Which Marketing Research Method Should You Use? A Decision Guide

Choose a marketing research method by identifying the evidence your decision still lacks. Observe behavior to find where something happens, interview people to explore why, survey a defined sample to estimate how common a view is, run a controlled experiment to test a change, or use credible existing sources for market context.

Start with relevant evidence you already have, then collect new evidence only for the unanswered question. For example, checkout events can show where customers leave, while interviews or usability observation can investigate why. Combining methods is useful when discovery, measurement, and causal testing are separate decisions.

Define the decision before choosing a method

A broad goal such as “improve conversion” does not yet say what to research. Define the decision, the people or events it concerns, the time period, and the finding that could change what you do. Then classify the question:

• Discovery: What is confusing users, and what explanations should we investigate?
• Measurement: How common is the issue in a defined population?
• Causal comparison: Did a specific change affect a defined outcome under test conditions?

For checkout, “Why are users abandoning?” is a diagnostic question, not a request for a conversion percentage. Specify the journey and period, inspect relevant funnel events, and decide whether you need to observe the experience or ask users about it.

Qualitative data describes experiences, language, and possible explanations. Quantitative data measures numbers. A survey, observation study, or other method can produce either or both, depending on its design. First-party customer and website records are operational data. Whether analyzing them counts as primary or secondary research depends on how the data was collected and how the current study uses it. Social listening has the same qualification.

Match the research question to the method

Decision question	Suitable evidence	Useful method	Cannot establish alone
What is confusing people?	Experience, language, follow-up	Individual interviews or usability observation	How common the issue is in a population
How common is this view?	Structured answers from a defined sample	Survey with a stated sampling approach	That reported preference predicts behavior
What do people do?	Observed actions or behavioral events	Usability observation, journey analysis, or internal behavioral data	Why a person acted that way
Did this change affect the outcome?	Comparable outcomes under defined conditions	Controlled experiment, such as an A/B test	That the result generalizes beyond those conditions
What is happening in the market?	Dated, scoped information from credible sources	Competitive analysis or public and purchased research	Competitors’ confidential plans or your customers’ views

Interviews allow individual follow-up. Focus groups can surface group reactions, but participants may influence one another. Surveys can combine closed and open-ended questions, but reported answers are not observed behavior. Behavioral data shows what happened without necessarily revealing motive. Experiments estimate the effect of a specific change under their test conditions. Secondary sources provide context only to the extent that their scope, date, and methodology fit your question.

For competitor or market research, record the publisher, source URL, publication or update date, scope, and methodology before using a finding. Treat social listening as evidence of available public conversation, not a complete record of customer opinion. HubSpot’s overview of marketing research methods is a useful editorial introduction, not a peer-reviewed methodology standard.

Four practical research workflows

The workflows below are illustrative designs, not tested deployments or turnkey integrations. Field names are proposed examples, not vendor-prescribed properties.

Trigger and input	AI or rule-based job	Validation gate	Action and fallback
Checkout diagnosis. A defined period, funnel events such as session_id, observed_step, and observed_at, plus usability notes or interview excerpts.	Rules group exits by journey stage. AI may suggest themes from anonymized notes and attach evidence spans.	Check that event and interview records refer to the same journey and period. A theme requires a cited passage and researcher approval.	Link the reviewed finding to funnel analysis and assign a product or UX decision. Unsupported themes, contradictory stages, or missing context go to the research owner.
Message feedback. A defined concept, research question, consent record, and responses with respondent_id, question_id, and response_text.	AI can suggest draft themes or summaries while preserving text spans. It does not estimate prevalence unless the sampling and counting design supports it.	Validate response identity, consent and access rules, allowed fields, evidence spans, and theme definitions.	Publish a reviewed research summary or revise the message. Sensitive, contradictory, or ambiguous responses remain with the researcher rather than entering a CRM conclusion.
Issue prevalence. A defined eligible population, survey question, allowed responses, period, and duplicate-response policy.	Use deterministic rules to count valid responses. AI is unnecessary for the calculation and may only draft a checked explanation.	Check eligibility, missingness, duplicates, question wording, period, denominator, and sampling limitations.	Store a period-scoped aggregate with its denominator and calculation version. The analyst resolves denominator ambiguity before reporting a percentage.
Design test. An intervention, eligible population, assignment identifier, control and treatment variants, primary metric, and test period.	Rules check assignment and exposure. No AI is needed to assign variants or calculate the metric.	Check instrumentation, exposure, outcome definition, competing changes, and the prespecified analysis rule.	Record ship, iterate, or inconclusive. An assignment fault or invalid outcome prevents the result being treated as a reliable comparison.

