Market Research and Customer Insights Questions

Direct answer

When public data on a category is thin, build the estimate from several kinds of secondary sources (research someone else already collected) and treat each as good at one job and weak at another. Before trusting any of them, check three things: how old the data is, how the source defined "the category", and who paid for or produced it.

Terms used in the table. An industry code is a standard number that classifies businesses by activity (NAICS in North America, NACE in the EU). A new category rarely has its own code: a company selling AI-based clinic scheduling might be filed under "software publishers" alongside games, so a count by code mixes in firms outside your category. A segment disclosure is a company's breakdown of revenue by business line. Audited means an independent accountant has checked the figures. A fiscal year is a company's 12-month accounting year, which can end in March for one firm and December for another, so "2024" can cover different months. Treating a number as a floor means using it as a minimum, because the true value is probably higher.

Source types, what each is good for, and the check to run

Worked example of the lag check

You find a government table counting small clinics in a region, released in March 2026, describing calendar year 2024. The data period ends December 2024, and today is October 2026, so the figure is about 22 months old. If the category you study is changing quickly, treat the number as a floor or a starting point and look for a newer proxy (for example current job postings from the same kind of clinic) before building on it.

Trade-offs and pitfalls

• Two sources that look independent may both rely on one underlying dataset, so count them as one.
• Paid does not mean accurate: ask how the figure was derived, and discount any number you cannot trace.
• A source that fits your category loosely (a broad industry code) can inflate the estimate; write down what you excluded and why.
• Record source, date and definition next to every number so a colleague can challenge it.

Direct answer

Analyst reports usually describe the companies that are still standing and still talking to analysts. I would first rebuild the full list of companies that entered the market, compare it with the list the report covers, and then restate the conclusion as a statement about all entrants, not only the winners.

Two biases are in play and they are different. Survivorship bias means the failures have dropped out of the sample, so what remains looks better than the market really was. Reporting bias means the companies that are present report their good news more readily than their bad news, so even the visible data is tilted.

How to check for each

1. Rebuild the denominator (the full population the report should have drawn from). Pick a time window, then assemble every entrant from sources that do not depend on the report: funding databases, company registries, app-store history, archived versions of websites, trade-show exhibitor lists from past years.
2. Find out what happened to the missing names. Closed, acquired in distress, pivoted, or still alive but small. Post-mortems, shutdown notices and former employees are the evidence.
3. Read the inclusion rules. Who chose the companies, and who paid? Rankings often include only vendors that met a revenue floor or took part in a briefing, which filters out small and failed players by design.
4. Speak to people on the losing side (customers who churned, ex-staff of failed firms, via expert calls) so the story is not told only by winners.
5. Look at the metric, not only the sample. Growth quoted as an average of survivors, or "customers" counted as logos signed rather than still active, flatters the picture even with a clean list.

Worked example (illustrative numbers)

A report profiles 12 meal-kit companies, all growing. A funding database and a registry show 40 companies entered the market in the same window. So the report covers 12 of 40, which is 30%, and 28 are missing. Checking the 28: 21 have shut down and 7 still operate but are small (21 + 7 = 28). Known exits are therefore 21 of 40, or 52.5%.

The report's headline, "the typical player grows 25% a year", describes only the 30% of entrants it chose to profile, a group selected because it is still visible and doing well. A defensible restatement: "Among companies that survived and were profiled, growth was strong, but about half of all entrants exited, so a new entrant should plan around a survival rate below 50% (19 of 40 entrants, or 47.5%, are still operating, and 7 of those 19 are small), with the growth figure as an upside case."

Adjusting the conclusion

• Change the claim from "what winners do" to "what separates survivors from exits", comparing the two groups on the same traits (channel, pricing, funding).
• Widen ranges and lower confidence on any number that comes only from the report.
• Keep the report as one input and say plainly what it cannot show.

Trade-offs and pitfalls

• The reconstructed list is itself incomplete (databases miss bootstrapped companies), so say which direction its error runs.
• Do not swing to the opposite error and dismiss the report. It is still good evidence about survivors.
• A company that vanishes from a list may have been acquired successfully, so classify each exit before calling it a failure.

Direct answer

I would assemble a lean toolkit by research task rather than by product name, and accept the trade-offs openly: free sources first for breadth and freshness, paid sources only where free ones cannot answer the question. For a new B2B (business-to-business) product that means five groups, one per row of the table: market background, demand signals, competitor monitoring, expert sourcing and a place to record everything.

Toolkit by task

Which task each tool serves, with the trade-off I accept

• Free market sources: I accept lag and loose fit in exchange for cost.
• Competitor monitoring: I accept seeing only what rivals publish, because it is cheap and ongoing.
• Expert calls: I accept higher cost per insight and a small, possibly biased sample (one person's view) because they carry tacit knowledge (know-how people have from experience but rarely write down) no document does.
• Paid databases: only after the free route fails a specific question, and tested first against questions whose answers I already know.

Worked example

Researching a compliance-software product for mid-size banks: week one uses filings and trade-association material to find named vendors and regulation; week two sets alerts on three competitors' changelogs and reads review-site complaints; then 5 expert calls (two former compliance officers, two consultants, one competitor's ex-employee) check what the documents cannot show. I would stay within each expert's confidentiality obligations and never ask for another firm's confidential data.

