Your team has interviewed customers, run a survey, read the reviews and studied the conversations happening around your category. The findings were useful. Then, a few weeks later, someone brings a new line to the meeting: “30% less sugar. Same crunch kids love.” Would that work? For whom? What might make someone hesitate?
You probably know more than you can recall in that meeting. The answer is scattered across projects, transcripts, charts and decks. Running another study for every variation is not realistic, but asking a general-purpose AI assistant to guess what customers think is not a substitute for research either.
That is the space Customer Models is built for. It is a new feature in Meaningful that turns research already in your workspace into an editable model of customer groups. You can test a message, product concept, ad idea or packaging direction in minutes; see a directional reaction from each group; and inspect the research behind the reasoning. If the decision needs a measured answer, you know what to take back to real customers.
What is a customer model?
A customer model is a set of research-grounded customer segments that you can ask questions of. In Meaningful, you choose existing research sources, review the segments the system drafts, and then explore how those groups might respond to a new idea. Each reaction is labelled directional. It is an interpretation of prior evidence, not a new interview, an observed response or a statistically representative prediction.
That distinction sounds small until you are deciding whether to spend money on a launch. “This message is likely to resonate because the group has repeatedly raised this concern” is a useful hypothesis. “63% of your customers will buy it” would be a very different claim. We do not turn the first into the second.
Why build this now?
There are hundreds of AI persona, digital twin and synthetic respondent products competing to answer the same question: How will my customers react? Their appeal is obvious. Brand and product teams need to make decisions between studies, and agencies need a way to make their accumulated knowledge useful between projects.
The problem is not that a fast, modelled answer is worthless. It is that an answer becomes easy to overtrust when its source, reasoning and uncertainty are hard to inspect. A polished paragraph, a precise-looking percentage and an avatar can make a simulation feel more measured than it is.
We think the more useful question is: Can we make existing customer evidence available at the moment a decision is being made, without pretending that the model has just surveyed anyone? Customer Models is our answer. It helps teams explore quickly, then identify the claims that deserve fresh research.
From a workspace full of research to a question you can ask
The experience follows the way a researcher would actually work:
- Choose the sources. Select research already available in your Meaningful workspace. Depending on what you have there, that might include interviews, surveys, social conversation research, reviews, uploaded studies or other customer-feedback sources. You decide what belongs in this model; the system does not silently pull in every file in the workspace.
- Build and review the segments. Meaningful analyzes the selected sources and drafts four customer groups. You can read their descriptions and supporting evidence, edit a group, add one or merge groups before making the model ready to use. A user-edited group remains a hypothesis to check, not a newly observed population.
- Try an idea. Choose a message or claim, product concept, ad or packaging direction. Describe it in your own words. You can also attach a supported image when the visual matters. An ad brief can be explored as text; an illustrative storyboard is not the same thing as testing a finished video with viewers.
- See the chart and the reasons. A single test produces a quick directional view of where each group leans, alongside what may resonate, what may hold the group back, why, and which pieces of the selected research informed that read. You can open the research behind a segment and ask follow-up questions or compare a new version.
- Keep the history. Results are tied to the model version used when you asked. When you refresh a model with new research, older answers do not quietly rewrite themselves. You can share a pinned model through a public link or bring its segments and directional learnings into a pitch deck, with the labels and source references intact.
Here is a concrete example. Suppose a snack brand tries the line “30% less sugar. Same crunch kids love.” A lunchbox-focused group may care about the sugar cue but still wonder whether children will notice a change in taste. Protein-focused shoppers may ask what else is in the bar. A value-led group may want the price per bar before it can judge the offer. A treat-led group may worry that “less sugar” sounds like less pleasure.
Those are possible interpretations of existing evidence, not claims that those customers saw this exact line. The useful output is the set of questions it gives the team: Which concern is already grounded in research? Which one is only a plausible inference? What should we actually put in front of people next?
Under the hood: a source-to-answer chain you can inspect
AI-generated personas often get discussed as if there were only two possible implementations: a big prompt stuffed with files, or retrieval-augmented generation (RAG) that picks a few similar passages. Real research demands a more careful account of what happened to the material between upload and answer.
In Customer Models, the chain is roughly selected source → source-specific analysis → segment evidence → saved model version → per-segment reaction. Each stage has a different job and a different failure mode.
1. Analyze the selected research before the first question
The build process works through the sources the user selected. It asks what needs, motivations, tensions, language and barriers each source actually reveals. It keeps the type of source in view: a participant interview, a survey result, a review, a social post, desk research and an AI-perception report do not all have the same evidential weight. An observation about public discussion should not become “customers told us” simply because it sits next to an interview in a model.
The per-source analysis is saved for the model build. The model also retains quality warnings when a source needs closer review; the system can continue drafting a model while treating the flagged material cautiously. This is especially important when the inputs are heterogeneous or incomplete. A source that says little about one proposed group should not be promoted into strong support for it.
