"Deep learning advertising platform" is a phrase doing two very different jobs. One of them is true and mostly invisible. The other is a slide.
The true one
The ad platforms are deep learning companies that happen to sell inventory. Every impression you buy is priced by a neural network estimating the probability that this person takes the action you asked for, multiplied by your bid, ranked against everyone else in the auction. Meta has published its progression through retrieval and ranking model rewrites, Google runs the equivalent under Smart Bidding and Performance Max. Advantage+ and PMax are the same idea pushed further: hand the model targeting, placement, budget split and increasingly the creative combination, and let it search.
That system already sees things you never will. It has your conversion signal, the user's entire session history, cross-surface behaviour, and a few billion comparable auctions from the last hour. It updates continuously.
So when a vendor says their deep learning will out-predict your ad performance, the honest follow-up is: out-predict what, exactly. The model that sets your delivery is on the other side of the API, and it is bigger than yours, closer to the data, and retrained more often.
The slide one
The second job the phrase does is decoration. A tool tags creative with a vision model, embeds it, clusters it, and the deck says deep learning platform. The tagging is genuinely deep learning. The claim stacked on top of it usually is not.
We tested the strongest version of that claim on our own data and published the result: pooled, creative embeddings separated high from low CTR at 0.94 AUC. Demeaned by brand and objective, so the model had to explain variance within an account, it fell to 0.36. Below chance. The full write-up is in Similarity is not prediction.
That is not an argument against neural networks in ad tooling. It is an argument about which question you point them at.
What that leaves for you
If allocation is being solved by a model you do not control, your remaining leverage is the input space. The platform can only optimise across the creative you gave it. It cannot test an angle nobody made, a format nobody shot, or an objection nobody answered. Advantage+ is a search algorithm, and you are writing its search space.
Which turns the useful questions into things that are actually checkable today:
- What is in market right now, ours and theirs, and what is decaying.
- Which angles, formats, hooks and personas this account has never produced.
- Which competitor territory we have no coverage in at all.
- How fast we can get the missing thing shot, written, and live.
Every one of those is diagnosis and coverage, not forecasting. You can verify each of them against the data the same afternoon, which is not true of a predicted win rate.
Three questions for anyone selling you a deep learning ad platform
- Demeaned, or pooled? Any performance-prediction number that has not been controlled for brand and objective is mostly base rates in a costume.
- What is the model's actual output? "Tags this creative" and "ranks which creative to run" are different products with the same brochure.
- What breaks if it is wrong? A wrong similarity search wastes a scroll. A wrong prediction moves budget.
Where we sit
Marxx uses deep learning for the things it is reliably good at. Vision and language models to read every frame, hook, face, scene and claim in your account and your competitors'. Embeddings for semantic search across the ad library and for measuring how close your creative sits to territory someone else owns. Language models to turn all of that into a brief, and a production graph to get the brief shot and published to Meta without leaving the tool.
What we do not do is tell you which ad is going to win. The system that decides that is already running, it belongs to the platform, and the most valuable thing we can do is make sure it never runs out of good options.