Sell audience research built from product comments

A product team deciding what to build next will usually pay an agency five figures to run twelve interviews. Twelve people, recruited by an incentive, talking to a stranger about a product they were prompted to think about. Meanwhile forty thousand people have already written down, unprompted and at length, what annoys them about that product category, under the review videos on YouTube.

The comments are worth about a hundred and twenty dollars to collect. The report built from them is worth thousands. This playbook is the gap between those two numbers, and the honest account of why the gap is smaller than it looks.

What you are actually selling

A decision, backed by counts. The client is not buying comments and would drown in them. They are buying a document that says: of the 41,000 people who commented on the top eighty videos in this category, 3,100 mentioned battery life, 2,400 of those were negative, and here are eleven of them in their own words.

The two things that make it worth money are the denominator and the quotes. The denominator turns an anecdote into a proportion, which is what survives being repeated to a sceptical executive. The quotes are what makes a room believe it, because a real customer sentence has a texture that a summary never does.

What you are not selling is sentiment analysis. Every social listening tool on the market sells a positivity score, and no product team has ever changed a roadmap because a number moved from 62 to 58. Specificity is the product.

The numbers

The collection cost is close to nothing. The reading is the work, and the pricing has to reflect that or the business is a treadmill.

Twenty hours is the honest figure for report one in a category you do not know. In a category you have already covered it drops to about eight, because the taxonomy is already built and you are only re-running it against newer comments. That is the whole economics of this business: the second report in a category costs a third of the first and sells for the same, so pick two or three categories and stay in them.

Forty thousand comments is more than you need for most questions. Ten thousand from the top twenty videos gets you within a point or two of the same proportions for $30, and is the right size for a proof-of-concept you give away to win the paid one.

How it works

1. Pick the category, not the product. “Robot vacuums” produces a report three manufacturers will buy. “The Roomba j9” produces one that one company will buy, and only once.
2. Build the video list by hand. Search the category, sort by view count, take the top eighty review and comparison videos from the last two years. Automating the selection is where these reports go wrong — an automated list pulls in unboxings and children’s content, and the comments underneath them are not from buyers.
3. Collect the comments, newest-first, capped per video. A cap of five hundred per video keeps one viral video from supplying a third of your corpus and skewing every proportion in the report.
4. Throw away the noise before you start. Timestamps, “first”, single emoji, self-promotion. This is typically a fifth of the volume and all of it is worthless.
5. Build a taxonomy from a sample, then classify the whole. Read three hundred comments yourself and write down the recurring complaints. That hand-built list is what the model classifies against. Letting a model invent its own categories produces twelve overlapping themes and no counts you can defend.
6. Count, then quote. Every finding needs a number, a proportion, and three verbatim comments. A finding without a quote gets argued with; one with three does not.
7. Write ten pages and one page. The ten-page version is what they paid for. The one-page version is what actually gets circulated, and it should be the first page rather than an appendix.

Deliver the classified data alongside the report. It costs you nothing, it lets the client check any number themselves, and being checkable is most of why you get hired again.

Where it gets hard

Comments are not a sample of your buyers. They are a sample of people motivated enough to type. That skews negative, skews towards the technically confident, and skews towards whoever the video’s audience already was. Every finding in the report carries this caveat, and stating it plainly is what separates you from the social listening dashboards. A client who finds the caveat themselves, later, stops trusting the rest.

Selling research is slow. Nobody has a budget line for a report they did not know existed. The reliable entry is a free one-page finding sent to a named person — a genuine, specific, slightly uncomfortable finding about their product — with the full report offered underneath it. Expect a low reply rate and a decent conversion rate on the replies.

The category runs out. After three reports in robot vacuums you have said the interesting things. The comments refresh, but the structural findings do not. Plan on a subscription that re-runs the same taxonomy quarterly at a lower price, rather than pretending report four is as novel as report one.

The work is reading. Not tooling, not automation. If you dislike reading four hundred consumer complaints closely, this business will feel like punishment regardless of the margin.

Limitations

Comment data answers “what do people say about this” and nothing else. It cannot tell you what people bought, what they returned, what they would pay, or what the people who never commented think. Presenting it as a substitute for sales data or a survey is the failure mode that ends the client relationship.

Coverage is uneven. Some channels disable comments entirely, and those are often the manufacturer channels you most want. Older videos accumulate comments over years, so a raw count across videos of different ages compares nothing; normalise by view count or by window, and say which in the report.

Comments are public but they are written by identifiable people. Reproduce the text of a comment in a report if it is on point, but do not reproduce the commenter’s name, do not aggregate one person’s comments across videos into a profile, and do not treat a public comment as consent to be contacted. The report is about the pattern, not the people.

Sarcasm, in-jokes and non-English comments all classify badly. Expect a real error rate on any automated pass and check the classification of anything you are about to build a headline finding on.

FAQ

How many comments do I need for the numbers to mean anything?

Ten thousand from twenty well-chosen videos is enough for proportions that hold up. Going to forty thousand tightens the estimate but rarely changes a conclusion. Video selection affects the result far more than volume does.

Can I sell the same report to competing companies?

Yes, and it is the main reason this business works — a category report has three or four natural buyers and the input cost is paid once. Say up front that the report is not exclusive. If a client wants exclusivity, price it at three times the standard fee, because that is what you are giving up.

What if the client already has a social listening tool?

They almost certainly do, and it is your best argument rather than your obstacle. Ask what decision it has changed. The tools produce scores and volume graphs; this produces eleven quotes about battery life and a count. Show one finding from their own category and the difference makes itself.

Do I need to know the category already?

No, and there is a case for not knowing it. A newcomer reading four hundred comments notices what the insiders stopped seeing. What you do need is enough domain vocabulary to build a taxonomy that a specialist will not laugh at, which is a day of reading, not a career.

Is this legal?

Collecting publicly visible comments and reporting aggregate patterns is ordinary research practice. The lines that matter are reproducing individuals’ identities, building profiles of named people, and using the data for contact rather than analysis — see the Limitations above. If you are working with EU consumers’ data at scale, take advice on your specific use rather than relying on a page like this.