What is the difference between a digital twin and an audience simulation?
A digital twin is a virtual model of a physical object or process. An audience simulation is a mathematical model of how a behavior spreads through a network of real people. The first models a thing you sell. The second models the system that decides whether to buy it. One is a replica of a product. The other is a replica of a market.
This distinction is not academic. A digital twin of a running shoe can tell you about its material stress and performance under load, a problem of physics. It cannot tell you who will wear it first, why they chose it over a competitor, or whether a new colour will be adopted by the community that matters. These are not questions about an object, but about a social system. Answering them requires modeling influence as a cascade problem on a graph of real people, a fundamentally different kind of engineering.
Why is this distinction important for consumer brands?
Because the hardest questions a brand faces are not about its products, but about its customers. What should we make next? How will this campaign land? Which complaint is a real signal and which is just noise? These are questions about future human behavior. A digital twin, which models an inanimate object, is the wrong instrument for the job. It contains no model of taste, status, influence, or need.
Audience simulation is built to answer exactly these questions by modeling the social physics of a market. For example, in our New York audience map, we observed how a specific pair of white sneakers appeared on 7 different accounts. The people wearing them also wore the same crew tee and jeans, and their connections propagated the trend to a woven leather bag seen on 4 other accounts. This is the mechanism: modeling not just who has what, but how influence propagates through a network of real people to predict the commercial outcome of a decision before it is made.
What is a digital twin, in simple terms?
The term "digital twin" comes from engineering, where it describes a virtual model of a physical object, like a jet engine, updated in real time with sensor data. While powerful for modeling predictable, physical systems, its application to consumer brands is often a category error. A dashboard showing sales data is not a digital twin. A model of a supply chain is closer, but it still models logistics, not the unpredictable human demand that pulls from it. A digital twin models a thing; it cannot model a market.
What is an audience simulation?
An audience simulation is a working model of a specific consumer audience, built from observed behavior and connections. It is not a model of a generic "consumer" but of the actual network of people who buy from a brand or its competitors. The model contains not just who the people are, but the structure of their relationships: who follows whom, who engages with what, and who is seen wearing the same products.
At Hugo, our audience simulations are built on a graph. The nodes are real people we have observed. The edges represent influence and co-occurrence, with weights learned from data. When we simulate a product launch or a marketing campaign, we are running a cascade problem on this graph, predicting how a new idea or product will spread from person to person. It is a model of social physics, grounded in observed reality.
Isn't this just another name for AI personas or synthetic respondents?
No. This is the single most important distinction. Most tools that claim to simulate consumers use a large language model (LLM) to generate responses. They write a prompt describing a persona, for example "a 35-year-old mother of two in Chicago who values sustainability," and ask the LLM to answer questions as that persona. An audience simulation built this way is an act of imagination, not observation.
Our position is that this approach is structurally flawed for predicting real consumer behavior. A language model is an engine for generating rational, conventional text. Consumers are often irrational and unconventional. The very mechanism of buying, which runs on mood, status, envy, and timing, is what language models are trained to suppress. We do not generate respondents. We observe real people and model the mathematics of their collective behavior.
What is the "one brain" problem with AI populations?
A population of ten thousand synthetic agents running on one foundation model is not a population. It is one brain wearing ten thousand name tags. They share the same underlying weights, the same training data, and the same post-training alignment. Their errors are not independent; they are correlated. When one agent misunderstands a concept or exhibits a bias, it is likely that many others will too.
In real research, averaging across a large sample cancels out random individual error. But with a synthetic population, averaging correlated errors does not cancel them out. It launders the model's systematic bias into a confident-looking number with a false sense of sample size. This is what we call the "one brain" problem, and it is a fundamental weakness of using a single LLM to simulate a diverse population.
Why are language models bad at simulating irrational consumers?
A language model is a reasoning engine. It is trained on vast archives of text, including books, scientific papers, and structured arguments from Wikipedia. It learns to construct coherent, logical explanations. When asked why someone bought a particular jacket, it will generate a plausible, rational account of a decision based on price, quality, and features.
