Adtech's blackbox bulge: blind AI

Everyone’s doing some sort of ML these days, nothing special to it. Let’s not focus on that or that for now. That’s a topic for another day.

Being a media buyer myself I am always stuck with the plethora of choices provided by adtech companies to get us all hooked. Creative A/B testing on the go with a dedicated platform? (think Revjet) The TradeDesk with their KOA? Tools like Adsbridge or Binom for LP optimization and A/B testing? Going further, there are platforms utilizing onsite .js to actually test some of the UI elements and adjust the LP delivery based on predictions PER USER. Nothing new here, I know.

I keep getting surprised with all these agency / client people trying to do it all at once. No one is opting out thus we are ending up stacking adtech black boxes with their proprietary algorithms with no back and forth communications between those. I might be wrong and there are (there definitely should be more advanced ‘stacking’ methods to make it all work. I’d say custom buying agents merging all the signals into one blackbox able to make buying decisions) but so far I wasn’t presented with any solution to actually make it all work in conjunction.

Every product is using some sort of ML while trying to optimize its own little part of the chain ‘assuming’ it is getting the very same input (in terms of features) i.e. TradeDesk getting third-party served ads assuming the assets to be the same. It is trying hard to optimize every single one of them per channel to ensure the best format/version is getting its prime time.

But if you happen to play around with ads on the go, if your LPs are rotating or the actual UI is changing based on user-agent?

The stack is swallowing it all for obvious reasons, everyone’s trying to get a piece of that cake (Hi Voluum!). But for God’s sake - how do you track it all at once? What’s the chance the optimizations you are ending up with are ideal. They can’t be by definition.

What if one product’s model wasn’t good enough and there is no way to tell as no platform is making its metrics public (by public I mean one model’s stats per campaign). How do we assign the weights? What’s actually contributing and what makes our CPAs lower?

As there is no feasible solution I would recommend withholding from taking all pills at once, especially after you get back from Dmexco. Adding more boxes to the stack will make it all more unstable and more expensive in the long run.

Consider using fewer tools for better transparency and predictability, i.e. I would still prefer a hammer to 'nail it' especially when this comes.

Do think about it.

I really do.