Most AI personas you'll see today are stories. A language model is handed a job title and a vague brief, and it produces a character sketch with a name, a stock photo, and a list of plausible preferences. The output reads well. It is also, more often than not, unrepresentative of the people the marketing team is actually trying to reach.
There is a quiet reason this is true, and it isn't the fault of the language model. The model has no information about the population. It has the prompt, its training data, and whatever it can plausibly infer. Plausible is not the same as accurate. Two prompts a few words apart can yield very different fictional people. Neither one is grounded in anything you can point at.
We built Cambium AI's personas a different way. The synthetic populations behind the product are sampled from the American Community Survey and the Public Use Microdata Sample, the same data sources statisticians, demographers, and the US Census Bureau use when they need a representative slice of the country. When you ask Cambium AI for a persona for, say, professional buyers of business software in Greater Boston aged thirty-five to fifty, the persona is not a story. It is a person drawn from a synthetic population that matches the real one in proportion.
That matters in two practical ways.
The first is base rates. A great deal of marketing strategy turns on knowing how common something is. How many of your target buyers commute by car. What fraction live in households with children. The income distribution at the percentiles you care about, not just the mean. LLM-generated personas tend to give back the median of every attribute, smoothed and centred, because that is what language models do by default. They produce confident averages. Real populations are not centred. They are skewed, segmented, and full of structure. If your strategy depends on a tail, an LLM-only persona will not help you see it.
The second is verifiability. Anyone in the room can check whether a Cambium AI persona is consistent with the underlying public data. The sample sizes, the margins of error, the source variables, all of it is traceable to the ACS. There is no equivalent paper trail for an LLM-generated persona. You either trust the model or you do not. Marketing teams that present personas to product, sales, or finance benefit from being able to point at a source.
This isn't an argument for replacing instinct or qualitative work. Personas built on real data still need an interview programme, a sales-call review, a competitive read. The point is to start from a representative base and add nuance, rather than start from a story and hope it is representative.
If you want to try this in practice, the personas product is built around three short prompts: who you're trying to reach, where, and at what life stage. The output is a persona panel grounded in ACS variables, with the underlying numbers one click away.
The takeaway for marketing teams is straightforward. AI personas are useful. AI personas built on a verified population are useful and right. Spend the extra minute on the second.
Create your personas for free here.