Focus Groups

Purpose

The focus group produces the attitudes, the perceptions, the beliefs and the reasons of a defined population, in the words that population uses itself. It is the technique to reach for when you do not yet know what to measure, and it is also the one that hands the next questionnaire its questions and its vocabulary. It steps in where numbers describe a behaviour without explaining it: a dashboard says what share of policyholders still post their invoices, it says nothing about what goes through those people's minds at the moment they reach for the envelope.

BABOK places the technique at three moments in an initiative, and each carries a different decision. On a product under development, the group's ideas are set against the requirements already stated: they correct some and bring others to light. On a finished product ready for launch, the group's reactions steer the market positioning. On a product in operation, the group reports the real satisfaction of users and gives direction to the next release. The technique therefore serves the elicitation of requirements and the evaluation of a delivered solution alike.

The deliverable is a findings report structured into themes and perspectives: what the group agrees on, what it splits over, the trends that run across several groups and the verbatim quotes that illustrate each. BABOK is explicit about the nature of that report: results are analysed and reported as themes and perspectives. The focus group makes visible what a population thinks, including where it contradicts itself, and it leaves the counting to the survey and the arbitration to the workshop.

The boundary with the workshop, the interview and the survey

That boundary is the first thing a business analyst has to hold, because all four techniques gather people around a subject and the confusion is expensive: a focus group convened to settle a trade-off produces a decision taken by eight people who had no mandate for it, and a workshop run like a discussion group ends without a deliverable. The GOV.UK Service Manual puts the distinction in these terms: research in small group workshops is not the same thing as a focus group, the latter serving to learn opinions, attitudes and likely reactions.

The coupling column is the one that earns its keep on the day you build a research plan: it says what each neighbouring technique becomes once you have chosen to start with a focus group.

The two pairings that follow are the ones a research plan should nearly always provide for. The focus group precedes the survey: it finds the questions and the words, the survey sizes them. It calls for observation: what a group says about a screen and what a user does in front of that screen are two different data, and the Nielsen Norman Group makes this its central warning, a focus group reports only what customers claim to do.

Usage

When to use it

• Attitudes and perceptions to understand: the why of a behaviour matters more than the how many.
• Behaviour measured but unexplained: the numbers establish the fact, nobody knows what produces it.
• A survey to prepare: discover the questions to ask and the respondents' vocabulary before drafting the questionnaire.
• A concept, mock-up or prototype to test: get the target segment's first reaction before the cost of building it.
• Satisfaction with a service in operation: collect lived experience and its reasons, in the users' own words.
• Cross-reactions are informative: the discussion surfaces what one participant alone would never have put into words.
• Positioning an offering before launch: test the pitch and the words on the target audience.
• Economy of elicitation: one session collects the material of six to twelve people, against as many interviews.
• Scattered audience, no travel budget: the online session remains workable, at the price of degraded reading of attitudes.
• Organisation with no data and no documented processes: the technique manufactures its own material and presupposes no history.

When not to use it

• A decision or a consensus is expected: the technique exposes divergences without resolving them, run a workshop (10.50).
• A number is expected: how many, what proportion, representative of what, those answers come from a survey (10.45).
• Sensitive, personal or confidential subject: trust does not hold in a group, prefer the interview (10.25) or an anonymous survey.

Description

From objective to plan

Everything starts from a single, precise, written objective. The guide's questions and the conduct of the session exist only to serve it, and a broadly worded objective ("understand our policyholders") produces a guide with no edge and an analysis with no conclusion. The objective is tested by its ability to sort: it must let you say of any given question that it serves the study or that it falls outside it.

The plan BABOK then describes holds seven headings. The purpose: the questions, the themes to cover, whether a discussion guide exists. The location: on site or online, in which room. The logistics: size and layout of the room, equipment, access by public transport, time of the session. The participants: the demographics sought, the observers, the moderator, the session scribe, the incentives. The budget. The schedule: when the sessions run, when the analysis is due. The results: how they will be analysed, communicated and used. That last heading is the one that gets forgotten, and its absence produces studies nobody reads.

Recruitment and screening

This is the most expensive item and the most frequent breaking point. The demographic criteria derive from the objective: if the study is about policyholders who do not use the app, recruiting among those who answer the insurer's emails means questioning precisely the population you are not looking for.

