Brand Positioning Maps from Text: Draw the Market Map Without Asking Anyone

In brief

A "perceptual map" shows where brands or products sit relative to each other in customers' minds. For decades the only way to draw one was a survey: ask people to rate similarities and attributes, then run a scaling method that turns ratings into a picture. That is slow and expensive, so most maps are stale by the time anyone acts on them. Since 2022, embeddings have offered a different route: turn text found online (reviews, forum posts, support logs, social media) into vectors, positioned so similar meanings sit close together, and read the geometry as a map. You get no fielding, no asking, and you can redraw the map whenever new text arrives. The evidence is real but narrower than the pitch would suggest. The strongest validity number, better than 75% agreement with survey data, comes from asking a language model directly, not from embedding a pile of reviews. The cleanest study that embeds reviews into a map reports no agreement figure at all. So the method can credibly recover the rough shape of a market and track how it shifts, while the specific claim that an embedded-reviews map matches a survey map has not yet been tested head to head.

What the theory says

The theory

Every brand wants to know the same thing: when a customer pictures the market, where does our brand sit, and who sits next to it. A perceptual map is that picture. Two brands close together are seen as similar, two far apart as different. The axes are whatever distinctions the market actually uses, like cheap to premium or practical to indulgent.

For decades the only way to draw one was to ask. A survey asks people to rate how similar pairs of brands are, or to score each brand on a list of attributes, and then a scaling method such as multidimensional scaling or correspondence analysis turns those ratings into positions on a page. This works, and established vendors still sell it. It is also slow and expensive, because someone has to design the study, field it, and wait. By the time the map is drawn, the market may have moved.

An earlier change to that constraint came well before the current LLM wave. Netzer et al. (2012) showed how to skip the survey and mine text people had already written, reading how brands and terms co-occur to build a market-structure map. That is the pre-embedding baseline: the idea that found text could stand in for a survey is not new.

What is new since 2022 is the instrument. An embedding turns a piece of text into a long list of numbers, a vector, arranged so that texts with similar meaning land close together in that number space. Feed in reviews, posts or support tickets, and every one becomes a point. The distances between the points form a geometry, and that geometry can be read as a map.

Transformer-based embedding models made this cheap and general enough to be a default tool rather than a fringe research project. The general-purpose vectorizers that most pipelines call, such as OpenAI's embedding models (OpenAI, no date), are documented and a few API calls away.

Embedding is not the only post-2022 way to turn found text into structure. Extracting the named things and the labelled links between them builds a queryable knowledge graph instead, a web of typed claims rather than a map of positions.

Two peer-reviewed results show what this instrument can actually do. Matthé, Ringel and Skiera (2023) built competitive-position maps for more than 1,000 listed firms and tracked how they moved over 20 years, using text similarity and a method designed specifically to stop the map jumping around from one period to the next. That is the track-it-over-time half of the claim, demonstrated at scale.

Wang et al. (2022) did the piece closest to the consumer version: they embedded product-review text, turned the pairwise distances into a map, and recovered a brand-level perceptual map from reviews alone. Together the two studies show the mechanism works.

Controversies

Neither study is a test of whether the resulting map agrees with a survey map.

The field has one strong validity number, and it is attached to a different method than the one being marketed. Li et al. (2024) tested whether a large language model could produce perceptual data that matches what a survey would produce, and found better-than-75% agreement on brand-similarity and product-attribute measures, faster and cheaper, for some categories. That is the best quantified evidence in the literature that a machine-made perceptual map can stand in for a surveyed one. But the perceptual data came from querying the language model directly, the same ask the machine what people think move covered in stated preference from language models, not from embedding a corpus of found review text into a geometry. Right destination, different vehicle.

Three different things all get called a perceptual map from text, and they are not the same job:

• Ask the model. You prompt a language model for the ratings a survey would collect. This is what carries the better-than-75% agreement figure (Li et al., 2024), but nothing was embedded. You queried a model, and whether it read your customers or its own training data is the open question that querying a model versus embedding a corpus is about.
• Embed producer text. You embed what firms write about themselves, such as regulatory filings, and map the result. Demonstrated over 20 years by Matthé, Ringel and Skiera (2023), but the text is the producer's, not the customer's.
• Embed consumer text. You embed reviews and posts and map those. Demonstrated by Wang et al. (2022), and this is the method the pitch actually rests on, yet it is the one with no agreement number attached.

The query side is an active research front in its own right. A 2025 preprint reports that language models can reproduce human purchase-intent distributions by eliciting ratings through semantic similarity (Maier et al., 2025). That is promising, and it is still a single preprint on the ask the model side of the line, not evidence about embedding found text.

