Audience research for social listening starts with the decision, not the data

Good social listening audience research starts with a population and a decision, not with the demographic labels displayed in a dashboard. Public posts can show what was observable through particular sources and queries. They cannot tell a researcher who stayed silent, prove that authors live in England or replace consented interviews and surveys.

This guide sets out a research design for an English audience. It does not report interviews, responses, personas or audience totals. Every finding would need to be produced through the documented method, then reviewed for privacy, research ethics, legal risk and statistical interpretation before publication or commercial use.

What to take away

Write the decision before collecting data

A useful brief names the person who will act, the choice they face and the evidence needed by a date. "Understand our audience" is too loose. "Decide which service questions an English customer-support team should investigate next quarter" is testable.

For an illustrative worked example, take an English utility that wants to reduce repeat contacts.

Turn assumptions into research questions. The GOV.UK Service Manual advises teams to prioritise questions and test beliefs about users. That guidance concerns government services. Its principle transfers well to commercial research, but it does not validate a particular business decision.

Record a stop rule too. Do not proceed if the decision owner cannot say how the answer would change the work, if the needed population cannot be reached fairly, or if the planned data use cannot pass qualified review.

Define the population and England boundary

The population is the full group about which the team wants to draw a conclusion. It might be adults in England who used a specified service during a stated period, staff responsible for brand-risk decisions at English local units, or customers whose contracted service address is in England. These definitions are not interchangeable.

Verify England residence in practice. Use a contracted service address, a billing postcode, or a self-report checked against a business address or invoice. For business participants, cross-check the registered address. If residence cannot be verified, code it as unknown and keep it outside England-only counts.

For business research, the ONS UK business activity, size and location dataset can provide registered-enterprise or local-unit frames by English region, industry and size. It does not identify individual participants, listening users or buyers. Choose either enterprise or local unit and document the extraction date and exclusions.

For England population denominators, the ONS Census 2021 and mid-year population estimates are official sources.

For public audiences, Ofcom's Online Nation 2025 supplies UK-wide evidence about internet and platform use. It may inform where recruitment or observation gaps could arise. It does not establish a representative England sample for a new study. Ofcom's Adults' media use and attitudes report provides further UK-wide context.

Build a sampling frame that matches the claim

A sampling frame is the list or route from which people or organisations can actually be selected. Existing customers, newsletter subscribers, a panel, venue visitors and publicly posting accounts each omit different people. Write down those omissions before recruitment.

Set inclusion and exclusion fields that answer the research question. An interview study about purchasing listening services might include people who recently specified, approved, used or rejected such work. It should exclude participants whose only connection is general social-media use. A study of service conversations needs criteria tied to that service and period, not presumed sentiment.

The Service Manual's participant recruitment guidance recommends recruiting actual or likely users and recording criteria, access needs and recruitment bias. Its suggested participant numbers are for iterative government user research, not a rule for market prevalence. Choose sample size for the method and intended inference.

Indicative sample sizes, not rules:

Keep a screening log with frame source, invitation route, eligibility result, non-response and completed participation. Do not add a name, demographic attribute or employer detail merely because it is easy to obtain.

Separate three evidence streams

Audience research often mixes unlike material. Keep these streams distinct:

Triangulation means checking whether these sources agree or conflict. It does not turn three biased samples into a representative one. Preserve conflicting evidence and explain which population each stream can support.

Treat participation and lawful basis separately

An invitation, consent to take part and the lawful basis for processing personal data answer different questions. The ICO's current research provisions guidance says ethical consent to participate is distinct from UK GDPR consent as a lawful basis. It also says the appropriate basis depends on purpose, context and organisation type.

Researchers should therefore document both the participation process and the controller's legal assessment. The GOV.UK guidance on managing participant data calls for early information, minimal recruitment data, secure handling and deletion when details are no longer needed. A private organisation still needs advice fitted to its own work.

If listening content includes personal data, public availability does not erase data-protection obligations. Record the source permission, purpose, fields, access, retention, sharing, rights process and any special-category implications. Automated labelling can add new inferences, so the review must cover outputs as well as raw posts.

Keep research invitations apart from marketing

The ICO says contact for genuine market research is not direct marketing. A message may cross that boundary if it promotes a service, collects details for sales or combines research with a later offer.

Use a research-only invitation, explain who is conducting the work and avoid sales follow-up unless a separately reviewed route permits it. Store recruitment choices independently from prospecting records. A person who discusses a problem in an interview has not thereby agreed to receive marketing.

