As a post-cookie alternative, contextual is gaining traction. But as we backtrack to targeted advertising, questions linger. Is it as effective?
Kasper Skou, CEO and Co-Founder, Semasio, believes it is a positive move that shifts the onus from publishers to advertisers. He speaks about two approaches of contextual targeting: top-down, based on the actual content of the page, and bottom-up, starting with a set of users and analysing their behaviour to project it onto pages.
Marketers are worried about a world without third-party cookies. How can contextual targeting help?
Contextual targeting focuses on the page and leverages no user-level identifier, which the third-party cookie is an example of. It is therefore being heralded by many as the future of data-driven advertising in a post-cookie world. We believe that perspective is a little simplistic and see great potential in the synthesis between the user and page-level targeting. The basic premise of this synthesis is to use post-cookie identifiers (like Unified ID 2.0, LiveRamp’s RampID, ID5’s Universal ID and others) to track user behaviour and then leverage this behavioural analysis to project the audiences onto contexts. We call this product Contextual Audience Extension, and other companies have similar propositions, which they’re rolling out as we speak.
Enterprises in the Middle East must cater to multilingual audiences; Is language a barrier for NLP models during semantics analysis? Where do you see this space moving?
There are two schools of NLP, the lexicographical and the statistical. The lexicographical approach is dependent on models, which formalise a certain natural language for processing. It requires a large, upfront investment for each language to develop these models.
The statistical approach treats a natural language as a set of ‘tokens’ and automatically builds a statistical language model. That means a new language can be added in a matter of days with minimal human involvement. One of the new developments in NLP fuse the two, basically combining the automatic generation of the statistical language model with the higher fidelity of a lexicographical model. This is a focal topic of R&D for us.
What are the advantages of the semantic approach over the more traditional taxonomical or keyword approach?
Here we need to be precise in our vernacular: A keyword-based approach can be semantic, but we choose to call our approach semantic, because it goes beyond keywords and puts them in context. The same keyword can have multiple meanings — and the same meaning can be manifested by multiple keywords. When you put keywords in context, they are disambiguated, enabling the client to target the right meaning.
Taxonomical approaches rely on pre-categorisation to capture the meaning of a given target. The problem is that it narrows your targeting possibilities down to the set that your vendor was able to imagine you might conceivably need in the future. Thinking about it this way helps us realise how reductive this approach is.
Which industries stand to gain the most from adopting this approach?
The semantic approach makes no assumptions about the topics clients want to target in the future and are thus very powerful in industries needing very specific targeting. In other words, the more specific your targeting requirements, the lower the probability that you will find what you need off the shelf from a taxonomically-based provider. Big verticals for us are automotive, pharma, finance, consumer tech and CPG, where clients go beyond what’s on the shelves to tailor-make targets — sometimes down to the individual marketing message level.
What’s in it for the publishers?
With the semantic approach, publishers don’t have to worry about how to package their inventory to satisfy individual client demand. Because the approach preserves all the information the individual internet user consumes, the publisher can satisfy a much broader range of client requirements than if the inventory had been pre-packaged like it is with a taxonomical approach.
How can marketers get started on using contextual targeting? Give us with a checklist.
There are basically two approaches to contextual targeting: top-down, based on the actual content of the page, and bottom-up, starting with a set of users and analysing their behaviour to project it onto pages. For top-down it is relatively simple: if you’re using audiences today, you expand these audiences semantically to pages. Let’s take a simple example: You want to target people who want to quit smoking or vaping. Your audience consists of users who consume this topic. Now create a contextual target, which includes the pages that contain this topic. We call this unified semantic targeting. For bottom-up, you simply start with the same set of users, which form the training set for your current look-alike modeling, and use their behaviour to create a Contextual Audience Extension.
The most important thing is to get started now. We need to go up this learning curve together in order to be ready for the post-cookie future. Don’t be one of the proverbial frogs left in the pot when the water boils.