Each group is made up of similar individuals, and within-group homogeneity is a design choice: people speak freely in front of their peers. Heterogeneity is handled by multiplying the groups. The Centers for Disease Control and Prevention put the rule this way: each group brings together similar individuals, and the number of groups depends on the number of population types whose views you want. A young urban professional and a rural retiree belong in two distinct groups, and the comparison between them is part of the findings.

Screening runs through a short screening questionnaire, which checks the objective's criteria and weeds out the professional respondent, the one who strings studies together for the incentive. The usual recruitment sources are existing files, interception at the point of use, nomination by third parties, chain referral, advertising and an institute's specialist services. BABOK adds an operational instruction worth taking seriously: over-recruit, inviting more people than there are seats, to absorb no-shows. In practice, you recruit ten people to seat eight.

Convening is a four-step procedure. Set the dates, contact each person by telephone or in person, send a written and personalised invitation, then call the day before. The incentive is the norm: BABOK notes that focus group participants are often paid for their time, and money is not the only lever, a pleasant and well-connected venue, a meal, taking part in something that matters to the person recruited all weigh too. The ICC/ESOMAR code governs the rest: the incentive, the recording and the presence of observers are announced to the participant before consent is given.

Group size, count and length

BABOK sets the size at six to twelve participants, and the CDC keep the same range. Research practice tightens it further: Richard Krueger recommends five to ten, with six to eight as the optimum. Both bounds are mechanically justified. Below five, the discussion stops being a group and reverts to a conversation among several people, where the dynamic you came for never sets in. Above ten or twelve, individual speaking time falls under the threshold at which a reserved participant can still enter the discussion, and the group splits into a talking core and a silent audience.

The number of groups is the parameter that plans systematically underestimate. One group is never enough. BABOK notes that it may be necessary to run more than one when many participants are needed, the Nielsen Norman Group insists on running more than one since a single session is not representative, and the CDC tie the count to segments: one set of groups per population type. In practice you plan three to four groups per segment and stop when an additional group brings no new theme. Saturation decides the final count; the budget follows what saturation demands.

The length is one to two hours per BABOK, and the CDC tighten it to sixty to ninety minutes, giving in passing the best quality test for a guide: a session that runs past ninety minutes probably contains too many questions or too many subjects. An over-long guide is fixed by cutting.

The group plan

The number of groups is therefore not decided as a single number, it is decided on a grid. A focus group is monolingual, the group dynamic being the instrument and surviving neither a language barrier nor an interpreter, so that in Switzerland the language region is itself a segmentation dimension. A French-speaking policyholder on the maximum deductible and a German-speaking policyholder on the maximum deductible are two distinct recruitment pools, each with its native-speaking moderator, its transcription and its first analysis. The unit of recruitment is therefore the cell of the grid, and it is to the cell that the rule of three to four groups until saturation applies.

The group plan of the health insurer's study reads as follows. Wave 1 opens one group per cell in the two largest regions, measures saturation, then extends where new themes keep appearing.

Those thirty-two seated participants and forty people recruited are exactly the inputs of the study's budget, and the cost of one additional group follows from them, which makes every empty cell of the grid immediately budgetable. The deferred Ticino cell is a decision. A national study that never opens it has a blind spot it knows about.

The discussion guide and the five question types

The guide is the prepared script of questions and subjects. BABOK assigns it a structure, from the general to the particular, and recalls that it also carries the welcome, the statement of objectives, the explanation of how the session will run and of what use will be made of the feedback. The constraint BABOK places on its execution is the heart of the moderator's craft: the moderator follows a prepared script, and the discussion must feel free and loosely structured to the participants. The whole art lies in holding both.

Krueger gives the guide the precise form BABOK sketches: five question types, in sequence, funnelled from the general to the specific. The opening question goes round the table, everyone speaks once, on something factual; its role is to break the silence and equalise speech, and a person who has spoken once will speak again. The introductory question opens the subject and puts each person in touch with their own experience of it. The transition question moves the group towards the heart of the study and deepens the frame. The key questions, two to five of them, are the study itself and absorb most of the time. The ending questions are an instrument in their own right and there are three: the all-things-considered question ("of everything we have talked about, what matters most to you?"), the summary question (the moderator gives a two-minute oral summary and asks the group whether it is faithful) and the final question (the purpose of the study is restated, and the group is invited to say what has been missed).

