What question does social listening not answer for fashion brands?
Social listening tools answer a question you can already name. You can track your brand name, a competitor, or a keyword like "quiet luxury". They return a measurement of volume and sentiment for that term. This is useful for campaign tracking and brand health monitoring. But it cannot find what you do not know to ask for. It is structurally unable to discover demand for a product that has no name yet, or a style that is spreading through images, not words.
The most valuable consumer signal for a fashion brand is not what people say, but what they wear. The limitation of keyword-based tools is that most of what people wear is never named. A person posts a photo from a weekend trip; the jacket they have on is the signal, but the caption is about the scenery. The question left unanswered is: what are our customers, and the customers we want, actually wearing today, and what does that mean we should make next?
Why is tracking what people actually wear so difficult?
The core difficulty is that the data is not created for researchers. It is just life. In our own analysis of fashion audiences, we found that only 2% of relevant images were explicit "outfit posts". The other 98% is the signal that every other method misses. It is a person in the background of a photo, a video of a concert, a candid shot at a market. The garments are there, but they are not presented, tagged, or described. To measure them is to solve a difficult data science problem at scale.
This is not a problem you can solve by hiring more analysts. The volume of visual data is too vast. It requires a purpose-built technology stack that can ingest millions of images and videos, identify the presence of a person, isolate the garments they are wearing, and do so without relying on any text. This is the hard part, and it is the reason most tools default to measuring keywords instead.
How does Hugo build a map of what a fashion audience wears?
We build standing infrastructure that holds a live model of a consumer audience. This is not a report or a one-time study. It is a permanent, always-on map that connects real people to the specific garments they wear. The map is the foundation, and it is what allows us to see demand before it shows up in sales data or search trends. We build this infrastructure from the ground up for each brand, focused on the specific audience they need to understand.
As we detailed in our research on mapping what consumers actually wear, this process has four distinct layers, built in order: reaching the right people, detecting their garments, clustering those garments into distinct items, and resolving those items to products. Each layer is a separate technical challenge, and the quality of the final map depends on solving each one correctly.
How do you find the right people for an audience map?
You cannot find a representative audience by typing "25-35 year old sneaker enthusiasts in London" into a search bar. The people who self-describe that way are a fraction of the total audience, and likely the least representative. To build a true map, we have to reach real people based on their connections, not their stated interests. We use a method from survey science called respondent-driven sampling, or chain referral.
We start with a small, credible set of seed accounts belonging to the audience. Then, our system traverses the social graph, following real connections to discover who they interact with, who they follow, and who follows them. This allows us to build out a map of an entire community, reaching the people at the periphery who are often the most influential but the least visible to traditional marketing.
What is chain referral and why is it better than a search?
Chain referral is a statistical sampling method developed to study populations that are hard to reach because no central list of them exists. It works by using the network structure of the population itself to guide the sampling process. You start with a few members and ask them to refer others, creating chains of connection. This approach is powerful because it finds people based on who they are connected to, not what they say about themselves.
For a fashion audience, this means we find people who are part of a style scene because they are genuinely part of it, not because they use the right hashtags. A keyword search finds the loudest voices. Chain referral finds the community. It is a slower, more computationally intensive process, but it produces a map of the audience that reflects its real structure and influence pathways.
How many people and posts does it take to map a city like New York?
Building a high-fidelity map requires significant scale. For our public map of the New York fashion audience, the process started with a handful of seed accounts and expanded from there. The final map was built from observing the public activity of 876 distinct accounts. This was not a sample of 876 posts, but the entire public output of those accounts over a period of time.
From those accounts, our infrastructure processed 4,464 individual posts, comprising both images and video. This depth is critical. A single post is a data point; thousands of posts from a connected community reveal the patterns of adoption, influence, and taste that define what the audience will want next. It is this scale that allows us to move from anecdotal observation to statistical certainty.
How do you identify garments from ordinary photos and videos?
This is the second layer of the infrastructure. Once we have the posts from the audience, our computer vision models analyze every frame to identify garments. The system first detects people, then segments the images to isolate what they are wearing: a jacket, a pair of trousers, shoes, a bag. Each isolated garment becomes a "crop".
