Market research is most useful when it resolves a defined business uncertainty. Start with the decision, identify what is not known well enough to act, collect evidence suited to that gap, and record what the evidence does and does not support.
For example, if qualified visitors leave a pricing page, first locate the drop-off in funnel data. Then interview a mix of converters and non-converters about a recent decision. Use a survey or experiment only if the team needs to estimate how common a candidate explanation is or compare specified alternatives.
This approach turns market research from a collection of reports into a defensible decision process. It also limits the role of AI: a model may organize approved evidence, but it should not silently convert a weak sample into a fact, estimate, persona, or CRM update.
What market research is, and what it should decide
Market research is the collection and analysis of information about customers, markets, and competitors to support a specific decision. It can inform whether to enter a market, which segment to prioritize, how to revise an offer, how to position a product, or what to investigate next. It cannot guarantee demand, product-market fit, or commercial results.
A topic is not yet a research question. “Understand our customers” is too broad to guide recruitment or analysis. “Why are qualified visitors leaving the pricing page before requesting a demo?” identifies a population, behavior, and evidence gap that can shape a study.
Before selecting a method, the decision owner should state:
- The decision: such as whether to revise pricing-page messaging.
- The deadline and consequence: when the decision is due and what could happen if it is wrong.
- The current hypothesis: for example, “implementation effort is unclear,” labeled as a hypothesis rather than a finding.
- The evidence gap: what the team does not know well enough to act.
Research begins with a decision and an uncertainty, not a tool or a pile of data.
Choose evidence to match the uncertainty
Primary research is collected directly for the question through interviews, surveys, observation, usability sessions, or experiments. Secondary research already exists, such as public datasets, published reports, customer reviews, or internal historical records. CRM, sales, support, and product data can provide useful context, but those records may have been collected for another purpose.
Use qualitative research to explore how people describe a problem, make a choice, or interpret an offer. Use quantitative research when you need to measure behavior, estimate frequency within a defined population, or compare specified groups. Interviews can surface a hypothesis. Structured measurement can test how prevalent it is. Neither method repairs a biased sample or an unclear question.
Combine sources when the decision requires both explanation and measurement. Product analytics may show where a behavior occurs without explaining why. Interviews may reveal several plausible reasons without showing which is common. Keep those contributions separate rather than forcing them into one unsupported conclusion.
Recruit beyond enthusiastic current customers. Depending on the decision, include prospects who evaluated but did not buy, customers who selected another solution, and people who match the intended market but have not encountered the product. A small interview sample can surface candidate themes, but it does not establish representative prevalence. Continue sampling when segments differ, evidence conflicts, or the decision has serious consequences.
For market sizing, define the geography, customer segment, market boundary, unit or revenue basis, base year, forecast period, inclusion rules, exclusion rules, and method. Label modeled values as estimates. The U.S. Census Bureau is a possible source of public demographic and economic data, not a ready-made answer for every market.
Example: diagnosing a pricing-page drop-off
- Use product analytics to identify where and when qualified visitors leave. Record the event definition, population, date range, and denominator.
- Interview converters and non-converters about what they were trying to decide, which alternatives they considered, and what they did next. Do not suggest a reason in the question.
- If the team must estimate how common a candidate explanation is, define the target population and denominator before running a suitable survey or test.
If analytics show that visitors leave before reaching pricing details, a pricing-preference survey may not explain the immediate behavior. Inspect the sequence first, then choose the next method.
Design a study that can produce usable evidence
A short research brief is the handoff between the decision owner and the research lead. It should specify the decision, target population, key unknowns, method, collection dates, owner, and result that could change the decision. Set eligibility rules before recruitment and record exclusions.
Recruiting only colleagues, loyal customers, or people easiest to reach can hide objections and alternatives that matter. The HubSpot market research guide recommends open-ended questions organized around background, problem awareness, solution evaluation, decision factors, and closing questions. Use that structure as a starting point, not as a substitute for a sampling plan.
For interviews, ask about a recent real experience: what triggered the search, which alternatives were considered, what the participant did, and what influenced the choice. Request examples. Separate a participant’s account from analytics or transaction records.