Trade-offs and pitfalls

• Buying many tools gives an illusion of coverage while nobody reads the output.
• A competitor-monitoring setup that nobody reviews weekly is noise.

Direct answer

I would not add interview findings on top of the model as extra adjustments. For each finding I would first ask which model input it affects and whether that input's number already reflects it. Findings that survive that test either change the model's structure (a new segment, driver or constraint) or move a specific input within a stated range. Every step goes into a log another person can audit.

Terms. Qualitative evidence is what people told you (interviews, quotes). A quantitative market model is a spreadsheet that turns inputs, such as number of target firms and adoption rate, into a demand estimate. Double-counting means the same effect is applied twice: once inside a measured number and again as a manual adjustment.

Method

1. List the model's drivers and where each number came from (survey, benchmark, past pilot).
2. Map every finding to one driver, or to none. Findings that touch no driver may still point to a missing one.
3. Run the double-count test. Ask: "If this finding were false, which input would be different, and was that input measured from data that already contains the effect?" If yes, do not adjust again.
4. Use interviews for what they are good at: direction, mechanism, new segments, and ranges. Interviews are not a sample of the population, so they should not set a precise percentage.
5. Show the model with and without the adjustment so reviewers see what the qualitative layer changed.

Worked example (illustrative numbers)

Model: 10,000 target firms x 8% adoption = 800 adopters. In 14 interviews, 9 say the lack of integration with their accounting system (ERP) blocks adoption.

• Where did the 8% come from? If from a benchmark of products that all had ERP integration, the blocker is not in the number for a product launching without it. Say 6,000 of the 10,000 firms use an ERP. Judgement: adoption there falls to 3% (range 2% to 5%); the other 4,000 stay at 8%.
• Base: 4,000 x 8% + 6,000 x 3% = 320 + 180 = 500 adopters (range 440 to 620, from 2% and 5%), against 800 before.
• If instead the 8% came from our own pilot, which had no integration, the blocker is already in the number. Adjusting would count it twice, so the finding only informs the roadmap.

Making it reviewable

An integration log, one row per finding: finding, evidence (number of interviews, quote IDs), driver affected, direction and size (low, base, high), source of the size, double-count check result, who reviewed it. Pair it with a versioned model, a toggle that switches the qualitative adjustments off, and a rule that any reviewer can trace a changed cell back to a quote.

Pitfalls

• 9 of 14 is strong evidence of a mechanism, but it does not mean 64% of the market is blocked.
• The 3% is a judgement. Label it as such and show the range instead of hiding it in a cell.

Direct answer

Triangulation means estimating the same quantity from sources that are independent and fail in different ways, then using the disagreement to learn something. I would put each estimate on a common definition, check whether the sources are truly independent, explain the gaps, weight by source quality and publish a range with every assumption written down.

Method

1. Align definitions: same geography, same year, spend versus revenue, actual versus forecast, software only versus software plus services.
2. Check independence: if two sources cite the same underlying data, they count as one.
3. Normalise: convert each estimate to the common definition (put them on the same basis) using stated, traceable factors, as the table below does.
4. Explain residual gaps (differences that remain after normalising) instead of averaging them away.
5. Weight and range: give more trust to sources with transparent methods, then report low and high with the assumption that moves the result most. Weighting means giving the sources different shares of influence on the central estimate.

Worked example (all figures illustrative)

Three raw estimates of annual global spend on a software category:

The raw figures span 2x ($0.9B to $1.8B). After normalising, they span $1.1B to $1.5B, about 1.4x. I would report "roughly $1.1B to $1.5B" and name the weakest assumption: the 60% coverage share in row B. If it were 50%, B would be $1.8B; if 70%, about $1.29B. That tells you the one fact most worth verifying next. The remaining gap is explained, not averaged away: B sits above A and C because it is the only figure scaled up by an assumed share (60 percent), so its error is the least checked. If listed vendors really hold about 70 percent, B drops to about $1.29B and the three sources nearly agree.

To pick a central figure, apply illustrative weights: A 30 percent (opaque method), B 30 percent (sound inputs, shaky 60 percent scale-up), C 40 percent (transparent method). The weighted central estimate is 0.3 x 1.17 + 0.3 x 1.5 + 0.4 x 1.1 = 0.351 + 0.45 + 0.44 = $1.24B. The range stays $1.1B to $1.5B; the weights only place the best guess inside it. In practice the 60 and 35 percent shares come from vendor market-share tables in analyst summaries, or from comparing how many vendors in a funding database are listed on an exchange. I would also note that A and C might both lean on one government dataset, which would make them less independent than they look.

Documentation

Keep a small log with source, date, definition, adjustment factor, and who produced it, so someone can change one assumption and see the range move.

Trade-offs and pitfalls

• Averaging raw numbers with different definitions gives a precise-looking figure that means nothing.
• A single point estimate hides uncertainty; a range with named assumptions invites useful challenge.
• Do not tune assumptions until sources agree. Agreement reached that way is not triangulation.