2. Draft segments from bounded, source-linked evidence
The first build produces four draft groups from those source analyses. A group has a description, the motivations and tensions that distinguish it, and references back to selected sources. The evidence shown in the model is drawn from source-linked analysis excerpts. A link to an analyzed excerpt is useful provenance, but it is not a guarantee that every interpretation has been manually verified against every line of the raw transcript or every survey record. For consequential claims, open the original source and check.
This is also why the review step is part of the product rather than a formality. A segment name can be misleading. Two proposed groups may really describe the same behaviour. A group may have too little support to deserve confident language. Your team can refine the model before using it, and can keep those human decisions visible instead of treating the generated draft as a discovered fact.
3. Answer against a saved version, not an invisible moving target
When you submit an idea, the system uses the saved model version and its segment evidence. It generates a response for each relevant group and associates the reasoning with evidence available to that group. A response can lean positive, mixed or hesitant; where the research is thin, the honest result may be mixed with a limited-confidence explanation.
This is a deliberate trade-off. The query is fast because it works from an analyzed, bounded model, not because it rereads every raw page and every respondent record for every prompt. Source-by-source analysis before segment creation is more useful for synthesis than treating each new question as a bare top-k search over isolated passages. But it does not make the system exhaustive. Summaries can miss minority or contradictory views; uploaded material can be partial; a model can only speak from the sources selected. If a conclusion matters, inspect the underlying research and collect fresh evidence where needed.
That is the right way to ask about the retrieval strategy. “Do you use RAG?” is less informative than: Which sources were included? What was analyzed before retrieval or generation? What was condensed or omitted? Could a small but important theme survive that process? Can I get back to the material that supports this sentence?
Qualitative insight is not a vote count
Customer research is often unstructured for a good reason. An interview can explain why a claim feels implausible. A review can reveal the words people use when a product disappoints them. A social conversation can surface a tension the original questionnaire never asked about. Reducing all of that to a single similarity score loses the point.
It is also easy to overcount. Ten excerpts do not necessarily mean ten different people agreed. One especially vocal participant, repeated social posts or a heavily quoted report can create the appearance of a dominant theme. And the respondent who contradicts everyone else may be the one who reveals where your positioning breaks.
Customer Models helps you inspect the source-linked reasoning behind a reaction, but it does not present an excerpt tally as a measured frequency. Look for the unit of analysis when you make a claim: Are you talking about participants, interviews, posts, reviews, reports or quotations? If you need a prevalence estimate, use data that can support one.
Where quantitative research fits, and where it does not
We added a chart because a quick visual comparison across groups is genuinely useful. But a bar-shaped or coloured result can be misread as a survey result, so the labels matter.
The directional chart shows how each modelled group leans on the new question. For a straightforward idea it may show a yes/unsure/no leaning. For a question with explicitly named alternatives, it can show which option a group appears to favour. These categories are not percentages of customers. Four group leanings do not establish market size, purchase intent, conversion lift or a representative share of the population.
If selected prior survey research contains a relevant, explicitly cited percentage and its sample base, Meaningful can surface that as an earlier measured finding next to the model response. It stays separate from the answer to the new idea. For example, an existing survey finding about how often respondents check price per bar may help explain why a group asks about value. It does not tell you what percentage would buy the packaging concept you just entered.
The model is not running a new weighted survey, a respondent-level cross-tab or a significance test when you ask it a question. It should not invent a denominator to make an inference look quantitative. If the decision needs a new percentage, a confidence interval or a defensible difference between groups, collect and analyze responses to that specific stimulus with an appropriate sample and statistical design.
| What you see | What it means | What it does not mean |
|---|---|---|
| Directional group leaning | An interpretation of selected research applied to a new idea | A measured response rate or a forecast of customer share |
| Source-linked qualitative reasoning | The evidence and inference behind a group’s reaction | A count of unique respondents who agree |
| Earlier survey figure with a sample base | An observed result from a previous question or study | A result from this new concept, ad or pack test |
The distinction is not an academic footnote. It is what lets a team move quickly without walking into the next meeting with an invented statistic.
A model should get more useful as research changes
A brand is not static. New interviews can expose a concern the previous model missed. A survey can clarify whether a theme is widespread or merely memorable. A new pack test can overturn an appealing hypothesis.
You can refresh a Customer Model from updated research and review the new version. Tests remain associated with the version that produced them. This makes it possible to ask an important question after the fact: What did we think before the new evidence arrived, and what changed? It is much more informative than letting an AI answer quietly drift as source material changes.
There is an evaluation practice here that we encourage, even though a directional model is not itself a validation study. Write down the stimulus and model version; note the expected reactions and the uncertain parts; then put the same stimulus in front of actual customers. Compare where the model helped, where it missed a minority view, and whether a prior survey figure was mistakenly taken as a prediction. That comparison can tell you where to collect better research next.
Six questions to ask any customer-model vendor
If you are comparing AI personas, synthetic respondents or customer models, a fluent demo is not enough. These are the questions we would ask a vendor, and questions we welcome about Meaningful:
- What exactly went into this answer? Ask which sources were selected, processed, skipped, condensed or cut off. “We use your research” is not a coverage report.