1. Diagnose checkout abandonment

An exit event can locate a problem, but it cannot identify a customer’s reason by itself. First inspect where exits cluster. Then observe the relevant journey or interview affected users. Keep behavioral events and interview observations as separate records, linked to their source.

A proposed output might contain a reviewed theme, the journey stage, an evidence reference, and counterexamples considered. If event and interview evidence point to different stages, resolve the mismatch before recommending a change. Measure the selected checkout outcome after the product decision, without treating the research finding itself as proof of improvement.

2. Collect reactions to a new message

Use interviews when individual follow-up matters, or a survey when structured comparison across a defined sample matters. A response record might include response_id, respondent_id, question_id, response_text, and collected_at. Keep the response separate from any later AI processing run.

AI may organize open-text themes, but the researcher checks supporting text, contradictory responses, and whether the theme answers the original question. If using a HubSpot survey, verify the account edition, seat, permissions, delivery channel, and record-association behavior before launch. The platform does not choose your sample or establish that responses represent the intended audience.

3. Estimate how common a reported issue is

Define the eligible population, sampling approach, response-quality rules, period, and denominator before collecting responses. Decide how missing answers, ineligible people, and duplicate submissions will be treated. A percentage without a clear denominator is not ready to report.

Store the result separately from response-level records. A suitable aggregate could include aggregate_id, population_scope, period_start, period_end, valid_response_count, denominator_definition, metric_value, and calculation_version. Do not apply a universal sample-size rule. Adequacy depends on the decision, sampling approach, uncertainty, and consequences of error.

4. Test a specific design change

Prespecify the eligible population, assignment rule, control and treatment variants, primary metric, test period, and decision rule. Retain experiment_id, assignment_id, variant_id, outcome_value, and the analysis version.

Check assignment, exposure, instrumentation, competing changes, and the outcome definition before interpreting the result. A before-and-after change without a controlled comparison is not causal proof. If the planned decision rule cannot be applied, record the result as inconclusive rather than allowing a readable AI summary to turn it into a winner.

Combine methods without confusing their evidence

A practical checkout sequence is to use funnel data to locate a drop-off, observe or interview users to investigate possible reasons, survey a defined population if prevalence matters, and test one specific change if the decision is causal. Each method answers a different question.

Triangulation means comparing evidence with different strengths, not forcing every source to agree. When results conflict, check whether the population, question wording, timing, instrument, or denominator differs. Keep stated preference, observed behavior, and experimental effect distinct in the final recommendation rather than averaging them into one conclusion.

Design the evidence record before automating it

Choose the record grain before creating CRM properties or an integration. One observation row should mean one atomic event or research observation, not a mixture of a survey response, citation, AI run, and period total. Keep the project, response or observation, AI processing run, citation, and aggregate as separate records or clearly differentiated entities.

The following is a hypothetical observation record, not a HubSpot-prescribed schema:

{
  "observation_id": "obs_0042",
  "project_id": "checkout_diagnosis_01",
  "method": "usability_observation",
  "observed_at": "2026-10-14T10:15:00Z",
  "source_record_id": "session_or_note_id",
  "raw_evidence_ref": "restricted_source_reference",
  "interpretation_status": "needs_review",
  "reviewer": null
}

Use stable source IDs when available. If there is no source ID, combine the identifiers that define the actual row grain, such as project, record type, source, and response, citation, or run ID. A project-plus-date key is not universal because multiple valid responses, citations, or processing runs can occur on one date.