Reading Li's 75% as proof that an embedded-reviews map matches a survey is the mistake the whole subject invites, and it is worth refusing out loud. The number is real. It just belongs to a neighbour.

A second, deeper worry sits under all three. Does an embedding of found text measure what people perceive, or only what they happened to write? Review and social corpora are self-selected and volume-skewed. The people who post are not the market, and the brands that get talked about are not the ones that matter equally. A map built from that text is a faithful picture of the conversation, which is not the same as a picture of perception. No published study directly contests the method on this point, so the concern is unchallenged in print rather than settled. That absence is worth reporting either way.

Then there are the vendors. A row of platforms sell market maps and conversation maps from social text. Those visuals are real product, but most are clustering and co-occurrence with a layer of generative summary on top, not an instrument checked against classic scaling. No vendor page publishes a head-to-head agreement number against a survey map. The word map is doing marketing work, and the evidence is a company describing its own tool.

Limitations

Even where the method works, four boundaries hold.

Fineness. Reporting their own limitation, Li et al. (2024) found the machine weakest on distinctions that turn on demographic or context, and expected improvement only with larger training data. So the honest claim is structure, not nuance. An embedding map can recover the gross shape of a market, which brands cluster and what the main axes are, more reliably than it can tell you how a specific segment reads a specific attribute.

Corpus. This is the one to worry about most. Because found text is self-selected, the map inherits whoever chose to write. A category dominated by a vocal enthusiast minority will map their view, not the quiet majority's. No citable post-2022 figure quantifies this skew, so it has to be argued rather than measured, but it follows directly from where the text comes from.

Time. Matthé, Ringel and Skiera (2023) built their method specifically to stop period-to-period maps drifting on noise, which is itself evidence that the instability is real. If raw text maps were stable, no correction would be needed. Track a market from text naively and some of the movement you see is the method twitching, not the market moving.

Instrument version. An embedding map is only as fixed as the model that drew it. Re-embed the same reviews with a newer model version and the geometry shifts, so a map and its follow-up a year later can differ partly because the vectorizer changed underneath and not because customers did. The embedding APIs document their model versions (OpenAI, no date), but no published source quantifies how much a version change moves a marketing map, and non-English behaviour is similarly unmeasured. Treat both as flagged boundaries, not settled numbers.

Open questions

Does a map built from found consumer text agree with a survey or scaling map, across several categories, at a correlation a reader can check? No published study answers that. The pieces exist separately: Li's agreement number from querying a model, Matthé's tracking on producer text, and Wang's map from reviews with no agreement figure. The direct head-to-head on the consumer-review method is missing. Every strong claim about that method rests on inference from its neighbours rather than on the test itself.

Will a text-only-versus-survey map result replicate across categories? No one has replicated one, so even the adjacent findings stand largely alone. And how far does embedding-model version drift move a real marketing map? No quantified figure exists, which leaves a working analyst unable to separate method noise from genuine market change.

Who decides whether an embedding-derived map is valid? The revised ICC/ESOMAR International Code (2025) governs how research is conducted and how data is handled, with a fresh pass at AI and synthetic data, but it says nothing about whether an embedding-derived map is valid. That question is decided in journals, thinly so far, not by any standards body. A vendor cannot point to a certification because none exists.

So what

An embedding map is a cheap, fast and repeatable way to see the rough structure of a market and watch it move. It is not a proven substitute for a survey on fine questions, so keep a properly sampled survey for decisions that carry money. The strongest validity number belongs to asking a language model, not to embedding reviews (Li et al., 2024). The text behind any map is a conversation, not a population, and the corpus you choose is itself a choice. Choosing it to get the map you wanted is the quiet way this method goes wrong.

For research practice

Run the embedding map first to narrow the questions, then spend the survey budget on the questions that carry a decision. Treat the map's clusters and main axes as hypotheses to check, not results to report. The specific failure to avoid: a clean-looking map arrives, agrees with what you expected, and the survey quietly gets cancelled. A map that only confirms your prior, built from text you never audited, is the cheapest way to be confidently wrong.

For companies

When a platform shows you a market map built from social data, ask what it is made of. Most commercial maps are clustering and co-occurrence with a generative summary on top, useful for spotting themes, but no vendor page publishes a check against a survey map. The map is real; the equivalence to a surveyed map is asserted. A tool that embeds reviews cannot borrow the querying-model number as its warrant.

For political parties

A map of public opinion built from social text is a map of who posts. Pick the platform, the hashtags and the time window, and you can make a coalition look united or split almost at will. That is a rigged sample moved to the corpus. Fix the corpus before you look at the map, not after. Then test anything that will drive a decision on real voters, especially the quieter ones the text leaves out.