The making of the questions obeys tight rules, and they are what separate a guide that produces data from a guide that produces noise. Questions are open. Dichotomous questions, the ones answered yes or no, are banned: they close the discussion at the moment it should open. Why is asked rarely, because it invites a rationalisation constructed after the fact; you ask instead about the attributes of the thing and the influences that played on it. Questions phrased as "think back to" return the participant to a real, dated experience, where a question in the future tense sends them off to speculate. The sequence runs from the general to the specific. And turns of phrase like "how satisfied are you" or "to what extent" import a scale into a method that has none, which produces answers unusable in both registers.

When speech alone is not enough, Krueger offers engagement strategies that make the discussion operative: choosing among presented options, listing, completing a sentence, writing on a blank card, placing oneself on a semantic differential, drawing, designing a campaign, role-playing. The most useful of all is also the best-established guard against group alignment: have each person write their answer before anyone speaks. Finally, the same guide serves every group in the study, and that is what makes the groups comparable at analysis time.

Three distinct roles

The moderator holds the session, knows the initiative, engages each participant, adapts, and BABOK describes the moderator as an impartial representative of the feedback process. Krueger makes that impartiality a discipline, and the craft is made of concrete rules. Light, discreet control. Presenting oneself as the participants do, in their register and their dress. The five-second pause after an answer, the silence that often brings out the most interesting contribution and neutral probes ("can you say more?", "can you give an example?", "I do not understand"). Controlling one's own reactions: no approving nod, no "very good", no "excellent", because a group learns within minutes which answer is expected and supplies it. Discreet control over the four difficult profiles, the self-appointed expert, the dominant talker, the shy participant and the rambler. Closing the session with the three-question ending.

The session scribe takes the notes, and this separation is structural: you cannot moderate and write at the same time. The scribe runs the recording, prepares the room, takes no part in the discussion, gives the oral summary at the end and debriefs with the moderator. The field notes carry the verbatim quotes with their author's initials, the salient points per question, the probes the moderator let pass, the scribe's own hunches and the non-verbal signals, nods, glances, passion, which will disappear entirely from the transcript.

The observers may attend the session and take no part in it. Their presence alters the dynamic, which raises a question every plan must settle: in the room or behind a one-way mirror. The ICC/ESOMAR code requires in both cases that participants be informed of their presence. The business analyst may take the moderator's role or the scribe's, never both, and in neither is a participant: no opinions are given.

When moderation is bought in, which is what the health insurer's study does, BABOK fixes the split: the moderator works with the business analyst to analyse the results and produce the findings reported to stakeholders. Concretely, the external moderator runs the session, hands over the recording, the field notes and the hot debrief and carries the context the transcript loses; the business analyst codes the transcripts, compares the groups, writes the report and signs it, because the analyst is the only one who knows the requirements the findings will touch. The moderator's cross-review of the themes is what stops the analyst projecting onto the room a reading the room never produced.

Running the session

The introduction follows a fixed shape: welcome, presentation of the subject, ground rules, first question. The ground rules are spoken out loud, and they are part of the instrument. There are no right or wrong answers, only different points of view. The session is recorded, so we speak one at a time. We use first names. Agreement is not required, listening is. Phones are off. The moderator's role is to guide, and the participants talk to each other. That last rule is what physically distinguishes a focus group from a series of interviews held in the same room.

Transcription and analysis

Analysis starts in the room. The moderator has vague comments made specific, offers a summary and asks the group to validate it. Immediately after, the session is debriefed while memory is fresh: the seating plan is drawn, the recording is checked to have actually recorded, moderator and scribe compare their themes and their hunches and the group is compared with the previous ones. Within hours, the audio goes to transcription; the analyst listens to the recording, reads the field notes and the transcript and writes a question-by-question account of that group, backed by the quotations that illustrate it. Within days, the groups are compared by category, the themes are drawn out question by question and then overall and the report is written, as narrative or as points, sequenced by question or by theme and checked with those who were in the room.