From the 4,464 posts in the New York map, the system generated 9,146 distinct garment crops. These are not perfect, studio-lit product shots. They are crops from real life: partially obscured, in motion, under varied lighting. The models are trained specifically to handle this kind of noisy, authentic data, which is where the real signal lies.
How can you tell if two garments are the exact same item?
This is the third and most crucial layer: clustering. Having thousands of garment crops is not enough. We need to know which crops show the same physical item. To solve this, we convert each garment crop into a vector, a mathematical representation of its visual features. Then, we use vector search to find crops that are close to each other in that mathematical space.
This process groups the 9,146 garment crops into clusters of visually similar items. A human expert then reviews these clusters to confirm which ones represent the exact same product. In the New York map, this process identified 169 distinct garments that appeared on multiple, unconnected accounts. This is how we find a trend: not when someone talks about it, but when the same item starts showing up across a community organically.
How do you connect a garment to a specific product or SKU?
The final layer of the map is resolution. This step bridges the gap between an observed item in the wild and a commercial product. For each confirmed cluster of a unique garment, the system attempts to match it to a known product in the market. This can involve searching retailer catalogues, brand websites, and marketplaces using the visual data from the cluster.
When a match is found, the node on our map is enriched with the product's name, brand, and, if available, its SKU. This transforms the map from a picture of what people wear into a demand surface. A brand can see not just that "denim jackets are popular," but that one specific denim jacket from a competitor is spreading through their target audience right now.
What does a finished audience map actually look like?
The finished map is not a dashboard or a chart. It is a graph. In this graph, nodes are either people (accounts) or garments (products). The edges are the relationships between them: this person wore this garment, this person influences that person, these two garments are worn by the same people. It is a live model of the social physics of taste for that specific audience.
For example, our New York map shows a specific pair of white sneakers appearing on 7 different accounts. The map shows that the people wearing those sneakers also tend to wear a particular crew tee and a specific style of jeans. Their connections, in turn, are seen wearing a woven leather bag that appears on 4 other accounts. The map reveals these constellations of products that define a real-world style.
Can you show a real example of the garment map in action?
Our public research on the New York map provides a concrete example. The map covers three major cities: Paris, London, and New York. In New York alone, we identified 832 distinct garments being worn by the 876 people in the audience. The most powerful findings come from the connections. We found 169 items that were the exact same product, worn by different people who often did not follow each other.
This is the signal. It is objective, verifiable proof of demand. You can click on any garment in our internal maps and see the evidence: the original posts, with handles and dates, where the item appeared. This auditability is fundamental. It is the difference between a black-box "insight" and a piece of verifiable intelligence you can build a decision on.
Why is observing real people better than simulating an audience?
Many AI tools are moving towards "synthetic respondents" or "AI personas" to simulate consumer audiences. This approach involves prompting a large language model (LLM) to act like a certain type of consumer. The fundamental problem is that a language model is a reasoning engine. It constructs rational explanations for behavior. Real consumer choice is not rational. It is driven by habit, mood, status, and what someone saw their friend wear.
As we explain in our research on audience simulations, an LLM asked to simulate a shopper will give you the most plausible, conventional answer for that persona. It is right when the answer is obvious and wrong exactly when the insight would have been valuable. A 2026 audit of 37 major models confirmed this, finding that a simple statistical model fitted to real human data was more accurate than any of them at predicting real human responses.
What does "one brain" mean for AI-generated consumer panels?
A simulated panel of ten thousand AI agents is not a population of ten thousand minds. It is one brain wearing ten thousand name tags. All the agents are built on a single foundation model, with the same weights and the same underlying patterns of thought. Their errors are not independent; they are correlated. When one agent is wrong in a certain way, they are all likely to be wrong in that way. Averaging their answers does not cancel out the bias, it launders it into a confident, but wrong, consensus.
Recent research measures this effect, calling it "persona collapse". Distinct profiles prompted into an LLM converge into a narrow band of stereotypical responses. A 2026 study found that prompting an LLM to adopt a persona often made its answers *less* like the real people it was supposed to represent. You cannot simulate what you have not first measured.
