How to research a multilingual audience without labelling people by language

Useful multilingual social marketing audience research starts with a decision and a recruitment rule, not a Census category.

The researcher must decide who can answer the question, how an England connection will be established, and what evidence would change the campaign or service. Participants describe their own language and format needs.

Never infer proficiency, literacy, ethnicity, consent or buying intent from administrative data.

No interviews, survey responses or audience findings are reported in this guide. It provides a method for collecting them. The method separates the commissioning buyer, internal users of research, people who may see social content and the individuals who volunteer to take part.

What to take away

Start with the decision

Write the decision in terms that can produce different actions. "Should the service commission a qualified reviewer for a Polish-language social explainer used by adults applying from Birmingham?" is researchable. "What do Polish people want?" is neither precise nor respectful.

State the source content, proposed language and locale, channel, England geography, intended action, evidence owner and decision date. Set a stop condition before recruitment. Examples include an inability to explain the study equivalently in each language, the absence of a lawful data route or too little coverage of the eligible population to answer the question.

A research question should ask about a recent task or barrier. It should not invite a participant to speak for a language, nationality or ethnic group. Nor should it ask a buyer to estimate national demand from their own customer list.

Keep five populations apart

A single project can involve several groups with different eligibility and data needs.

A participant may sit in more than one population, but the analysis should not silently combine their answers. A procurement manager's view of translation quality is not evidence of audience comprehension. An employee who speaks the language is not automatically representative of service users.

Use official statistics to design, not label

The ONS Census 2021 language bulletin reports main language for usual residents aged three years and over in England and Wales on 21 March 2021. Its detailed variable definition describes main language as a person's first or preferred language.

This may help identify locations where an inclusive recruitment plan needs closer attention. It cannot show which social platforms a person uses, whether they can read the language, what language they prefer for a particular topic or whether they wish to be researched.

Question wording also matters. The ONS language-question development report records difficulties for some bilingual respondents and the development work behind the Census questions. Quote the variable, age base, geography and reference date whenever the data informs a sampling choice. Keep it out of individual screening and targeting decisions unless a qualified reviewer identifies a lawful, ethical reason for using it.

Business statistics have a different role. The ONS UK business activity dataset can support a broad, stratified frame of organisations by English region, size and industry. It does not reveal a multilingual campaign buyer, a language requirement or purchasing authority.

Write the sampling and recruitment protocol

Define the eligible population in observable terms. For audience work, that might be adults resident in selected English local-authority areas who completed a named service task within the previous three months and who self-describe a need for one of the study languages. A business study might require people who approved a comparable commission during a fixed purchasing period.

A recruitment record needs:

The GOV.UK Service Manual page on finding user-research participants recommends clear recruitment criteria, varied routes and attention to people with access needs. It is government-service guidance, not a statistical recipe. A named social researcher should select a method that fits the decision and explain why it can support the intended inference.

Convenience recruitment through staff, an agency panel or one community organisation may be practical, but it narrows the population represented. Report that boundary. Do not attach a margin of error to a non-probability sample as though every eligible person had a known chance of selection.

Let people describe language and access needs

Screen for the experience required by the study, then ask the person which spoken or written language and format they want for the session. Do not ask a recruiter to infer those answers from a name, accent, postcode, nationality or profile.

Record locale and script only when they affect the materials or analysis. Ask whether an interpreter, support person, captioning, telephone route, large text or another adjustment would help. A person may prefer one language for conversation and another for reading technical information. Preserve that answer rather than forcing it into a general proficiency label.

GOV.UK guidance on running research with disabled people asks researchers to plan around individual needs and accessible consent. Disability and language are separate dimensions; providing translated text does not prove that the material is accessible.

Control source and target research materials

Freeze the source-language information sheet, screener, consent wording, discussion guide and questionnaire before translation. Give the language professional the research purpose, terminology, reading context and any sensitive concepts. Track each target-language version to the exact source version.

A qualified native-level reviewer should examine meaning, register and usability for the specified locale. Add a subject specialist for legal, health, financial or other technical content. Bilingual status by itself does not establish research, translation or domain competence.

Use a small cognitive or comprehension check with people eligible for the research before wider recruitment. Ask what they think the question means and how they would answer. This is a test of the instrument, not a finding about the market. Amend every affected language version, retain the change log and pause fieldwork if versions no longer ask comparable questions.

If machine assistance is proposed, document the tool, model or service, version, input rights, retention, human review and manual fallback. Do not enter personal or confidential information until security and privacy specialists accept the actual route.

Make participation information equivalent

A translated invitation is not enough. Participants need equivalent information about the research body, purpose, activities, data collected, observers, recordings, sharing, retention, withdrawal, complaints and any intended quotation or promotional use.