Reading a transcript obeys criteria that are in no way statistical, and they are what explains why you do not count a focus group. The words: the participants' vocabulary is itself data. The context: a remark takes its meaning from what triggered it. The internal consistency: a participant who changes position after hearing the others is a major finding, and the movement counts for more than the final position. The frequency and extensiveness: what is said often counts, what is said by different people counts more. The intensity: passion is heard in the voice and the pace, and it is nearly invisible in text, which is why you listen to the audio again. The specificity: a detailed answer lived in the first person weighs more than a generality in the third. Finally, the big ideas: you take a day's distance and name the three or four findings that matter.

What makes the exercise fail

Group alignment is the method's structural trap. The first strong opinion voiced in the room anchors the discussion, and those who follow fall in behind it out of conformity; the group then produces an apparent consensus that exists in nobody's head. The guard rests on three measures: the opening round of the table, which installs a plurality of voices before the subject is opened; the answer written individually before anyone speaks, which fixes each person's position outside the influence of the others; and a moderator who goes looking for the disagreement ("does anyone see it differently?").

The dominant participant is the embodied version of the same problem, and BABOK states it plainly: a single voice can steer the outcome of the whole group. The control is discreet and exercised through the body and through speech: breaking eye contact with the talker, turning physically towards the quiet ones, naming a person to give them the floor, returning to a written answer that a reserved participant produced before the discussion.

The moderator's bias is the most insidious because it leaves no trace in the transcript. A nod, a "very good" or a leading question is enough to teach the group the expected answer, and the final report then faithfully reproduces the sponsor's expectations. The CDC classify the focus group among the methods sensitive to facilitator bias, and that is the principal reason why the project sponsor never moderates their own study and why the guide is reviewed by someone other than its author.

The gap between the said and the done is the technique's best-documented limitation. BABOK puts it thus: what people say may diverge from how they actually behave. The Nielsen Norman Group draws the operational consequence, a focus group reports what customers claim to do and never how they use the product, and observing one user at a time remains indispensable alongside it. The practical rule is absolute: you never validate an interface with a focus group. Watching a demonstration in a group and using a screen alone at home are two unrelated activities.

The homogeneity trap is a tension to manage. BABOK warns that an over-homogeneous group produces answers that do not cover the full set of requirements; yet within-group homogeneity is precisely what loosens tongues. The way out is structural: homogeneous within the group, heterogeneous between the groups, one set of groups per cell of the group plan.

Trust sets a limit that moderation does not cross. BABOK says it: participants may have trust reservations or refuse to discuss sensitive or personal subjects. No ground rule makes a room safe for talking about your health, your salary or a conflict with your line manager in front of seven strangers. It is a criterion for choosing the technique.

The numerical reading of a qualitative result is the commonest fault in reports. Writing "seven participants out of eight said that" manufactures a statistic from a sample of eight people who were not drawn at random, and the number then travels through presentations shedding its method on the way. The CDC are blunt: this information is not measured numerically and does not hold at the individual level. You report extensiveness across the groups and intensity, in words. If a number is required, the survey is the technique that produces it.

The single group treated as the truth follows from the previous point. Eight people gathered on a Tuesday evening are a sample of eight, and the first question a clear-headed executive will ask is "do the others think the same?". The honest answer, after one group, is that nobody knows.

The online session has blind spots of its own, and BABOK names them: the moderator cannot read body language and therefore struggles to establish attitudes, and the platform mechanically limits interaction between participants, which is to say the instrument itself. The correctives: smaller groups, cameras on, a moderator who names the silences out loud and an on-site session whenever the rebound between participants is at the heart of the study.

Deferred analysis finishes off what the organisation paid dearly for. Field notes interpreted days or weeks later, once the memory of the session has faded, lose the context and the intensity, which are two of the six analysis criteria. The immediate debrief and transcription within hours are what save the data.

AI considerations

The soundest contribution bears on the discussion guide, and it is auditing more than drafting. A language model given the five question types as a grid reliably spots the precise defects the method fears: dichotomous questions, leading questions, "how satisfied" and "to what extent" phrasings that import a scale into a qualitative instrument and a sequence that skips the transition to land straight in the key questions. Having candidate questions generated is useful; having the guide reviewed against the grid is where the real gain sits, because those defects are invisible to the guide's author and obvious to a systematic reviewer.

