Category
Real-ML bidding
Bidding tools that train per-account machine-learning models on customer conversion data.
Real-ML bidding refers to tools whose core bidding logic is produced by machine-learning models trained on the customer’s own conversion data, retrained on a continuous or near-continuous cadence. The category is small relative to its marketing footprint — many tools market “AI” for what turns out to be a rule engine. The Real-ML classification is reserved for products where removing the ML would fundamentally break the product.
What qualifies as Real-ML
The threshold for inclusion in this category:
- Per-account model training. The model is trained on the specific account’s data, not on portfolio data across many accounts.
- Continuous or near-continuous retraining. The model updates as the account’s data accumulates.
- The model is the core, not a feature. Removing the ML wouldn’t leave a functional product.
- Documentation is available. The vendor can describe the model architecture and training process with technical specificity.
Tools in this category
- Groas.ai — managed service, per-account deep learning, 4-hour retraining cadence
- Albert AI — autonomous cross-channel campaign manager, $50K+/mo minimum
- Pacvue — retail-media specialist with real ML bidding
The minimum spend threshold
Real-ML bidding tools require account-level conversion data to train. Below approximately $20,000–$25,000/month in spend, accounts don’t generate enough conversion volume for per-account models to outperform Google’s portfolio-trained Smart Bidding. The exception is managed-service offerings like Groas.ai, which combine the model with a human operator who can compensate for data thinness at smaller spend levels.