What Customer Research Is and How to Run One in 2026

Last updated: 18 September 2026

Quick Answer: Customer research is the structured study of who buys, why they buy or refuse, and what would change that. It runs as a project in five stages, from a written decision to an in-market read. Teams reach saturation around 12 to 15 interviews, then size the findings with a survey, in-house or commissioned.

In a 2008 post-election survey, Pew Research Center asked Americans what mattered most in their vote. Offered a list, 58 percent picked the economy. Asked with no list, only 35 percent volunteered it, and 43 percent of them named something absent from the closed list. Same electorate, same week, two findings. No customer research service, and no in-house team, analyzes its way out of that.

Customer research fails more often on who you asked and how you asked than on the analysis. A sample drawn from your own list tells you what engaged customers think, not what the market thinks, and extra responses do not repair it.

The Four Questions Customer Research Has to Answer

Every project, whatever its method, reduces uncertainty on four questions. One that moves none of them is data collection, not research.

1. Who buys, and who does not? Segmented by something that predicts behavior, not something easy to ask.
2. Why do they buy, or refuse? The job, the trigger, the objection, the thing that almost stopped them.
3. What do they think of you against the alternatives? Including doing nothing, the competitor least often listed.
4. What would change their behavior? A change to product, price, message or experience, stated so someone can act.

The fourth is where most studies quietly fail. A report that describes customers accurately and names no change anyone can make has answered three out of four, and the fourth is the one that paid for it.

What Are the Types of Customer Research?

Textbooks split the types three ways: primary versus secondary, qualitative versus quantitative, attitudinal versus behavioral. None tells you which study to run, because all three describe the data, not the decision. The cut below is by decision, alphabetical by type, implying no ranking.

These map onto the standard categories in research types and when each earns its place.

How Do You Run a Customer Research Project End to End?

Run it as five stages, each with an exit test you can fail. The common failure is leaving a stage early, not skipping one.

1. Write the decision down. The decision, the options, the date it gets made. If nobody can name a choice that would change on the result, stop there. Exit test: you can name the choice and the date.
2. Mine what you already own. Support tickets, churn reasons, site search, returns data and analytics record what people did, not what they remember. Exit test: you can say what that already answers, so you do not re-buy it.
3. Interview before you count. Depth interviews produce the answer options a survey later measures, and skipping them is how teams reproduce the 2008 problem: a closed list missing the real answer. Ask about specific past episodes. "Walk me through the last time you bought one" beats "would you buy this." AI-moderated interviews changed this stage's arithmetic, not its logic. Exit test: new interviews stop producing new themes.
4. Size what you found. A survey finds how common those things are, in the respondents' own language. Who answers matters more than how many: Pew Research Center's telephone polls fell to a 6 percent response rate in 2018. Exit test: the numbers separate the stage 1 options.
5. Test before you commit. Put the preferred option in front of behavior, not opinion: a live concept, a prototype, a landing page, a limited launch. Synthesized evidence on the intention-behavior gap reports a medium-to-large change in intention yielding only a small-to-medium change in behavior, d = 0.36. Exit test: you observed a behavior, not an intention.

How Many Customers Do You Need to Interview?

For qualitative work the answer is saturation rather than a number, and published guidance clusters tightly. A 2024 concept analysis of the saturation literature reports a common recommendation of 12 to 15 participants in relatively homogeneous populations, rising as the population varies. Three distinct segments means three of those groups, not twelve interviews split across all three. For quantitative work, sample size follows the precision you need: two concepts five points apart need a sample that can resolve five points.

Sample quality decides more than sample size, and the drift is measurable. The US Census Bureau reports the weighted response rate for the 2025 Current Population Survey ASEC at 62.0 percent against 69.0 percent in 2019, and finds the consequence rather than assuming it: since 2020 the survey-only income estimates have run about 2 to 3 percent above estimates corrected with administrative records.

A high response rate is not proof of a clean sample either. AAPOR is explicit that response rate alone does not reliably separate accurate from inaccurate data, and its Standard Definitions exist so that two vendors quoting 60 percent are quoting the same fraction of the same denominator. Where respondents came from is the more useful question to put to a provider, and sourcing models differ more than the pitch decks suggest.

When Is a Customer Research Service Worth Commissioning?

A customer research service is worth commissioning when reach, method design or the capacity to land the result is genuinely missing, and not before. Those are the three things a buyer either already has or buys.

1. Reach. Access to the right people, including the ones who are not customers and never will be on your list.
2. Method design. A study built to answer the stage 1 decision, with the guide written before fielding rather than assembled ad hoc.
3. Capacity to analyze and land it. Someone to interrogate the data and get a recommendation in front of the person making the call.

The Run-It-Yourself Test

Keep it in-house when the question is narrow, the people you need are already on your list and willing to talk, and someone owns the follow-through. A team that interviews ten of its own customers this week usually learns more than one that spends six weeks scoping a project.