The GOV.UK guidance on informed consent for user research says the information must be understandable and usable, and covers withdrawal and recording. Its legal context is government research, so a commercial team still needs a named privacy and research-ethics review.

Research participation, the UK GDPR lawful basis for processing, permission to use a quotation or testimonial, and later sales contact are different decisions. The ICO's direct marketing guidance applies where the later purpose is direct marketing. Do not make an incentive conditional on a marketing opt-in.

The Government Social Research ethical-assurance guidance is written for government social and behavioural research. Its principles include informed consent, enabling participation and minimising harm. A private study can use those principles as a benchmark, while documenting its own governance.

Children and adults who may be at heightened risk require a specific safeguarding assessment, appropriate consent or assent route, trained staff and a disclosure procedure. If those controls are not ready, exclude the population explicitly and state how that affects the decision.

Run interviews without creating false evidence

Open with identity, purpose, recording and withdrawal reminders in the participant's chosen language. The interpreter or translator should have a documented role, confidentiality terms and a route to flag ambiguity. Researchers should distinguish the participant's words from an interpreter's clarification and an analyst's interpretation.

Ask about a recent event: what the person was trying to do, what they saw, what they understood and what happened next. Avoid asking whether an invented campaign idea is "good" or whether a whole community would like it. Never create polished quotations from notes. Any published extract needs the permission and contextual review specified in the study materials.

The Service Manual's advice on taking notes and recordings treats notes and recordings as potentially personal data and recommends minimising identifying detail. Access, storage, transcription, translation, retention and deletion should be recorded for every copy.

Analyse within the sample boundary

Build a codebook in the source language and document how target-language material enters it. Retain the original passage, translated analytical text, code, reviewer and uncertainty note. Where a word has no stable equivalent, keep the ambiguity rather than forcing agreement.

For every survey result, publish the eligible population, invitations, usable responses, question wording and language, numerator, denominator, period, response mode, missing answers and exclusions. Compare language versions only after checking that the questions and routes remained equivalent. Small or purposive samples can reveal problems and hypotheses; they do not estimate prevalence across England.

Counter-evidence belongs beside each finding. A barrier seen through one mobile route may disappear with assisted support, while an interview observation may not appear in service data. State what evidence would overturn the interpretation and when the finding expires.

Use personas as temporary evidence summaries

A research persona should be a hypothesis card, not a fictional biography. Name it by decision-relevant situation, such as "recent applicant needing spoken-language support", and list supporting records, contrary cases, unknowns, access needs, applicable tasks and an expiry date. Do not add a stock photograph, invented quote, assumed ethnicity or imaginary spending power.

Keep buyer hypotheses separate. A budget holder, platform operator and native-language reviewer have different jobs and risks even when one person holds several roles. Validate each through records and interviews with eligible people.

Select tools after the protocol

A survey platform cannot repair a weak sampling frame or a mistranslated question. Test candidate services against the same synthetic questionnaire and language versions. Check language routing, right-to-left display, accessible keyboard and screen-reader use, exports that preserve script, administrator access, data location, sub-processors, retention, deletion, contracts and exit.

Supplier pages describe features, not independent performance. Ask for the exact product, edition and contract. A claims page, security certificate or accessibility statement cannot alone prove suitability for the study. Keep a manual or no-new-tool route so procurement failure does not remove the research control.

Turn evidence into a bounded decision

The final evidence pack should identify the decision, populations, frame, recruitment, materials, language professionals, field dates, response flow, analysis method, uncertainties and counter-evidence. It should also show the privacy, ethics, equality, accessibility, safeguarding, security, translation, AI, IP, advertising, employment, contract and finance reviews as separate decisions.

Finish with one of three outcomes: proceed to a controlled content test, redesign the research, or stop. A finding should name the action it supports and the date it must be reconsidered. If the study cannot provide equivalent information, recruit a defensible sample or protect participants, the correct outcome is stop, not a more confident persona.

Before you act

Common questions

What should a research question ask about?

It should ask about a recent task or barrier. It should not invite a participant to speak for a language, nationality or ethnic group. Nor should it ask a buyer to estimate national demand from their own customer list.

Can Census main language data show which platforms people use?

No. The ONS bulletin reports main language for usual residents aged three and over in England and Wales on 21 March 2021. It cannot show platform use, reading ability, topic preference or willingness to be researched.

Does speaking a language make someone a suitable reviewer?

No. A qualified native-level reviewer should examine meaning, register and usability for the specified locale, with a subject specialist for technical content. Bilingual status by itself does not establish research, translation or domain competence.