- What is the unit of analysis for qualitative claims? Is a recurring theme supported by distinct participants, interviews, posts or merely several excerpts from the same person? How are contradictory and minority cases preserved?
- How does retrieval work for a large, messy corpus? Is the system synthesizing each selected source first, or only fetching a small number of semantically similar passages when you ask a question? What are the cost and context limits, and where might an important but differently worded theme disappear?
- Where did a number come from? Was it calculated from respondent-level data, copied from a prior survey table or generated as a model estimate? If calculated, can you inspect the denominator, filters, weighting, treatment of missing answers and statistical method?
- What does a citation actually verify? Does it link to a raw response, an original report or an AI-generated analysis of that report? Does the quoted passage support this specific claim, including its strength and scope?
- How do you test whether the model was useful? Can you freeze its answer before new evidence arrives, expose real customers to the same stimulus, inspect where the directional read missed and update the model without rewriting the historical record?
These are not trick questions. They are how you tell the difference between a useful research interface and a plausible story about customers.
Useful for brands and for the agencies that know their customers
For a brand team, this is a way to explore early ideas without waiting for the next research cycle. It can help narrow five message routes to two worth testing, identify the objection an ad brief fails to answer, or make a packaging discussion more specific than “I like the green one.”
For an insight agency, it is a way to keep a client’s existing work accessible between projects. The model does not replace the agency’s interpretation or the fieldwork it does best. It gives the team and client a shared place to interrogate that work, find evidence gaps and decide which questions deserve a new study.
For both, the value is in a clear handoff between exploration and measurement. A directional read helps formulate a better research question. A real test tells you what respondents did or said about the actual stimulus. Neither should be marketed as the other.
Sharing without losing the caveats
You can share a ready model through a public link. Someone who does not have a Meaningful account can open that pinned version and run a new query. They do not need a demo call or a sign-up to understand the experience. The link can be revoked, and its availability depends on the underlying sources still being usable.
There is a practical responsibility here: a public model can expose the segment descriptions and displayed research context to anyone who has the link. Review the model and the source material before sending it outside your team. Sharing is useful precisely because another person can ask a real question, but it is not a way to make sensitive research private by default.
For a presentation, the pitch-deck export carries forward the segments, source references and directional labels. Review generated slides before using them with a client. If you attached a creative image to a test, check the exported material separately; the editable deck does not automatically embed that image.
What Customer Models is not
It is not a panel of synthetic respondents pretending to be interviewed people. It is not a digital copy of a specific individual. It is not a survey emulator that returns a statistically valid forecast for a new concept. And it is not a reason to stop commissioning primary research.
It is a practical way to turn the research you have already paid for into a better conversation about the next decision. Sometimes the answer is “this might land, and here is why.” Sometimes it is “we do not know enough about this group.” Both are more useful than unjustified certainty.
Frequently asked questions
How is a customer model different from an AI persona or synthetic respondent?
An AI persona is often a generated character that can converse. A synthetic respondent is usually presented as someone who can answer a survey or interview question in place of a participant. Customer Models is built around research-grounded groups, source-linked reasoning and explicitly directional answers. The groups can help you explore a hypothesis, but their reactions are not new responses from real customers.
Can I use existing qualitative and quantitative research together?
Yes, if those sources are available in your workspace and supported in the source-selection flow. The system can use qualitative material to explain motivations and objections, while relevant figures from previous survey research can appear separately with their original sample base. A quote is not a percentage, and an old percentage is not an answer to your new question.
Does it analyze my entire research corpus every time I ask a question?
No. You select the sources for a model. They are analyzed during the build, and queries use the saved version and its segment evidence. This makes follow-up exploration practical while preserving a versioned basis for the response. It also means you should not assume that every raw passage was reconsidered for every new prompt, or that a bounded summary captured every exceptional case.
Can I test an image, an ad or a video?
You can describe an ad concept or attach a supported image when testing visual creative, including a packaging direction. A written 15-second ad brief or illustrative storyboard can help explore the idea. That is not the same as measuring how viewers react to a finished video, and we would not describe it that way.
Do the chart and percentages predict how many people will buy?
No. The chart shows directional leanings of modelled groups, not a population forecast. Any earlier survey percentages shown are explicitly separated and labelled with their source and sample base. A purchase or preference estimate for a new idea requires research with real respondents exposed to that idea.
What if the sources disagree or the model has a gap?
That is worth seeing, not hiding. Read the supporting excerpts, inspect the original source where a claim matters, and treat weakly supported or user-edited groups as hypotheses. If a gap would change the decision, it is a good candidate for a new interview, survey or concept test.
Try it in your workspace
Open Models in your Meaningful workspace, select the research you want to build from and review the proposed customer groups. Then try the message, concept, ad or pack you are discussing right now. You will get a directional chart and the reasoning together, with the research close enough to challenge the answer.
If you want to explore before using your own data, the Models tab includes a sample UK snack-shopper model. Its research and reactions are fictional and illustrative, including the example ideas in this article. You can also watch the illustrative product demo.
The point is not to replace the next study. It is to make the work you have already done useful while you decide what that study should be.