How do you personalise ads for people you know nothing about? The honest answer: at the individual level, you don't. But you can absolutely understand the people who are interested in a category, and match new users to a suitable persona from a handful of signals. That is the heart of a workable cold-start strategy.
How hard the industry is hunting for alternatives to classic audience data shows in recent reporting from Digiday: synthetic audiences are moving from experiment into everyday planning and strategy work. Aveeno Body used Vurvey's synthetic research to cut the time needed for a strategy brief from around 30 days to three. Goodway Group uses Arima's ‘synthetic population’ of consumers from the US, the UK and Germany for media planning.
What is the cold-start problem in advertising?
The term comes from the world of recommender systems. A system that learns from behaviour needs behaviour to learn from – and a new user doesn't have any yet. So the system serves up the average: bestsellers, generic messages, the ad that works a little for everyone and properly for no one.
In advertising, the problem gets worse. Several browsers block third-party cookies by default, consent banners are part of everyday life in Germany, and not everyone clicks ‘Accept all’. The classic way out, bought-in segments, has become trickier under data protection law and reveals little about why anyone buys anyway. In the GDPR-shaped DACH region (Germany, Austria and Switzerland), ‘we'll just buy in more data’ is no longer a strategy.
Why more data doesn't solve the problem
Data quality research has a sober formula for this: data is only as good as its fitness for purpose. Accuracy alone isn't enough; data has to be relevant, current and suited to the specific question. In the literature, this is known as ‘fitness for use’.
A bought-in segment such as ‘women, 25 to 34, interested in beauty’ is precisely worded. For the question ‘What would make this person try our product?’ it is nonetheless almost worthless. It describes attributes, not motives. Engagement is born from understanding, not numbers.
For the cold start, that means the crucial resource isn't the individual user's history but a robust picture of the motives, worries and language in the category. And that picture can be drawn from publicly available content.
How are personas built from public content?
People say a surprising amount in public about why they buy – or don't: in comments under brand posts, in product reviews, in creator videos, in forum threads. This content can be collected, clustered by theme and tension and distilled into personas – with real quotes instead of invented profiles.
The difference from the classic workshop persona: it is based on what people actually say, not on what a team assumes about them. Digiday describes a related approach: for the brand Sacheu, Limitless Labs analyses thousands of TikTok and Instagram videos from its creator ecosystem to build personas.
Important for the German market: personas like these work at the level of patterns and groups. They don't need a profile of individual users to choose a first message. The legal assessment of your specific set-up still belongs with your data protection adviser.
What approaches are there to the cold start?
| Approach | Data basis | Strength | Weakness |
|---|---|---|---|
| Bought-in third-party segments | External profiles, cookies | Quickly available | Little insight into motives; tricky under data protection law |
| Contextual targeting | Environment, time, location, device | Works without personal data | Says little about motivations |
| Zero-party data | Answers from quizzes or onboarding | Direct and consent-based | Only from those who take part; friction in the funnel |
| Synthetic audiences | Models based on survey, statistical and panel data | Fast simulation of reactions | Pull towards the average; quality depends on the input data |
| Personas from public content | Comments, reviews, creator content, competitor ads | The category's real motives and language | Needs signals to match individual users |
No approach solves the problem alone. In practice, personas from public content supply the hypotheses, contextual signals the matching and first-party data the confirmation.
How do you put cold-start personalisation into practice in six steps?
- Collect category conversations. Comments, reviews, creator content and competitor ads from recent months – not just about the brand, but about the whole category.
- Build personas and tensions. Cluster the content by motive and worry. Three to five clearly distinct personas are enough to start with.
- Write one hypothesis per persona. Which message, which hook, which offer fits? Develop at least one dedicated creative variant for each persona.
- Define matching signals. Landing page, search term, context, campaign source, first clicks: decide which signal points to which persona.
- Serve with exploration. Reserve part of delivery for variants that don't fit the assumed persona. That way the system keeps learning instead of merely confirming its own assumption.
- Feed first-party data back in. Every interaction refines the matching. Update the personas regularly, because conversations change.
How useful are synthetic audiences for the cold start?
Quite useful – if you know their limits. In Digiday's reporting, even those involved voice reservations: Vurvey CEO Chad Reynolds says large language models deliver the top of the bell curve, in other words the most common answer. Julie Clark of TransUnion warns about personas built on poor input data. And Stef Hoffman of Code & Theory puts it neatly: the results inform judgement, but they don't replace it.
This is exactly where public content earns its keep. It isn't a simulation; it is evidence. If a persona claims that people in your category mainly fear buying the wrong thing, you should be able to see the posts that claim comes from. A persona you can't trace back is just an assertion.
Defy Monitor takes this route: audience content is clustered into personas and tensions automatically, and planning starts from real conversations rather than demographics. It is one building block of the Defy platform, which gives independent agencies the same analytical capability the big networks have spent heavily to build.
How do you know your cold-start strategy is working?
Compare the persona-based variants with a generic control group. The most telling metrics are new users' early numbers: interaction with the first ad, depth of the first visit, time to first conversion. If the persona variants aren't ahead, that isn't failure but a clue: either the personas are off, or the matching signals are too weak.
