The Data Trap: Deterministic vs Probabilistic
Programmatic TV promises advertisers something traditional television could never fully deliver: the ability to target very specific groups of people or households. But there is a potential problem. Not all audience data is created equal, and failing to understand the difference between deterministic and probabilistic data can lead advertisers to significantly overestimate the precision of their targeting.
Deterministic data is based on something that is actually known. A household may have purchased a particular product, registered a vehicle, subscribed to a service or provided information directly. There is a verified connection between the data point and the audience being targeted.
Probabilistic data is different. It uses signals, behaviours and modelling to determine that someone is likely to possess a particular characteristic or behave in a certain way.
There is nothing inherently wrong with probabilistic targeting. In fact, it can be extremely valuable. The danger comes when advertisers treat a probability as a fact.
Consider a campaign targeting people who have recently bought a new car. A deterministic dataset might identify households known to have registered or purchased a vehicle recently. A probabilistic audience might identify people whose online behaviour suggests they are likely to have bought one.
Those audiences could perform very differently.
The problem becomes particularly important when comparing programmatic TV opportunities. Platform A might offer an audience of 500,000 households based on deterministic data, while Platform B offers two million households apparently matching the same description. On the surface, Platform B looks considerably more attractive.
But does it really have four times as many relevant households?
Possibly not. It may simply be applying a predictive model much more broadly.
This also makes CpM comparisons potentially misleading. Paying £30 per thousand impressions against a tightly defined deterministic audience may represent better value than paying £20 against a much larger probabilistic audience if a significant proportion of the latter don’t genuinely meet the targeting criteria.
Advertisers therefore need to look beyond the audience label.
When assessing programmatic TV proposals, ask where the data originates, whether the characteristic is directly observed or inferred, how recently it was collected, how frequently it is refreshed and what proportion of the audience is modelled.
Programmatic TV can deliver extraordinary targeting capabilities. But greater sophistication also creates greater potential for misunderstanding.
The question advertisers should ask isn’t simply, “Can you target this audience?”
It is: “How do you know these people are actually in it?”