Transcription is the item where AI changes the economics of the technique. Automatic speech recognition takes the eight hours of typing per hour of recording down to a few minutes of processing, plus the cost of the service. The trap is diarisation, the attribution of speaking turns to each speaker: that is precisely what automatic recognition gets wrong in a room of eight people talking over each other, and a quotation attributed to the wrong participant destroys the analysis, since extensiveness, the number of different people who expressed a thing, is one of the criteria that give a theme its weight. A human pass to correct the speaker labels is mandatory, it is done with the recording in your ears and it is what survives in the budget once the typing disappears.

First-pass thematic coding is an honest use: the model proposes a coding frame from the transcript and groups the comments into candidate themes, which the analyst then checks against the audio and the field notes. The gain is greatest at the cross-group comparison stage, where the volume of text is largest and the work most mechanical. The Swiss case adds a specific constraint: when the study runs in French, German and Italian, machine translation into a pivot language makes the comparison workable, but attitude lives in the phrasing and translation flattens it. The rule is therefore to code in the original language and translate only the theme labels.

Three limits are hard. First: a simulated panel is not a focus group. A model asked to play eight policyholders produces a plausible distribution of declared opinions, drawn from its training corpus. Yet the technique's best-established weakness is precisely that declared opinion diverges from real behaviour; a synthetic panel reproduces that gap with no ground truth to correct it, and it adds a defect of its own, it cannot be surprised. The value of a real group is the remark nobody expected. AI prepares the study and the analysis; it does not populate it.

Second: moderation is not automatable in principle. Reading the room, hearing the silence that follows a question, noticing the participant who went quiet when their manager's name was spoken, all of that is intensity carried by tone, pace and body, and intensity reads poorly from a transcript alone. BABOK already notes that an online moderator loses body language; an artificial moderator has none.

Third is legal. A focus group transcript is sensitive personal data: it names identifiable people and often carries information about health, income or employment. The Federal Act on Data Protection (FADP, SR 235.1) classes health data among sensitive personal data and imposes a duty to inform: the data subject is told in advance of the purpose of the processing and of what becomes of what is collected. The ICC/ESOMAR code further requires that these data be protected against unauthorised access. Pasting the transcript into a consumer AI service whose terms allow training on inputs breaks the consent the participants gave. You anonymise before any processing, or you process on infrastructure covered by the consent obtained.

One last reflex to contain: asked to summarise a transcript, a model counts. The "three participants mentioned" it produces spontaneously is exactly the numerical reading the method forbids, and none of the criteria that give a theme its weight, extensiveness, intensity, specificity, internal consistency, is a count.

Examples

A health insurer active in several cantons finds that a large share of its policyholders still posts their treatment invoices, although its app has allowed them to be sent as a photo for two years. Under the tiers garant system, the insured person pays the provider and then claims reimbursement from the insurer, and the app exists precisely for that step. Usage statistics establish the fact; they explain none of it.

Two amounts recur in the groups and they come from the Health Insurance Act (LAMal). The deductible, the franchise, is the annual share of costs the insured person pays out of pocket before the insurer starts reimbursing, and an adult chooses it between CHF 300 and CHF 2'500, the premium falling as the deductible rises. Above the deductible, the insured person still bears a co-payment, the quote-part, which is 10 % of the remaining costs, capped at CHF 700 a year for an adult. The running total of those two amounts is what a policyholder tries to hold in their head every time they send an invoice, and that is why it comes out of the groups as an information need.

The first wave of the group plan opens four groups, two in French-speaking Switzerland and two in German-speaking Switzerland, one per deductible segment, eight participants each, all policyholders who have posted at least one paper invoice in the past twelve months. Four wave 1 groups are not enough to declare saturation: they open the cells, they do not exhaust them and the study's budget gives the price of one more group. The technique produces two artefacts here, and they bracket the session: the guide before, the findings report after.

The completed discussion guide

The guide is the real instrument of this study: five blocks sequenced from the general to the specific, each with its role and its time budget, the whole fitting inside the ninety minutes the CDC set as a ceiling. It is identical for all four groups, which is the condition of comparing them.

