Systematically querying a language model with structured prompts to map its latent biases, brand associations, and internal knowledge graphs without relying on traditional rank trackers.
Model probing is the practice of systematically querying a language model with structured prompts to reveal its internal biases, brand associations, and knowledge-graph relationships — without relying on traditional rank tracking or search-result observation. The model itself is the measurement instrument.
Two core prompt patterns drive the method. In the Brand-to-Entity direction (B→E), the model is asked what it associates with a named brand: "List ten things you associate with [Brand]." In the Entity-to-Brand direction (E→B), the query is reversed: "List ten brands you associate with [keyword or concept]." Running both variants repeatedly, across multiple models and over time, surfaces a brand's position in the model's latent knowledge graph.
Each response is normalised and ranked. The position at which a brand first appears, how often it appears across runs, and which competing brands appear alongside it together quantify what the model believes — before any grounding or live retrieval has occurred.
A model that does not associate a brand with a topic will not surface it in answers regardless of how well that brand ranks in traditional search. Probing makes that gap visible and measurable. It also reveals which competitor brands the model treats as default answers, and which associations are weak enough to shift through targeted content or citation work.
At DEJAN we built AI Rank around this method, probing multiple models daily to track how brand associations change over time — the AI equivalent of a rank tracker.
What probing reveals is largely primary bias: the ungrounded confidence a model formed during training. Because these beliefs precede any live retrieval, improving them requires influencing the training signal itself — through consistent co-citation, authoritative sourcing, and the kind of associative embeddedness that comes from being named alongside the right entities at scale.
