There are two core approaches to cross-device tracking and attribution – probabilistic and deterministic, but how different really are these two methods? Here Vinnie Khurana, head of sales UK at Admo.tv, discusses the ins and outs of the two approaches, the specific benefits of each, and whether the two can coexist.
Khurana will be on stage at New Video Frontiers in London next week, on a panel discussing the topic ‘Unifying Cross-Platform Audiences: Are We There Yet?’
There is a division in the TV analytics market between two approaches: deterministic and probabilistic – what’s the difference between the two?
The deterministic approach is based on TV set-top boxes and their associated household data. So you have tech sitting within the set-top box which allows whoever is doing the analysis to understand what you’re watching, when you’re watching it, and in what context. They will have some pre-determined household data , and they can use that to understand the impact the ad is having so that they can target TV more accurately. It’s more associated with the addressable side of TV.
There are some prominent players now like Sky AdSmart and Samba TV and a few others. They each do it differently, but they use set-top boxes and household data, and extrapolate out from the audience sat within that panel.
With the probabilistic approach, we’re instead looking at the business impact. We don’t have tech within the household itself, instead we analyse a live stream. So we use image and audio recognition to identify when an ad is playing and in what context (first in break or last in break, who the advertisers before and after are, what competitors are doing on other channels). Then we do spike analysis – we look at the behaviour of a brand online.
So for example, when a BMW ad plays on TV we see an increase in online activity straight after. We then add an attribution model and look at behaviour at scale to identify someone who has been exposed to TV, as opposed to someone who hasn’t. So if there’s a spike right after an ad, we might say “of those 200, we think 50 people came from TV” via our probabilistic modelling.
What are the main targeting features and goals when it comes to a deterministic and probabilistic methods?
On the probabilistic side, we look at user-centric analysis. So we’re looking at what someone who was interested in the ad and has gone online is likely to do, so that’s things like downloads and conversions. The idea is partly to have like-to-like measurement against digital. One part of our service is we do econometric modelling, so we look at the entire media mix to see where TV sits within the grand scheme of things. So over the six weeks that your TV campaign has been running, what has been the direct effect and what’s been the indirect effect?
With deterministic, what you’re doing is looking to get better targeting for your TV campaigns. When your ad plays on a given channel, what does that mean? What kind of impact does that have, is it driving more sales? So it’s for buying TV more efficiently.
What type of activation/amplification can you do with both methods?
A lot of TV is bought quite far in advance, maybe six months to a year. So the findings from both methods can take a while to implement – if we see one creative is resonating with people in Manchester a lot more and driving sales, it can take a while to change the media plan to reflect that.
With amplification, we’re using digital as a way to amplify your TV campaign, which might be increasing reach or re-targeting. We think that’s more common on the probabilistic approach, where we might link a display or video campaign with a TV commercial. Going back to the BMW example, every time a BMW ad plays you may choose to bid-boost your BMW digital campaign, so that if someone is searching for 4x4s, you’re highly ranked.
Obviously to do that all the time is expensive, so with amplification you can choose to do this at strategic times based on your TV campaigns. And you can do the same thing with display and video, and here you can do re-targeting too. So maybe if someone goes online after they see the TV ad, you might re-target that person four or five days later with a discount, because you know they’ve shown intent.
And you can also do these things against your competitors’ campaigns, where every time their ad is on, you might choose to boost your visibility online.
What are the pros and cons of both approaches?
With deterministic, you know exactly who has seen the ad. So through the set-top box you know who is in the household, you know when the ad is on and who has watched it. When you’re targeting, it’s people who you know have seen the ad. One of the other big pros of the deterministic approaches is that its usful for addressable TV ads. It is a contextual way of showing an ad to an individual within a household.
The flip side is that it’s based on a sample of your audience – it’s dependent on having access via the set-top box. So if everyone you’re looking to target doesn’t have your tech in their set-top box, then you have to do some modelling and extrapolate out, which means its a panel.
With probabilistic on the other hand, it’s in its nature that you’re saying (with a high degree of accuracy, but with a five percent margin of error), that you’re looking at the likelihood of someone seeing an ad, rather than knowing exactly who has seen the ad. But it counts everybody – this approach measures everyone who has had some sort of digital contact with the brand.
The other pro with the probabilistic approach is how it aggregates and analyses anonymised user data gathered from distinct behavioural patterns. As I described earlier, using this information it is possible to do user-centric profiling, and from there you can create a TV Segment within your own attribution modelling (for example Google 360) opening the doors for more amplification and conversion mechanisms.
Can the two approaches coexist within the market?
They are very different approaches, so I think they can coexist. You might be using AdSmart as part of your TV campaign where you’re looking to target specifically, and with the probabilistic approach you might be wanting to measure the efficiency of that. So you can have both working in tandem, you can use the two together to get an understanding of who exactly is watching your ad on TV, what the household is doing etc, and then also look at the impact on your business in terms of sales, downloads, or whatever metric you use.
So they don’t need to live exclusively in my opinion, they’re just different.