For surveys, define the population and denominator before launch. Pilot the wording and response options. Do not describe a convenience sample as representative. For observation or usability work, record the task, starting condition, completion rule, and interpretation limits.
Protect interview transcripts, survey responses, support records, and CRM data with access controls. Remove unnecessary personal information before sharing material with research or AI tools. Keep the original response separate from any coded theme or summary.
Turn raw material into traceable findings
Keep source records, analysis runs, citations, and aggregates distinct. A source record is an interview, response, review, or event. An analysis run records one particular review, model execution, prompt version, and input set. A citation is evidence supporting a finding. An aggregate summarizes a defined population and period. These records have different grains and must not be substituted for one another.
For a finding, retain a project ID, finding ID, source record ID, source type, source URL or internal location, observation period, segment, finding text, supporting excerpt or measurement definition, analysis or model identifier, confidence or limitation label, reviewer, and review status.
For counts and percentages, retain the numerator, denominator, date range, and population. State whether a count refers to participants, responses, mentions, coded excerpts, accounts, or events.
One interview excerpt can support a candidate theme. A count of participants expressing that theme is a different measure. A monthly segment percentage needs its own population, numerator, denominator, and period. Retain the source citations and calculation basis so the percentage can be checked rather than used as a substitute for its evidence.
AI can summarize approved transcripts, suggest candidate theme clusters, or locate excerpts. Give it a bounded question and require source IDs and exact excerpts. A researcher must check those excerpts against the original material, verify the counting grain, and approve any consequential finding.
Use deterministic rules for fixed categories, dates, identifiers, deduplication, required fields, and arithmetic. Use AI for ambiguous language and candidate interpretations. When evidence conflicts, preserve the conflict and create a follow-up question instead of allowing a model to choose one explanation.
The following is an illustrative finding record, not a vendor-defined schema. Its declared row grain is one finding supported by one source record. A separate citation record would be needed when a finding has multiple supporting sources.
{
"row_grain": "one_finding_one_source_record",
"finding_id": "pricing-page-study-01-finding-014",
"research_project_id": "pricing-page-study-01",
"source_record_id": "interview-014",
"source_type": "interview_transcript",
"observation_period": "2026-10-10",
"segment": "evaluated-but-did-not-buy",
"theme": "implementation effort",
"supporting_excerpt": "Checked excerpt from the source record",
"count_unit": "participants",
"frequency_count": 3,
"denominator": 8,
"review_status": "pending_human_review"
}In this example, three of eight interviewed participants expressed the theme. It is not a market-wide estimate and does not mean three mentions. Keep raw transcripts in an access-controlled evidence repository and put approved findings in a decision memo. HubSpot advertises upload-based analysis and prompts in its gated Market Research Kit, but the landing page does not document the evidence fields or review contract above. Organizations designing controlled synthesis workflows can also consider AI agent services.
Compare methods by the decision they support
Set the proof standard before collecting evidence. Interviews are suited to explanations. Behavior data shows actions. Structured measurement can estimate prevalence within its defined population. Documented secondary sources provide market context.
| Decision | Evidence and AI job | Validation and action | Fallback |
|---|---|---|---|
| Why did qualified visitors leave pricing? | Funnel events plus interviews with converters and non-converters. AI suggests themes and locates cited excerpts. | Verify event definitions, population, date range, and excerpts. Test a messaging change and monitor the same funnel measure. | If the drop-off occurs before pricing content, investigate navigation or comprehension before testing price preference. |
| Which price or offer should we test? | Stated preferences alongside selected plans, completed purchases, cancellations, refunds, and discount exposure. AI organizes open-text reasons. | Report test conditions, sample, denominator, and observed behavior separately from stated willingness to pay. | Route inconsistent price signals to a controlled test or additional interviews rather than treating survey intent as demand. |
| Is an opportunity worth further investment? | Defined market boundary plus compatible public data, reports, or bottom-up inputs. AI organizes assumptions and detects definition conflicts. | Check source dates, arithmetic, currency treatment, boundaries, and assumptions. Present a labeled estimate to the decision owner. | If sources use incompatible definitions, narrow the estimate or document the method change instead of combining them. |
For competitive research, count indirect substitutes too. A spreadsheet, manual workflow, another category of tool, or doing nothing may compete with a purchase. Do not combine incompatible market definitions without explaining the change in scope or method.
