Why this manifesto exists, and why we wrote a second one for SaaS founders.

Consumer research is an industry built on a single, rarely-examined compromise: the research is slower than the decision. A brand manager deciding whether to buy €4M of Q3 inventory has to choose between shipping blind or waiting eight weeks for a study whose findings will land after the decision is already made. Everyone in the market has quietly accepted this. The research happens. It is dutifully filed. The decision is made on instinct anyway.

Synthetic consumer research, the use of statistically-grounded populations of simulated humans to predict reactions to stimuli, dissolves that compromise. Not because simulated humans are better than real humans (they are not), but because a calibrated simulation produces a usable signal at the speed of the decision. The question worth asking is no longer “is synthetic research viable?” It clearly is. The question is “which synthetic research is actually research, and which is decoration with an LLM on top?”

For forty years the research industry lived in a world where each of its main tools, focus groups, survey panels, in-home ethnography, retail tracking, had obvious limitations, and no emerging technology meaningfully threatened any of them. That world ended in 2023.

Three things happened in rough succession. First, language models crossed a threshold where a well-prompted model could convincingly hold a consumer persona for a long conversation. Second, embedding models got good enough to retrieve and anchor behavioural fragments to a generated agent in real time. Third, and this is the one the industry discusses least, the unit economics of frontier inference flipped. Generating 10,000 agent reactions, end-to-end, with a multi-model ensemble and a calibration pass, became something you could do for less money than a single research-industry overnight turnaround used to cost.

The combination of those three things is what makes synthetic research a category, rather than a parlour trick. Any one of them on its own was a novelty. Together, they redraw the industry's cost curve. A cost curve re-draw is how every software market actually begins.

The first generation of synthetic research products demonstrated plausibility. Ask a model to roleplay a 34-year-old fitness-tracker buyer, and it will do so compellingly enough to fool an untrained reader. Put a hundred such roleplays in a dashboard and you have something that looks like insight. Sell it at enterprise prices and you have a company.

What this generation did not do, because it was not cheap, easy, or at that moment competitively necessary, was prove that the plausible output was useful. “Useful” is a narrow word, and we mean it narrowly: a prediction is useful only if its error bar, against ground truth, is small enough that it can meaningfully improve a decision.

Most of the products in this category cannot answer that question. They cannot answer it because their error bar is, empirically, larger than the decision margin of the problems they are sold against. They cannot answer it because they have not, in most cases, even measured it against dated public ground-truth data. They charge €60,000 a year, put an 89% accuracy claim on the homepage, cite no sample, and go to trade shows.

This is the opportunity. The research industry has survived on a compromise for four decades. The first wave of synthetic tools exploited the speed half of that compromise and quietly dropped the accuracy half. There is a whole profession of researchers, often the sharpest people in the building, who have spent two years trying every synthetic tool on the market and concluding that none of them pass basic methodological scrutiny.

That audience is not going to be convinced by another “90% accurate” homepage claim. They are going to be convinced by accuracy numbers that are dated, sourced, auditable, and, when they are bad, visible anyway.

Every synthetic research platform is, under the marketing language, the same architecture: a population-generation step, a stimulus-reaction step, and an aggregation step. What separates them is what happens inside each step, and whether those internal steps are known, numbered, and accountable.

Our working thesis is that the accuracy stack, not the interface, not the dashboard, not the go-to-market motion, is the product. A customer is not buying a tool. They are buying an error bar. Everything we spend engineering time on is, directly or indirectly, a reduction of that error bar.

Which is why Prism is built as an explicit eight-layer stack: real-data seeding, multi-model ensemble, calibration, revealed-preference weighting, behavioural consistency, noise injection, distribution-shape matching, public validation. Each layer closes a specific, named failure mode of the layer below it. Each layer is individually measured. If any one of them were removed, the stack would visibly degrade against ground-truth, and we'd rather know that.