But real purchasing decisions are rarely so rational. They are driven by habit, impulse, what a friend was wearing, or an aspirational feeling a brand evokes. The reason is often invented after the fact to justify an emotional choice. A language model, by its very nature, is built to do the opposite: it starts with reason. Furthermore, models are deliberately aligned to suppress prejudice, status judgment, and in-group signaling. For a social scientist, these are not bugs to be fixed; they are the very mechanisms of consumer behavior.
What does the evidence say about synthetic respondents?
The academic literature is beginning to measure the gap between synthetic and real responses, and the findings are stark. One 2026 preprint study across 70,000 real survey answers found that prompting a model with a persona often made its answers less like the real person's, not more, and pushed the error onto specific subgroups. This confirms that synthetic respondents often do little more than mimic data already collected, rather than generating genuinely new predictive insight.
Another 2026 audit of 37 different models, including those from OpenAI, Anthropic, Google, and twelve open-weight families, put a number on the "one brain" problem. It found that the models resembled each other more closely than any of them resembled the actual humans they were supposed to be simulating. In that same audit, a simple statistical model, with no language capabilities, did a better job of matching the real human answers than all 37 language models combined. The conclusion is that synthetic respondents are flattened caricatures that are wrong exactly when a finding would have been worth paying for.
How does Hugo build an audience simulation without language models?
We observe, we do not imagine. Our process starts by building a detailed map of a real audience. We then fit a mathematical model to their observed behavior to predict what they do next. The core of our simulation is not a language model, but a graph neural network that learns the coefficients of influence from how trends have actually spread in the past.
As we detailed in our research on building simulations without LLMs, this involves three steps. First, we identify candidate trends from the audience's conversation. Second, we model the network of influence between people. Third, we fit a model that predicts adoption based on who is exposed to the trend and how influential they are. The social physics sits in the mathematics of the model, where it can be measured and validated, instead of inside a black-box language model where it can only be asserted.
Where does the data for Hugo's simulations come from?
The data comes from observing what real people do and say in public. For our work in fashion, this means mapping what consumers actually wear. As detailed in our research on mapping consumer audiences, we build a map that connects garments to the people who wear them. For our New York map, we identified 832 distinct garments across 876 real accounts from 4,464 posts. We found that only 2% of the images people post are explicit "outfit posts." The other 98% is the unprompted, everyday context that traditional research misses.
We reach these people not through keyword searches, but through chain referral, a survey science method for finding populations that are not listed on any register. Starting from a few seed accounts, we follow the real connections between people to build out the network. This gives us a map of the audience's real social structure, which becomes the substrate for the simulation.
How do you measure if a simulation is accurate?
We test every method on a world where we already know the answer. This is called backtesting. We take a historical period of data, use the first part to train our model, and then test its ability to predict what actually happened in the second, held-out part. This is the only honest way to validate a predictive model.
For our audience simulations, we did this with a real German audience of 40,212 people. We trained our model on adoptions that happened up to a certain date and tested its ability to predict the 7,151 adoptions that happened afterward. We measure accuracy using a metric called AUC (Area Under the Curve), which scores how well the model ranks a person who will adopt a trend above a person who will not. A score of 0.5 is random chance; 1.0 is a perfect prediction.
Can you give a real example of Hugo's simulation beating a language model?
Yes. In our published research, we ran a head-to-head test. We took the audience of 40,212 real people and their measured behavior. We gave our mathematical model the data on the network structure and past adoptions. We gave a frontier language model the same information, translated into a detailed text prompt. Both models were then asked to predict who would adopt which trends next.
Our fitted mathematical model achieved an AUC of 0.86. The language model, given the same context, scored 0.78. Both are significantly better than chance, but the model built on observed social physics was meaningfully more accurate than the one built on prompted reasoning. This is not a failure of the language model; it is a demonstration that it is the wrong tool for this specific job.
What commercial questions can an audience simulation answer?
