Next Event: Gaming & Advertising, 10th June Register Your Place>

Can Conversion Modelling Solve Attribution Headaches in the Post-Cookie World?

Tim Cross  03 September, 2020

The death of third-party cookies in digital advertising is set to pose all sorts of headaches for marketers. The industry is scrambling to find alternative solutions for things like behavioural targeting and frequency capping.

And the death of cookies will make it harder to measure advertising’s impact too. Without cookies, measuring conversions (when a user interacts with an ad, and then goes on to buy the product/download the app/sign up to the newsletter etc) becomes much harder. And this causes problems for attribution.

Third-party cookies provide a straightforward method for tracking conversions. For example, they can show when a user has visited a publisher’s site where an ad was shown, and track them from that ad across to the brand’s ecommerce store, where they make a purchase.

But without third-party cookies, it becomes harder to link up this activity.

One solution to this problem is ‘conversion modelling’, where machine learning is used to estimate conversions in cases where they can’t be measured.

Philip McDonnell, director of product management at Google, posted a blog last week outlining what conversion modelling is. “Conversion modelling refers to the use of machine learning to quantify the impact of marketing efforts when a subset of conversions can’t be observed,” he said. “With a modelling foundation in place, observable data can feed algorithms that also make use of historical trends to confidently validate and inform measurement.”

Essentially, in cases where cookies can’t be used to track users across sites, marketers can plug whatever data they have about the audience into a machine learning algorithm to estimate conversions.

This sort of modelling is already used in cases where cookies aren’t available. For TV attribution, advertisers aren’t able to directly track viewers from TV sets where they’ve seen ads across to other devices where they might buy the product. Or even if a user sees an ad on a web browser where cookies are available, they might buy the product in a physical store.

Ron Jacobson, co-founder and CEO of attribution specialist Rockerbox, said he expects these sorts of techniques to carry over to the digital world as cookies disappear.

“The death of third-party cookies will definitely accelerate the use of machine learning in conversion measurement,” he said. “But there are still some pretty clean data sets that you’re able to use. Things like click based-data, actions happening on your website, information means you buy a customer post-purchase, promo code data. So there are still deterministic data sets that you can use. And I really think the future of attribution measurement overall will be using these dataset, and then leveraging machine learning to fill in the blanks.”

Jacobson said conversion measurement and attribution techniques used in the TV world can carry over to digital environments, albeit with some changes.

“There’s a bit of adapting, because the datasets you’re using are very different,” he said. “Every channel has very unique datasets, so there’s a three step process. First you have to normalise that data. Then you have to figure out how you measure if someone has engaged with that media. And lastly you have to work out when someone did engage with that media, how much did it contribute to the conversion?”

JP Pereira, SVP of marketing science at video measurement company VideoAmp, agreed that conversion modelling will become more commonplace. “I think machine learning is going to become even more critical, because when we have less availability of persistent IDs, we have gaps in the data,” he said.

Pereira said that machine learning actually already plays a role even where persistent IDs are available. Companies like VideoAmp use modelling to predict whether the ID they’re seeing actually is a persistent ID or not, and therefore whether it can be used to link up a consumer’s activity or not.

And modelling is also used to try to measure the effect of an ad’s creative material versus the effect of the ad’s targeting. Machine learning can help advertisers understand if an ad drove results because the creative was impactful, or because it was accurately targeted at people who were likely to buy the product anyway.

But Pereira said that instead of using conversion modelling, advertisers may simply just buy inventory wherever they are still able to track users.

“When we talk about traditional display and video impressions that are bought on the open web, those are going to be harder to track,” he said. “But if you started looking at the walled gardens that have emails as identifiers, like Google, Facebook, Hulu, Pandora, and Spotify, those companies are serving ads to authenticated users who can be identified by their email address.”

“So now, multi-touch attribution becomes less about relying purely on the open web and tracking impressions through third party cookies or device IDs. It’s really about establishing those relationships with publishers that are able to provide impressions from authenticated users in a safe, privacy-compliant way,” he added.

Go to Top