A surrogate model in AEO is a lightweight predictor — typically an XGBoost gradient-boosted tree — that maps a page's content features to a predicted mention rate per buyer question. It is what makes the simulate-before-publish step computationally feasible: a real LLM scan costs a few cents and takes seconds; the surrogate runs in milliseconds and lets the team test 50 hypothetical edits in the time a real scan takes to test one.
Why a surrogate at all
Without a surrogate, every 'what would happen if I added this paragraph' question requires an actual LLM scan: 50 samples per model, four models, real cents per call, and seconds of latency. That cost forecloses interactive simulation. A surrogate trained on past scan data approximates the LLM's behavior cheaply enough that the team can run hundreds of simulations in a sprint and pick the highest-predicted-lift edits to actually ship.
What features the surrogate uses
Two feature classes combine. Brand signals (mention rate history, position rate history, sentiment rate, share of voice trend, competitor signals on the same questions) describe where the brand stands today. Content features (entity density, definition-first paragraphs per page, schema coverage, internal link count, freshness signal, comparison-page presence) describe what the page actually contains. A typical AskRanker surrogate trains on 14 to 21 numeric features per page-question pair.
Why XGBoost specifically
XGBoost is the practical default for surrogates of this shape because it handles tabular features well, trains in under a minute on data sizes typical for a single brand's scan history, requires minimal hyperparameter tuning to be reasonable, and produces SHAP values for free, which feed the explainability of every prediction. Neural alternatives offer marginal accuracy gains at the cost of much higher MLOps overhead. Linear models lose meaningful nonlinearity. XGBoost is the sweet spot.
How accurate the surrogate has to be
The surrogate does not need to predict absolute mention rate — it needs to predict the lift from a proposed change. Mean absolute error on lift in the 3 to 5 percentage point range is the practical threshold for the predictions to be load-bearing. Below that, the simulator is sharper than the noise band of the actual scans. Above 5, predicted-vs-actual comparisons stop being interpretable. AskRanker brands typically reach sub-5 MAE within 30 days of daily retraining.
Daily retraining and drift
Categories evolve. New competitors launch, new comparison articles go viral, the LLMs themselves get updated. A surrogate trained once and frozen drifts out of accuracy in weeks. AskRanker retrains the surrogate every day on the running 90-day scan history at roughly $0.30 of compute per brand, which keeps drift bounded. The retrain log is part of the verify step: when MAE creeps up, the team gets a flag and can investigate whether something structural changed in the category corpus.