SHAP values (Shapley Additive Explanations) decompose a surrogate model's prediction into per-feature contributions, so you can explain why the model expects a given page edit to lift mention rate by 6 points instead of 2. Without SHAP, the predicted lift is an oracle output the team is asked to trust; with SHAP, the prediction comes with a receipt showing which features drove it.
What SHAP actually computes
For each feature in the prediction, SHAP computes how much that feature contributed to moving the prediction away from the model's baseline. Positive SHAP values pushed the prediction up, negative ones pushed it down. The sum of all SHAP values equals the difference between the model's baseline prediction and the actual prediction for this input. The decomposition is mathematically guaranteed to be additive and consistent — no other feature attribution method has those properties together.
Why explainability matters in production AEO
Two reasons. Trust: the team has to commit content investment based on the prediction, and 'because the model said so' fails the first stakeholder review. Detection: when a prediction looks wrong, SHAP localizes the feature responsible — usually surfacing either a feature-engineering bug or a category-specific quirk the team needs to handle. Programs that ship predictions without SHAP eventually erode trust to the point where the surrogate stops being consulted.
How to read a SHAP chart correctly
Three patterns to internalize. A long positive bar on entity density means the model attributes most of the predicted lift to the entities you added. A long negative bar on freshness suggests the page's age is dragging the prediction down — the simulator is telling you the edit alone may not be enough if the page is also stale. Bars near zero mean that feature is not load-bearing for this prediction; do not over-invest in it for this page.
Common SHAP misreadings
The most common mistake is reading SHAP as causal. SHAP says 'the model thinks this feature explains its output' — which is not the same as 'changing this feature in the world will change the outcome.' Causal claims still require the verify step. The second mistake is comparing SHAP values across different brands or different prediction horizons; the values are local to one input and one model, and cross-input comparisons need careful normalization.
How AskRanker uses SHAP day-to-day
Every Execute playbook recommendation ships with the SHAP decomposition behind the predicted lift. The team can see, for each suggested edit, which content features the model thinks will drive the lift, and how much each contributes. After the verify step, the residual gets attributed back to the SHAP features so the team learns where the model's explanations were faithful and where they were not. Over a quarter, the brand's SHAP-features-vs-reality map becomes a real artifact, separate from the model itself.