Auto-build every mix of your ad pieces and keep feeding them to Facebook

Eric Seufert… this. I mean, yes, you can go from like 10 creatives to 200, but a lot of times it's just people taking their 10 creatives and those themselves are variants, and they're getting 200 variants of one concept, right? That's not actually doing anything for you. There's just diminishing returns on taking 20 variants of one concept and going to 200, and and they're they're like vanishingly small gains, right? What you really care about is the concept.

Eric SeufertAnd actually, it's coming up with concepts that you yourself couldn't come up with, because if you could just come up with them, the AI is not doing that much more to sort of influence the performance. Sort of like aha moment with this was the last company I worked at, I built this tool called Draper, and that's just what it did. It just it just created variants of of ads. And this is like 2018, so people weren't really talking about like AI at that point. This wasn't even like machine learning, it was just created a bunch of variants, like all permutations of these different ads. What I would do is we would just deploy these on Facebook

Eric Seufertall the time. Like there's this constant cycle of this deployment, and then we had like a stand-up every every week with the whole company, and I would say, "Here's the ad that worked the best this week. I have no idea what it's going to be."

David BarnardLaughs.

David BarnardYeah.

Eric SeufertRight? There's no point. What you should be interpreting is if when you get a win or the win rate increases, the process worked. Now, maybe the process took a new input, and that's the learning, right? But it's not the output, because that was random. That was utterly random. Why that worked was utterly random. If you try to sort of deconstruct it and take a learning from that, you're just wasting your time. What you should do is, "Okay, how am I sort of changing the inputs such that I'm getting a higher win rate?" and let the machine do its thing. That output is irrelevant. That output is is cannot be interpreted by you. You can't understand it.

Eric SeufertYou can't understand why that worked. Don't even try. If it worked though, what did you change about the inputs? And that's what you learned, right? The process worked, not the ad.