The findings report

The report reports themes, the weight of evidence that carries them, a verbatim quote that embodies them, the divergence observed and the consequence for requirements. The weight of evidence is expressed in words: in how many groups the theme appeared, with what intensity and with what specificity. No row converts a count into a proportion, because eight people gathered one evening are not a random sample and "seven out of eight" is a manufactured statistic. A participant named and followed through a change of mind is, by contrast, data: it is Krueger's internal-consistency criterion, and the episode is reported as it was observed. A theme that appears only in the French-speaking groups is a finding.

This report gives the business owner five themes, their relative weight, the policyholders' own words for naming them, and it makes visible that the deductible question does not arise in the same way depending on the amount chosen. The rest of the research plan follows from the report itself: a survey of a representative sample sizes the five themes, and a usability test on submitting a two-page physiotherapy invoice checks what policyholders do, which the group could not say.

Visualisations

The technique's two artefacts each have their form, and the form follows from their nature. The discussion guide is a spatial object: a sequence whose scope narrows while the time widens, and the block that is narrowest in subject, the key questions, eats fifty of the ninety minutes. That tension is the lesson, and a numbered list destroys exactly it: it gives the order and loses the proportions. The guide is drawn. The findings report is the opposite: themes down the rows, criteria across the columns, text in the cells. It is a grid, therefore a table, selectable, resizable and read aloud by a screen reader, and printing it as an image would lose all of that for no gain. The group plan is made of the same wood: segments against language regions, a number in each cell.

The funnel also serves as an audit instrument: set beside a real guide, it yields four checks, and each is a question you put to the guide's author.

• Does the opening block exist? A guide that starts straight on the subject never equalises speech, and the first strong opinion voiced in the room will anchor the discussion for the ninety minutes that follow. The factual round of the table is a guard.
• Do the key questions absorb more than half the clock? If they take only a quarter of it, the guide is a twelve-subject tour of the horizon dressed up as a study, and the CDC have already given the verdict: a session that overruns ninety minutes contains too many questions. Counting the minutes per block is enough to see it.
• Does the ending block carry its three questions? A guide that finishes on a thank-you never had the summary validated by the group, and the analysis then rests on the analyst's interpretation alone, without the check the room could have applied while it was still there.
• Does every question pass the dichotomy test and the imported-scale test? A question answered yes or no closes the discussion; a question phrased as "how satisfied" or "to what extent" imports a scale into an instrument that has none and produces answers unusable in the qualitative and the quantitative registers alike.

Cost

The budget prices the first wave of the health insurer's study. The unit prices are calculation assumptions: they are set so that each line can be recomputed and the reader can substitute their own. The wave has four ninety-minute sessions, eight participants seated per session, so thirty-two participants and ten people recruited per group to seat eight, so forty people recruited. Transcription is priced here as manual, a high assumption and the reference scenario, so that the line stays comparable with the interview line; the automatic-recognition variant is priced right after.

Under automatic speech recognition, the transcription line changes nature. The typing disappears, and what remains is the cost of the service, a few francs per hour of recording, plus the human pass to correct the speaker labels, which does not automate and which runs to one or two hours per hour of recording, audio in the ears. Over this wave's six hours of recording, the line therefore falls from forty-eight hours to about ten, and the item drops below a thousand francs. That is the scenario a real study will adopt. The budget keeps the manual case as its reference because that is what makes the comparison with interviews honest, the two scenarios being unmixable in one and the same arithmetic.

The comparison with the alternative turns BABOK's economic argument into arithmetic. Interviewing the same thirty-two people in one-hour individual interviews produces thirty-two hours of session and thirty-two hours of recording, against six hours of session and six hours of recording here. At the same manual transcription rate, eight hours of typing per hour of recording at CHF 80 an hour, transcription alone rises from CHF 3'840 to CHF 20'480, that is 32 × 8 × 80. The counterpart is real: the interviews would have given individual depth and would have allowed subjects nobody raises in a group. The choice between the two is made on the nature of the data sought, and cost arbitrates only afterwards.

The cost of one more group

Four groups open four cells of the group plan, they do not saturate them. The minimum study runs to three to four groups per cell, therefore per segment and per language region, and the path from wave 1 is priced exactly, on the same assumptions.