How does Hugo's simulation method differ from using language models?
We do not generate respondents. We observe real people and then simulate the dynamics of their real network. Our simulation work stands on the infrastructure of the audience map. We start with the graph of 40,212 real people in a measured audience and the trends spreading through it. We then fit mathematical models of influence and adoption to that real data.
The result is a model that predicts what will spread, who will adopt it, and why. We tested this against a frontier LLM given the same data. Our method, built on the mathematics of graph propagation, was significantly better at predicting which real people would adopt a trend next, achieving an AUC score of 0.86 versus the LLM's 0.78. We simulate the physics of the system we have already measured, we do not invent people from whole cloth.
What decisions can a fashion brand make with this infrastructure?
The purpose of the map is to enable better and faster commercial decisions. Because the infrastructure is live and the data is at SKU level, the agents we build on top of it can answer specific, high-value questions. They work continuously, around the clock, and report their findings directly into a brand's Slack.
This includes identifying catalogue opportunities by finding products a brand already sells that are gaining traction with a new audience. It includes recommending new products to make based on observed gaps in the market. It finds product feedback by analyzing what people say in the context of wearing an item. And it can match a brand with the single creator who has the most authentic influence over a specific audience for a specific product, based on the measured network graph.
What are the current limitations of this mapping approach?
This method has clear boundaries, and it is important to state them. First, it can only see what happens in public. It cannot measure private conversations or what happens inside a walled garden. This means it is not a replacement for every other form of research.
Second, the signal skews towards the delighted and the annoyed. People tend to post about experiences at the extremes, so while we can accurately measure the presence and spread of a product, we cannot use this data alone to calculate market share or the exact incidence of a complaint in the total customer base. The frequency is directional, not a census. Finally, it cannot test a pure counterfactual for a product that has no existing analogue in the world.
How does this infrastructure run continuously for a brand?
A traditional research project is a snapshot. An agency is commissioned, they conduct a study over six weeks, and they deliver a report. The findings are dated the moment they are printed. Our infrastructure is not a project; it is a permanent installation. The audience map is always on, continuously updated as new posts and interactions occur.
This means the intelligence is live. An agent built on this infrastructure is not answering a question about last quarter. It is working with data from today. This continuous operation is what allows it to spot a trend as it emerges, not after it has peaked. It is the difference between reading a history of your market and having a live feed from it.
What is the difference between a map and an agent?
The map is the hard part we own. It is the proprietary infrastructure that holds the live model of the consumer audience. It contains the people, the garments, the products, and the network of influence between them. The map itself is the foundational layer of intelligence. It is what we build everything else on.
The agents are the product. An agent is the infrastructure pointed at a single commercial decision. For example, a "New Product" agent constantly scans the map for unmet needs and emergent styles, and recommends what the brand should make next. A "Creator Matching" agent uses the graph to find the partner with the highest measured influence for a specific launch. The map provides the intelligence; the agent applies it to a job.
How does a brand work with a Hugo agent?
The agent does not live in a separate dashboard that someone has to remember to log into. It reports its findings and recommendations directly into the brand's own Slack. When the agent identifies an opportunity, it opens a thread in the relevant channel, for example #product-planning, and presents the finding.
The finding includes the recommendation, the reasoning, and the evidence, with links back to the original source posts. The team can then discuss the decision in that same thread. This closes the loop between intelligence and action. The goal is not to deliver a report for someone else to interpret, but to deliver a decision into the exact place where the decision is being made.
What is the first step to mapping my brand's audience?
Every brand's audience is unique. The process begins with defining the audience you need to understand. This could be your existing customers, a competitor's customers, or a demographic you want to win. We work with you to identify a set of seed accounts that represent the core of that community.
From there, our systems begin the process of building the map: expanding the audience through chain referral, ingesting the visual data, and running the layers of detection, clustering, and resolution. The first results, showing the initial structure of your audience's taste, can often be available within weeks. The first step is a conversation about who you need to know.