The goal of a simulation is to let a brand test a decision before committing to it. The questions it can answer are the ones that a brand argues about in a meeting, but now backed by a quantitative forecast.
- Product Release: If we launch this new sneaker, which communities will adopt it first, and how far will it spread?
- Marketing Campaign: Will this campaign message resonate beyond our core followers, or will it die in the first circle?
- Creator Partnership: Which specific creator will have the most impact on driving adoption for this particular product within this target audience?
- Price Change: If we raise the price, which customers will leave on budget, and which new customers might arrive because the price now signals the status they want?
How does this compare to traditional consumer research like surveys or focus groups?
Traditional research is elicited. It asks people questions. This is valuable, especially for testing a concept that does not exist yet. But it has three structural limits. It answers the question someone thought to ask. It arrives as a report, weeks or months after the question was asked, often after the decision window has closed. And it returns a document that a human still has to interpret and turn into a decision.
An audience simulation built on standing infrastructure is different. It is always on, because the audience model is live. It can surface a finding without being asked, because it can detect when a new trend is propagating. And because it is part of an agent-based system, it can report its finding directly into a brand's Slack, recommend a decision, and close the loop between insight and action.
What about trend-watching tools that use keywords?
Keyword-based tools, like social listening platforms, are designed to measure things you can already name. If a brand needs to track mentions, monitor a campaign hashtag, or measure sentiment around a known topic, these tools report on the volume and context of those specific terms. They are instruments for counting known quantities.
Their structural limit is that they cannot find what has no name yet. A new aesthetic, an unmet need, or a novel use for a product emerges in conversation before it has a tidy label. Because our infrastructure maps the audience first, at the level of people and garments, we can detect these patterns as they form, before they are coherent enough to be a keyword. We find the trend by observing the behavior, not by searching for the term.
Can a simulation predict who will be influential?
Yes, and it often finds that influence does not correlate with follower count. In our study of the German audience, one account dwarfed all others in followers. But when we measured influence by looking at who actually led whom in adopting new trends, that account's score shrank. A different account, with only 99 followers, was the source of 596 separate adoption pairs in the network. They were a true originator, but a follower count would never have revealed it.
This is a critical function of a simulation grounded in a real network. It moves beyond vanity metrics to measure effective influence: the proven ability to cause a behavior to spread. For a brand deciding who to partner with, this is the difference between paying for reach and paying for results.
What are the limitations of audience simulation?
Our approach has clear and important limits. The first is that we can only simulate what we can observe. It works only for consumer categories where people talk and show their behavior in public. It is not suitable for private B2B populations or for decisions that happen entirely offline.
The second is that it cannot test a pure counterfactual for a product that has never existed. Because our model is trained on real adoptions of real things, it needs some historical data to work from. It is poor at concept testing for a completely novel invention with no precedent. For that job, traditional elicited research like a survey or a focus group is genuinely better.
Finally, the signal we read skews toward the delighted and the annoyed. The frequency of mentions is directional, not a representative measure of incidence in a whole population. We have not solved this, and that is the honest state of this layer.
Does this work for products other than fashion?
Yes. We began with fashion for one reason: the data is unusually rich. Almost every public photo or video of a person contains clothing, providing a dense and continuous stream of signal to build our models. It is where the data is best, not a pilot or a testbed.
The underlying infrastructure, however, is general. The same method for mapping an audience, modeling its network structure, and simulating cascades can be applied to any consumer category where people share their experiences and preferences publicly: beauty, home goods, food and beverage, travel, consumer electronics. The core challenge is the same across all of them: to understand what customers want to buy next, before the sales data catches up.
How does a brand get started with an audience simulation?
It starts with a specific audience. For example, a brand could use our map of New York, built from 876 real accounts and 4,464 posts, to test a product launch. The brand provides the product details, and the simulation runs a cascade on the real network graph of that audience. The output is not a dashboard, but a predictive answer to a commercial question. For instance, the simulation would forecast which specific communities within the New York map would adopt the product first, how far it would spread, and which individuals act as the true originators of the trend, moving beyond simple follower counts.