When concurrent processing is possible, a lookup followed by create can race. Use a database-enforced uniqueness constraint or a supported transactional upsert keyed by a configured unique property. HubSpot documents batch upsert using a unique property, but the target object property must actually be configured and unique. Check the HubSpot batch upsert documentation and current API version guidance before implementation.

Keep aggregates in separately scoped records. A survey percentage or share-of-voice total needs a period, population or competitor scope, denominator, and calculation definition. Do not store it as if it were one individual observation. Preserve source references separately from AI-generated summaries, and retain processing and review details where AI is used. Teams defining record ownership and CRM writeback can explore CRM systems consulting.

Use AI for bounded support, not research judgment

AI can suggest themes from open-text responses, extract candidate topics, or draft a source-linked summary. It should not turn an exit event into a motive, make a convenience sample representative, or decide an experiment winner. Before collecting or processing responses, define consent, privacy, access, retention, and any restrictions on sending customer data to an AI service.

Use deterministic rules for required fields, allowed method values, date formats, numeric ranges, duplicate keys, prohibited personal data, and approval status. Validate AI output against a schema before it enters a workflow or database. A hypothetical output contract could be:

{
  "topic": "checkout_payment_step",
  "summary": "Draft theme, subject to review",
  "evidence_span": "reference to supporting note or excerpt",
  "confidence": 0.82,
  "pii_detected": false,
  "needs_human_review": true
}

Reject missing evidence references, invalid fields, or unsupported values instead of silently accepting them. Route ambiguous, contradictory, sensitive, or low-confidence outputs to a named researcher. Do not let an uncorroborated public post overwrite a customer-level CRM field.

HubSpot’s reviewed documentation supports survey workflows and notifications, but does not establish a prebuilt end-to-end AI classification, confidence gate, human review, provenance, and CRM writeback process. Teams evaluating a bounded AI-supported workflow can review AI agent consulting.

Check platform limits before choosing a collection channel

HubSpot custom feedback surveys are documented for email, shareable link, web page, and workflow delivery. The current guide lists Service Hub Professional and Enterprise availability, requires an assigned Service Hub seat to create, edit, and publish surveys, and specifies additional marketing access and publishing permissions for email surveys. Responses can be associated with contacts, companies, and deals in the HubSpot survey context. External survey data does not automatically map to CRM records.

For shareable-link surveys, HubSpot documents cookie-based behavior: submissions from the same device may overwrite previous submissions. This is not a general respondent identity or research deduplication strategy. If repeat responses must be retained, capture an appropriate study-specific respondent identifier and define the duplicate policy before launch. Review the current HubSpot custom survey guide and test the intended association.

HubSpot’s documented social keyword feeds cover X under current constraints. Feeds begin after creation, the X feed has a seven-day display window, and updates may take up to four hours. Connected-account social reporting varies by network. These features should not be described as complete historical monitoring across all platforms or all public mentions. Check current permissions, data availability, privacy requirements, and whether the channel reaches the population the study needs. Teams reviewing HubSpot workflow and CRM structure can explore HubSpot systems consulting.

Make the decision and record what would change it

There is no universal threshold for sufficient research. Match the evidence requirement to the stakes, uncertainty, target population, evidence quality, and cost of being wrong. Record the finding, sources and limits, decision owner, action, and evidence that would trigger a review. Then assess whether the research answered its original question and whether the resulting action produced the intended operational outcome.

The practical rule is simple: use behavior to locate, conversation to explain, structured measurement to estimate, experiments to compare, and credible secondary research to add context. Treat each result according to what its design can support.