When looking back on summer 2018, it’s hard to ignore the optimism that’s been in the air. Sunny weather? Check. England football triumph? Almost! AI as the next big thing in digital marketing? Try and count the number of articles, blog posts and sound bites that you’ve encountered over the last month which cite AI in a hype-tastic way.
Now we’re all for a bit of well-reasoned optimism, and there is no doubt that AI is an extremely powerful toolkit that will positively impact all kinds of socio-economic activity. But we’re not so sure about the true value of AI in the context of digital marketing, and specifically for international paid media.
Back to basics
Cutting through the hype, let’s start by looking at exactly how AI and machine learning work in the context of international paid media. For example, on a keyword level, how much and what kind of data are needed for AI to make a good decision?
Well, Google’s machine learning product Smart Bidding states that it “enables you to tailor bids based on each user’s context. Smart Bidding includes important signals like device, location and remarketing lists for better automation and performance”.
This implies that the signals required by the algorithm can be culled from the sum of users’ behavior, and that its “learning capabilities quickly maximize the accuracy of your bidding models to improve how you optimize the long-tail [by evaluating] patterns in your campaign structure, landing pages, ad text, product information, keyword phrases and many more to identify more relevant similarities across bidding items to effectively borrow learnings between them”.
This suggests that the ‘go to’ source of data is our own campaign. But what are these patterns, how long is ‘quickly’, and how on earth can landing page data would help with bid management?
Staying with bid management as an example, we think it works like this:
- Primary data: the algorithm looks back at historic direct interactions with a keyword within a client campaign, and makes a cost/position decision based on pre-defined goals like ROI or CTR, and of enough data.
- One way to address a possible data volume problem would be to look back a long way. But this would ignore seasonality, promotions and changes in consumer behaviors over time.
- Secondary data – the algorithm has insufficient data to make a ‘good’ decision on the primary basis, so uses corroborative data (performance indicators from other campaigns which have similar characteristics (e.g. same vertical, same language) to make decisions.
Do we even have enough data?
The question is if, aside from very high-volume big category campaigns (think car insurance, credit cards), there is enough primary data to power effective AI decision making. AI needs a huge amount of data to be effective. When IBM’s Deep Blue learned chess, for instance, the developer relied on 5 million data sets. Most industry experts believe that AI’s biggest limitation will be access to high-quality data of enough scale.
We also have no idea what a ‘good’ volume of data looks like. This is even more unlikely for international PPC, where campaigns are often very granular, multi-language, and designed to include lots of long tail keywords (which by definition do not have much volume).
When it comes to secondary data, how relevant can the corroborative data be? For maximum relevance, taking CLIENT X as an example, we’d have to assume that the algorithm is quickly assimilating data from CLIENT X’s direct competitors and using that to better inform the bid management strategy.
Surely that kind of cross-fertilized data would power all auction players’ bid tactics, creating a loop where no player has an advantage?
If competitor data is not used, then what kind of secondary data is sufficiently relevant to power good AI decisions. This would easier if we knew definitively how the rules of the algorithms were constructed, but of course, we never will.
Time for a reality check
To recap, if we knew that 10, 100 or even 1,000 interactions were enough to deliver superior efficiency via AI, we’d be delighted. Campaigns could be planned and executed to use the optimum blend of AI and human capabilities, with best results for ad platforms, agencies and clients. AI could focus on brand and category level interactions, with human oversight and detailed management of long tail.
It seems unlikely that adequate transparency as to how AI actually works, how much data is needed, how the ‘rules’ work, will be forthcoming unless significant changes in business models or practices occur.
Instead, AI is optimistically overhyped as digital’s next big thing while blithely ignoring the basic premise of AI and the current practicalities of both domestic and international digital paid media