The Commission Test

Buy it when you need non-customers, several markets or languages, a sample you can defend to a board, a method your team has not run, or speed you cannot staff. Language alone is a bigger constraint than most single-country plans assume: the Census Bureau counts 67.8 million people in the United States, almost one in five, speaking a language other than English at home.

What a provider supplies that an in-house team usually cannot is the unglamorous half: recruitment and screening to a quota, incentive rails that actually pay people, moderation in the respondent's own language, and an audit trail. Once it has the client's brief, Alchemic tailors the discussion guide before fielding rather than leaving that to the buyer, runs managed fieldwork or bring your own recruitment across 14 markets including the USA and the UK, runs in-home use tests where the product has to be lived with for days, and keeps the findings queryable afterward instead of retiring them with the deck.

Commissioning the same study every quarter is no longer a project, it is a program, and a standing research operation runs on different infrastructure and a different cadence.

Which Customers Never Reach Your Research?

The standard method skips this question, and it decides whether the answer is true. Browser links, email invitations and in-product prompts reach people already connected, already on your list, already engaged enough to respond. Everyone else goes missing systematically rather than randomly, which is why a larger sample does not help.

The scale is documented. The ITU reports 2.2 billion people still offline, with 85 percent of urban populations online against 58 percent of rural ones, and 94 percent internet use in high-income countries against 23 percent in low-income ones. Even in a fully connected market, your list excludes everyone who considered you and chose otherwise, usually the group the decision depends on.

The gap is domestic too: in 2021, 82 percent of US households earning $150,000 or more had a computer, smartphone, tablet and home broadband, against 29 percent of those under $25,000.

What Each Channel Covers

Channels are alphabetical, with who each reaches well and who it structurally misses.

Alchemic was built around this gap: interviews run natively inside WhatsApp with no link and no app, outbound AI phone calls for people a browser never reaches, and it publishes 57+ languages including Spanish, Arabic and Mandarin, so the interview happens in the respondent's own language. Practical routes to respondents without smartphones need no platform at all, and often suffice for a single-market study.

How Do You Turn Findings Into a Customer Research Report?

A customer research report is judged on one thing: whether a decision-maker can act without reading the appendix. Keep it to five parts, in this order.

1. The decision and the recommendation. One page. The stage 1 options, the recommended one, and the confidence behind it.
2. The evidence for each option. What was found, how common it was, and which segment it came from.
3. The verbatim layer. Quotes and clips attached to each claim, so findings can be interrogated rather than believed.
4. What the sample excludes. Who was not reached, and which conclusions weaken as a result.
5. The open questions. What the study did not settle, and what would settle it.

Part 4 is most often dropped and it is the part that protects the report. A finding that survives a reader knowing exactly who was missing is one the business can build on, though moving an insight to a decision is a separate discipline from producing it.

What Goes Wrong in Customer Research Projects?

Four failures account for most wasted studies, and each is a choice made before anyone reads a chart.

• Leading and loaded wording. Pew Research Center's own examples are stark. Support for taking military action in Iraq ran at 68 percent, and fell to 43 percent once the same question added that US forces might suffer thousands of casualties. One view, two numbers, and one of them ends up in a deck.
• Hypothetical questions treated as forecasts. What someone says they would pay, or whether they say they would buy, is not a prediction. Anchor pricing and demand to observed behavior instead.
• Reviews and social posts read as clean data. Unprompted sources are valuable and also manipulated. The Federal Trade Commission's final rule banning fake reviews and testimonials, announced on 14 August 2024, prohibits six practices, among them AI-generated fake reviews, undisclosed insider reviews and the suppression of negative ones.
• Consent and participant duties handled loosely. The ICC/ESOMAR International Code on Market, Opinion and Social Research and Data Analytics is the profession's global standard, and the Insights Association's Code of Standards states the operational duties plainly: participation must be voluntary and based on accurate information about the purpose of the research, and personal data may only be used or shared for the purpose it was collected for. Research run outside those duties is a legal exposure before it is a data problem.

Where Customer Research Falls Short

It is a poor guide to genuinely new categories, because people reason about unfamiliar things by analogy and the analogy is usually wrong. It cannot establish causation, and it inherits every limit of its sample. Open links make the last one worse: a 2024 study of a small language community that promoted its survey on social media classified 83.1 percent of 1,774 responses as suspected fraud.

Three cases where something other than customer research is the better answer:

1. Measuring what people do at scale. Product analytics and transaction data beat asking, and are already paid for.
2. Proving a change worked. A randomized in-market test or a holdout beats any interview, and costs less than a study that cannot settle the question.
3. Reading embodied or situational behavior. In-home observation and in-store shopper work beat remote modes when context is the variable, a real limit of every remote method including the ones here.

A study that admits what it did not settle beats one implying it settled everything.