Recruit for coverage, not convenience

Recruitment should reflect the population and the decision. Start with a matrix of relevant experience, role, region, access need and any characteristic necessary to test the question. Then show how each invitation route can reach those groups.

Customer lists may overrepresent current users. Social followers omit people who do not follow or post. Professional networks can cluster around the researcher's contacts. Paid panels have their own membership and response patterns. None is automatically wrong, but the resulting claim must stay within the achieved sample.

The 2023 MRS Code of Conduct applies to MRS members and accredited company partners. It covers participant welfare, transparency and reporting when a sample is described as representative. It is a professional code, not proof that an unaccredited project complies with law or good practice.

Design interviews around real decisions

An interview guide should move from context to a recent, specific episode. Ask what decision arose, what information was available, which sources were trusted, what was missing, who was involved and what happened next. Avoid asking participants to predict whether they would buy an imagined service.

The Service Manual's in-depth interview guidance recommends open, neutral questions, real examples, a tested discussion guide and follow-up when meaning is unclear. Do not claim this makes an interview unbiased. Interviewer behaviour, recall and recruitment still shape the material.

Keep a versioned guide and note any departure from it. Record observations before interpretation. Quotes should be verbatim, consented for the intended use and checked for identifiability.

Use surveys only for questions they can answer

A survey can count responses within its achieved sample. It estimates a wider population only when the sampling design, response pattern, weighting and uncertainty support that inference. A large convenience sample can still be biased.

Write one idea per question, use neutral wording, provide sensible response options and offer "not applicable" where forced answers would mislead. Test the form with people who match the intended participants. Check routing, keyboard use, mobile display, privacy text, completion time and export before launch.

ONS guidance on survey uncertainty distinguishes sampling error from problems such as non-response, inaccurate answers and processing error. Report the base for every percentage, missing responses and any confidence interval that the design justifies. Do not attach a margin of error to a convenience sample as if it were random.

Analyse without inventing a persona

Start with evidence units: a source post, interview observation, survey response or operational record. Code them using a defined scheme and retain links to the original material under the approved access rules. Have another researcher review ambiguous or high-impact coding.

The Service Manual's analysis guidance separates observations, findings and actions. Use that distinction in the report. "Three interviewed users described delayed escalation" is an observation about a sample. "The process lacks a clear escalation owner" is an interpretation that needs supporting evidence. "Assign an owner" is a proposed action.

A persona should be an evidence summary, not a character. Use a functional label such as "weekly insight commissioner" and attach the source studies, dates, inclusion criteria, decision, recurring needs, constraints, contradictory evidence and confidence. Do not add a fictional name, portrait, biography or quotation. If two groups cannot be distinguished reliably, keep them together and record the uncertainty.

Bound tool and competitor evidence

Competitor research needs similar discipline. Public websites, filings and advertisements show what was published on a date. They do not prove internal process, product performance or customer experience. Mark supplier claims as first-party and reserve "tested" for work actually performed under a declared protocol.

Companies House warns that its public register is not a comprehensive source. Use it to verify filed identity and status, not competence or market position. If a resulting comparison is used in advertising, the CAP Executive's verifiability guidance says identifiable comparisons need objective, checkable support.

Report confidence and decision use

Every finding should carry five labels: population, achieved sample, evidence period, method and inference limit. For counts, give the denominator and missing data. For qualitative themes, show how many participants or records supported and contradicted them without treating frequency as population prevalence.

End the report with decisions, not decorative insight. The owner can proceed, narrow the question, gather different evidence or stop. Note which observation changed the choice and what evidence would reverse it.

Revisit the work when the audience definition, service, platform access, privacy guidance or commercial purpose changes. Until real research is completed and named reviewers approve the method and interpretation, this framework supports planning only.

Before you act

Common questions

Why is a stop rule needed before data collection begins?

The article says do not proceed if the decision owner cannot say how the answer would change the work, if the needed population cannot be reached fairly, or if the planned data use cannot pass qualified review. A stop rule prevents research that cannot inform a real decision.

Can a profile location prove someone lives in England?

No. Verify England residence with a contracted service address or a self-report checked against a business address. If it cannot be verified, code it as unknown.

Does public availability remove data-protection duties for listening content?

No. The article says public availability does not erase data-protection obligations if listening content includes personal data. Researchers should record source permission, purpose and retention, and review automated labelling outputs as well as raw posts.