Live vs predictive

Lokrix produces a visibility score two ways: by asking the real engines (live) or by asking a calibrated model that predicts citation probability from page, brand and web features (predictive). Both feed the same score contract; they differ in cost, latency and the kind of confidence they give you.

Live runs

A live run queries the real answer engines with browsing/grounding enabled, captures the answer and citation cards, and detects mentions of your brand (URL match, brand NER, semantic match). It is the ground truth — what an engine actually said today. Because answers are non-deterministic, every prompt is sampled K times. Live runs consume credits — roughly one credit per engine response — and their cost is multiplicative.

Predictive runs

The predictive model estimates citation probability without paying for a live call. It is a calibrated model (calibrated with Platt/Isotonic and evaluated with Brier score, ECE and AUC) that scores cheaply and near-instantly, and also powers cheap what-if simulation for the optimization engine. It never consumes credits.

When each is used

Lokrix defaults to the predictive model and escalates to live runs deliberately — for example to establish a baseline, to verify a predicted uplift, or on a scheduled cadence for the prompts that matter most. This keeps ongoing measurement affordable while reserving expensive live calls for the moments they add the most certainty. Every result is labelled with its provenance — live, cached or mock (sample) — see Data honesty.