The number checks out in one line: the wave 1 total, less the five days of analysis which are a per-wave lump sum, gives 32'880 less 6'000, that is 26'880 for four groups, and 26'880 divided by 4 is 6'720. The one line that escapes this calculation is the analysis, which lengthens with the number of groups since the cross-group comparison is what absorbs the volume and which therefore has to be reopened as soon as the grid extends. Taking both regions to three groups per cell costs four more groups, that is CHF 26'880 in direct costs, and opening the Ticino cell costs two, that is CHF 13'440. The floor of a national study is read cell by cell like this, and it is voted before wave 1.

One last cost factor is properly Swiss and it is structural. A focus group is monolingual. The Federal Statistical Office records German as the main language of about 62 % of the population, French of about 23 % and Italian of about 8 %, residents being allowed to declare several main languages since 2010, so that the shares total more than 100 % and none can be deduced from the others. An organisation active across the whole country therefore has at minimum two, and realistically three, distinct recruitment pools, each demanding its native-speaking moderator, its transcription and its first analysis. The themes are then compared across regions before being merged.

Tools

On site

A circle layout and name cards, which make participants address one another and name one another. A flip chart, indispensable as soon as the guide uses the strategy of a list drawn up together. An audio recorder, duplicated, because a lost session is a session to run again from scratch. An observation room with a one-way mirror when observers must attend without weighing on the dynamic, it being understood that their presence is announced to participants in every case.

Remote

Teams, Zoom or Webex with recording will do for a synchronous session, with the blind spots of the online session, body language lost and interaction between participants limited. Specialist online qualitative platforms open a variant videoconferencing does not allow: the asynchronous group over several days, where participants answer and answer one another on a moderated forum. It is the only reasonable way to run a study across three language regions without setting up four rooms, and it trades the immediacy of the rebound for time to think.

Recruitment

A market research institute, whose trade this is and which holds the pools. In Switzerland, the relevant quality signal is the "Market & Social Research by SWISS INSIGHTS" label, whose member institutes commit in writing to the ICC/ESOMAR code and guarantee in particular that no interview is conducted with an advertising or commercial intent. For internal groups made up of employees, recruitment goes through Human Resources, and the hierarchy problem then applies in full.

Transcription

Automatic speech recognition of the Whisper class or a service hosted in Switzerland when the consent obtained requires data localisation. Diarisation remains the weak point, and the human pass to correct the speaker labels is an integral part of the job.

Analysis

Computer-assisted qualitative analysis software, NVivo, ATLAS.ti, MAXQDA or Dedoose, codes the transcripts and compares the groups. For a four-group study, a spreadsheet honestly suffices. The manual method Krueger describes, printed transcripts, scissors, coloured pencils and one large sheet per question, remains the clearest description of what those tools automate, and an analyst who has practised it once knows what they are asking the tool to do.

Sources

• IIBA, A Guide to the Business Analysis Body of Knowledge (BABOK Guide) v3, §10.21 Focus Groups: the definition, the purpose, the elements (objective, plan, participants, guide, moderator, recording), the strengths and the limitations.
• Richard A. Krueger, Designing and Conducting Focus Group Interviews, University of Minnesota: the five question types, the making of questions, the engagement strategies, the moderator's conduct, the convening procedure and the analysis criteria.
• Centers for Disease Control and Prevention, Data Collection Methods for Program Evaluation: Focus Groups (Evaluation Briefs no. 13): the size and length of sessions, within-group homogeneity, facilitator bias and the non-measurable nature of the results.
• Nielsen Norman Group, Focus Groups in UX Research: the gap between what customers claim and what they do and the need to complement with observation of one user at a time.
• GOV.UK, Researching in small group workshops, Service Manual: the boundary between the workshop and the focus group.
• ICC/ESOMAR International Code on Market, Opinion and Social Research and Data Analytics: informed consent, disclosure of recording, incentives and observers and protection of the data collected.
• Swiss Confederation, Federal Act on Data Protection (FADP), SR 235.1: health data as sensitive personal data and the duty to inform the data subject.
• Federal Statistical Office, Languages: the distribution of main languages in Switzerland, which sets the number of recruitment pools for a national study.