For government and policy

The people least represented in found text are often the ones a public body most needs to hear. An embedding map of social or forum text systematically underweights whoever does not post, so presenting it as public opinion overstates what it can claim. In any document that informs a decision, record the corpus and its known gaps and state the evidence tier plainly. No standards body rules on validity here (ICC/ESOMAR, 2025), so the accountability sits with the analyst, and it should be visible in the record.

How to use this

Before you rely on an embedding map, ask four questions.

• What text is it built from, and who is missing from that text?
• Which method drew it: did someone embed a real corpus, or query a model and call the output a map?
• Is the claim about shape and movement, which the method supports, or about a fine segment-level distinction, which it does not?
• If the map drives a real decision, has a properly sampled study checked the part that matters, given that no head-to-head validation of the review-embedding method exists yet?

A map that survives these is a cheap advantage. The method has not made the survey optional, and the moment a good-looking map convinces you that it has is the moment to slow down.

Case studies

EvoMap (Goethe University Frankfurt and UNC Kenan-Flagler). The cleanest public instance of building a positioning map from text and tracking it over time is an open-source tool, EvoMap, tied to the Matthé, Ringel and Skiera (2023) work (EvoMap, no date). It maps and animates the competitive positions of more than 1,000 listed firms across 20 years, with worked trajectories for companies like Apple, Walmart and Capital One as they repositioned.

Two honest caveats travel with it. The text it reads is firm filings, what companies write about themselves, not consumer reviews, so it demonstrates the tracking mechanism rather than the found-consumer-text version of the claim. It is the anchor authors' own tool, a research instrument rather than an independent commercial deployment.

As a runnable, inspectable example of the method it is the best there is. As evidence that a review-based map agrees with a survey, it is not that.

Black Swan Data, now part of Mintel. On the commercial side, Black Swan Data uses AI and social data for trend and culture mapping, with named clients including PepsiCo, Mars and General Mills (Black Swan Data, no date). It is a real deployment of embedding-adjacent text analytics at scale.

The caveats are the tier. The client list is self-reported, and no public case study validates one of its maps against a survey or scaling map. Treat it as evidence of what is offered and bought, not of whether the map is right.

References

Black Swan Data (no date) Black Swan Data. Mintel. Available at: https://www.mintel.com/black-swan-data/ (Accessed: 18 August 2026).

EvoMap (no date) EvoMap: dynamic mapping of evolving market structures. Available at: https://evomap.io (source code: https://github.com/mpmatthe/evomap) (Accessed: 18 August 2026).

ICC/ESOMAR (2025) ICC/ESOMAR International Code on Market, Opinion and Social Research and Data Analytics. Available at: https://esomar.org/ (Accessed: 18 August 2026).

Li, P., Castelo, N., Katona, Z. and Sárváry, M. (2024) 'Frontiers: Determining the validity of large language models for automated perceptual analysis', Marketing Science, 43(2), pp. 254–266. Available at: https://doi.org/10.1287/mksc.2023.0454 (Accessed: 18 August 2026).

Maier, B.F., Aslak, U., Fiaschi, L., Rismal, N., Fletcher, K., Luhmann, C.C., Dow, R., Pappas, K. and Wiecki, T.V. (2025) LLMs reproduce human purchase intent via semantic similarity elicitation of Likert ratings. arXiv:2510.08338. Available at: https://doi.org/10.48550/arxiv.2510.08338 (Accessed: 18 August 2026).

Matthé, M., Ringel, D.M. and Skiera, B. (2023) 'Mapping market structure evolution', Marketing Science, 42(3), pp. 589–613. Available at: https://doi.org/10.1287/mksc.2022.1385 (Accessed: 18 August 2026).

Netzer, O., Feldman, R., Goldenberg, J. and Fresko, M. (2012) 'Mine your own business: market-structure surveillance through text mining', Marketing Science, 31(3), pp. 521–543. Available at: https://doi.org/10.1287/mksc.1120.0713 (Accessed: 18 August 2026).

OpenAI (no date) Embeddings. OpenAI platform documentation. Available at: https://platform.openai.com/docs/guides/embeddings (Accessed: 18 August 2026).

Wang, X., He, J., Curry, D.J. and Ryoo, J.H. (2022) 'Attribute embedding: learning hierarchical representations of product attributes from consumer reviews', Journal of Marketing, 86(6), pp. 155–175. Available at: https://doi.org/10.1177/00222429211047822 (Accessed: 18 August 2026).