Vendor case studies can illustrate possible workflows, but their reported outcomes are not transferable benchmarks without test design, baseline, control conditions, time period, and independent verification. For example, the Strella case study reports a Ritual pricing workflow and results, while the result remains vendor-reported.
Move findings into a business system without losing ownership
A CRM update is a possible downstream use, not an automatic result of analysis. First decide whether the finding belongs on a contact, company, deal, custom object, or external evidence store. A persona label used to segment outreach should not be treated as validated merely because an AI system proposed it.
Before a write, verify the object, property meaning and type, allowed values, permissions, source freshness, reviewer approval, and stable record identity. Decide whether to overwrite a value, append information, or create a related evidence record. If a short CRM field cannot hold provenance, retain an evidence reference in a separate record or repository.
- Confirm the target record with an existing CRM object ID or a verified unique property.
- Check that the property meaning, type, and allowed value match the approved finding.
- Retain the evidence reference, reviewer, source freshness, and approval status outside a short field when needed.
- Use a logged write operation and reconcile rejected or partially completed writes.
- Send ambiguous identity or disputed classification to a named human owner.
A lookup-then-create sequence is not safe when concurrent workers can process the same record. Two workers may both find no match and create duplicates. Where concurrency matters, enforce uniqueness in the integration’s staging database or use a transactional upsert, then write with a verified unique CRM property or existing object ID.
Keep separate keys for citation-level findings, analysis runs, raw observations, CRM events, and aggregates. For example, a proposed citation key could combine research_project_id, source_record_id, finding_type, and finding_version. A run-level key could combine project, run, source, and model identifiers. A daily aggregate needs its own observation date and aggregation definition. These are illustrative design patterns, not vendor-published schemas.
HubSpot documents a Persona contact property, unique-property behavior for contacts, and batch upsert documentation. The batch-upsert reference is labeled legacy, and current API documentation is date-versioned. Check the current endpoint, object requirements, permissions, and property configuration before implementation. These sources do not establish a turnkey research-to-HubSpot integration. See HubSpot’s Persona property documentation, its API overview, and the legacy batch-upsert reference. For teams defining record ownership and CRM controls, CRM systems consulting is relevant.
Close the loop with a monitored decision
Turn an approved finding into an action, named owner, baseline, success measure, and review date. Track research quality separately from business results. Source coverage, participant mix, unresolved contradictions, and sampling limits describe the evidence. Conversion, retention, cost, or adoption describe outcomes after an action.
Choose a measurement period suited to the action and record other changes that could affect the result. If customer mix, product behavior, market conditions, or the buying process changes, identify which new evidence updates or challenges the earlier conclusion. The decision owner should review the result on the agreed date and continue, revise, or reverse the action.
Frequently asked questions about market research
How many interviews do I need?
There is no universal number. A small initial sample can surface themes, but continue when segments differ, evidence remains varied, contradictions persist, or the decision is consequential. Five to ten interviews can be an initial stopping point when themes repeat, not proof of representative prevalence.
When should I use a survey instead of interviews?
Use interviews to understand language, experiences, and possible reasons. Use a survey to quantify responses across a defined population after the concepts and answer options are clear. Specify who could respond and what the denominator represents.
Can ChatGPT do market research?
An AI assistant can help draft questions, summarize supplied material, or suggest candidate themes. It cannot make a sample representative, verify unsupported facts, or replace primary evidence. Check consequential summaries against their source records.
How should I test willingness to pay?
Separate stated willingness to pay from observed choices. Report the price and offer shown, sample, test period, denominator, selected plan, completed purchase, cancellation, refund, and discount exposure.
How often should personas be refreshed?
Review them when customer, product, market, or buying-process evidence materially changes. Record what new evidence supports or challenges the current profile rather than relying on an arbitrary universal schedule.