“If you throw a lot of stones at a target, you’re bound to hit with one” - Preciso on how AI is freeing advertisers from wasteful targeting

The dirty open secret of digital advertising is the incredible volumes of wasted ads that are bought in search of every successful conversion. Anoop Ramachandran, chief technology officer of smart-bidding platform Preciso, explains how machine learning enables precise media buying – and how low-waste targeting can transition to the cookieless age.

Retargeting has traditionally been a remarkably approximate science, with a very low hit rate. Clearly, you feel it can be improved?

Basically, we came from an advertiser environment, where we were managing multiple direct clients and were very, very close to advertisers. And when we started our DSP, what we realised is that these platforms were actually built for publishers, and their objective is to make us spend money with those publishers, not necessarily to provide the best inventory for the advertiser. That frustrated us a lot.

So that’s where we started to think about our own data, and buying directly from exchanges and so on. So while building this platform, which we call our Smart-Bid Platform, one of the main requirements for us was to build it for advertisers, not for publishers. We didn’t want to buy all the traffic, but just the traffic that we really, really want. That’s a core role for us.

How do you know what traffic you really want?

If you throw a lot of stones at a target, you’re bound to hit with one. And it’s the same with retargeting. If you bid a lot with some sort of basic algorithm, you’re going to target the relevant person in the end, but there will be a lot of waste, and the advertiser pays for that.

Using machine learning, but also using human input, we look a lot harder at the traffic that is fed to us, and we are extremely selective with our bids.

In fact, if we bid, we plan to win that particular bid, and 90% to 99% of bids we send out, we actually get. It varies by country – the US is tougher, but in Europe and Asia, very high figures are very possible.

In the end, the results we get, in terms of conversions and sales, are almost the same as platforms that bid on everything, but the advertiser would have had to spend a lot to achieve that, whereas we have a better aim, and a lot more of our stones hit the target. We buy fewer impressions, but we maintain the same or higher conversion rate as anyone else, without spending every last bit of an advertiser’s budget.

What is the role that machine learning plays in this?

Piero Pavone, our CEO conceptualized the idea and authored our bidder’s ranking system to qualify/grade users and ad-slot and worked hand in hand with the machine learning team based in Portugal. It’s a third-party team led by one of the big people in the machine learning area, and on the back of that we train our own team of data scientists in India.

We have built our bidder’s rank system, which is a machine learning system with a human touch to it. Ranks one to five are factors that a human can change, according to what we want from a particular campaign. The team can move those values according to their needs to make the campaign perform better.

Then 5a to nine is the machine learning prediction – a mathematical formula that creates a probability about all these ranks, before making the final decisions about what we’re going to bid on. And each of those ranks filters in a different way. So, rank 5a identifies good users and bad users – it will identify bots and remove bots, or users that have only come only one time to the website but do not travel to multiple pages – all these kinds of filters. Rank six calculates the probability of whether a person will click or not.

Rank seven calculates the probability of a sale, and then eight calculates the optimum bid amount for this particular traffic. That considers a lot of stuff like time of the day, the weather and season and the location. And finally it comes to rank nine, which is product recommendations – identifying products that are good for this particular user based upon their previous behavior. And then all this data is put through a mathematical algorithm with ranks one to five, takes everything into consideration and makes a prediction about the value of that user and makes the bid.

How specifically can an advertiser tailor a particular campaign?

Well, we have a thing called a rule engine with defined rules for each campaign. We call it retargeting plus.

So, when we have campaigns for hotel clients, or when there is a seasonal sale going on, we can write different rules for that particular campaign that control the conditions that have to be met before we will bid. You can even set rules for a specific person – for example, if that user has already been in the checkout page, do not bid for three days. Or if the user has just booked a hotel, we will target them for the return flights rather than for hotels.

And then of course there are all the more straightforward types of targeting – targeting people on 4G connections only, or people on the latest iPhone or the latest Android only; blacklisting and whitelisting of domains; using exchanges, not using exchanges. We are also doing increasing amounts of geo-targeting.

Is this something any advertiser can plug into, without necessarily bringing data of their own?

At the basic level, you can just plug in.



But, the ones who are most interested are the big players, who have a lot of retargeting experience and are trying to find ways to improve their efficiency and get more from their budget, and often we’ll work with their first-party data as well, and combine that with custom machine learning to create the rule engine.

Obviously, aspects of what we do involve cookies, but from an R&D perspective we are working hard right now on natural language processing for contextual, and also on converting first-party data into prediction systems. Those are things that will be important in the coming years, as we move into a cookieless environment.

Also Published in: AIThority