations with real potential buyers will beat any amount of reading.