When you ask a model a question with web search enabled, you get one answer. When you ask the same question with web search disabled, you get a different answer. These are not two versions of the same thing — they are two fundamentally different narratives, shaped by different mechanisms, carrying different risks. Deep Narrative Intelligence measures both, because a complete picture of your AI narrative requires understanding both.
The two interrogation modes
Modern AI assistants offer two modes of operation. In no-search mode, the model answers purely from its training data — the static representation baked into its weights at training time. In web-search mode, the model augments its training data with real-time retrieval from the web, pulling in current articles, pages, and sources before generating its answer. Most users do not consciously choose between these modes; the model may default to one or the other depending on the query, the platform, or the user's settings. But from a DNI perspective, the two modes produce narratives that can diverge significantly, and both must be interrogated.
Latent narrative vs. live retrieval
The no-search answer reveals the latent narrative: what the model believes based on its training data alone. This is the narrative that persists across sessions, that is consistent for every user who asks the same question without search, and that changes only when the model is retrained. It is the deep narrative — the one most resistant to correction, because it lives in the weights, not in a retrievable source. The web-search answer reveals the live retrieval narrative: what the model assembles from current web sources at the moment of the query. This narrative is more volatile — it can change from day to day as web content changes — but it is also more correctable, because it responds to changes in the source layer in real time.
Don't know what AI says about you today?
A SetRecord audit queries supported AI models about your organisation and measures the gap with verifiable facts — within five business days.
What each mode reveals
Each mode exposes a different risk. The latent narrative is the risk of a deep, persistent distortion — a falsehood or stale fact that has been baked into the model and will not change until the next training cycle. If the model believes your CEO resigned last year (when they did not), that belief will surface in every no-search answer until the model is retrained, regardless of what you publish today. The live retrieval narrative is the risk of a volatile, source-driven distortion — a falsehood that a recent web source has introduced and that the model amplifies because it trusts the source. If a low-quality aggregator publishes a false claim about your company today, the web-search mode may surface it tomorrow. These are two different risks requiring two different correction strategies: the latent narrative requires training-cycle-level interventions, while the live retrieval narrative requires source-level interventions.
Why you need both
An organisation that measures only the no-search narrative understands its deep exposure but misses the live distortions that web-search users encounter. An organisation that measures only the web-search narrative understands its current surface exposure but misses the persistent distortions that will outlast any individual web source. You need both because your stakeholders use both modes — some deliberately, most unknowingly — and the narrative they receive depends on which mode the model used. A DNI programme that measures only one mode is measuring half the picture. SetRecord runs dual-interrogation panels as a standard practice, because the gap between the two modes is itself a diagnostic signal: where the latent and live narratives agree, the distortion is deep and stable; where they diverge, the distortion is recent and potentially correctable through source-level intervention.
How DNI structures the dual interrogation
The dual interrogation is methodologically straightforward but operationally precise. Each query in the interrogation panel is run twice: once with web search enabled, once with it disabled. Both answers are recorded, compared, and scored against the verifiable record. The comparison produces three outputs: the latent narrative score (no-search), the live retrieval score (web-search), and the divergence metric (the gap between them). The divergence metric is particularly valuable — it identifies topics where the model's training-data narrative and its retrieval narrative are in tension, which is where correction efforts are most likely to produce movement. A high divergence on a critical topic is a flag that the source layer and the training layer need different interventions, and that the correction strategy must be two-track to be effective.