The alternative, “we use a proprietary model”, is not a product. It is a flag to walk away from.

A smart engineering team at a consumer-goods company could, in theory, build something like Prism. The components are all documented. The models are all available on API. The calibration math is sophomore statistics. The infrastructure bill runs into four figures a month, not seven.

They will not, and it is worth being honest about why. First: the ground-truth datasets needed to calibrate are not sitting in a single S3 bucket. Assembling them is a year of patient work and a legal process with each source. Second: the weekly re-calibration cadence requires a dedicated pipeline, monitored, alerted, and re-deployed, not something a general-purpose data team picks up between priorities. Third: auditing clusters against new ground-truth arriving from disparate datasets requires a team that speaks both LLM evaluations and survey methodology, and there are perhaps a hundred people in Europe who speak both fluently.

Synthetic research is a specialist's tool. It will consolidate into specialist companies the same way database infrastructure consolidated into Snowflake and document infrastructure consolidated into Notion. The question is which specialists.

A Prism Pulse test costs €2,000. An equivalent engagement with a traditional research supplier or a first-generation synthetic incumbent costs between €15,000 and €40,000. We are, on pretty much any axis you want to measure, an order-of-magnitude cheaper.

This is not a discount. It is not a land-grab. It is the actual cost structure of synthetic research run competently, revealed for the first time, in a category that has been pricing against focus-group-era assumptions.

Every market that re-prices by 10× does the same thing: volume expands by roughly the same factor. Teams that were commissioning four studies a year suddenly commission forty. Teams that were never research customers because they couldn't justify the €20k minimum become customers at €2k. Research moves from being the big-question question to being the many-small-question question. Which is a different shape of business.

We think this is the inevitable destination of the category. Price collapses precede usage explosions precede category consolidation. We want to be the default tool by the time that wave peaks.

We will not publish an accuracy claim we cannot audit. If we say a cluster is 88% accurate, there is a dataset, a date, and a sample size on the validation page. If a cluster drifts below 80%, we pause it, silently letting a degrading cluster run is the single fastest way to destroy a research brand's credibility, and that credibility is the whole asset.

We will not claim to replace real research. A well-run synthetic study approximates what a well-run real-panel study would tell you, at one-tenth the cost and one-hundredth the time. It is a precision tool for mid-consequence decisions. For the most consequential decisions, a rebrand, a regulatory filing, an M&A-scale positioning bet, real humans should still weigh in. Prism's role is to make those real-human studies better targeted, not to eliminate them.

We will not build dashboards that look impressive at a glance but lie by omission. No dashboard will ever show a single headline number without the distribution, the sample size, and the last-audit date.

Within three years, every brand marketing team worth the salary it is paid will run synthetic research as a pre-flight check on every consequential decision. This will not be controversial. It will be as default as spell-check. The teams that do it will simply win more of the decisions they run.

Within five years, every major consultancy will have a synthetic-research practice. It will be hollow theatre for one or two years, prestige firms signing white-label deals with whatever vendor has the most presentable dashboard, and then it will get serious, because their clients will demand it and their competitors will have built it.

Within ten years, the phrase “synthetic research” will sound as quaint as “online research.” It will just be research.

We are early. The company is small. The product is good today and will be meaningfully better in six months. The accuracy stack gets one new layer per quarter, announced publicly with the audit data. The validation page gets one new cluster per month until it has every cluster our customers need.

If you are a researcher who has spent the last two years frustrated by the gap between what synthetic tools promise and what they deliver, we are the company you were hoping would exist. If you are a brand leader who has been told by your insights team that every synthetic platform they trialled failed on methodology, we are the one you ask them to trial next. If you are a competitor and some of the things we have written here apply to you, you know what to do next; publish your accuracy data, date it, source it, and stop quoting 92%.

Synthetic research is inevitable. The question is whether the category will be built by people who care about accuracy or people who care about optics. We are betting on the first. If you are too, talk to us.