LLM Perception Drift Audit

When this service becomes relevant

This service becomes relevant when an organisation no longer asks only whether it appears in AI answers, but how it is reconstructed. A brand can be cited, visible or mentioned while being described with the wrong framing, placed in a broad category, compared with the wrong competitors or reduced to an outdated version of itself.

That is the problem Pagup calls AI perception drift. The emerging market term is often LLM perception drift. On Pagup, this phenomenon is treated as a digital readability problem: the organisation exists, but the systems reading it do not stabilise the right representation.

This service is especially useful after a redesign, repositioning, brand merger, new offer, fast-growing corpus or period in which public content evolved faster than its governance.

What the audit actually measures

The audit is not designed to produce a decorative score. It compares generated readings against a reference baseline. Without a reference, there is no observable drift. There are only impressions.

We observe, among other things:

• how AI systems describe the organisation in one sentence, one paragraph and comparative contexts;
• the categories they place it in;
• the competitors or semantic neighbours they associate with it;
• the differentiators that survive or disappear in generated answers;
• outdated formulations that persist despite the current offer;
• the organisation’s recommendability across several query types;
• consistency between generated answers and the site’s canonical surfaces.

The work connects external observation to internal structure: source pages, proof, service hierarchy, public vocabulary, entities, links, structured data and machine-readable surfaces.

Interpretive variability dimension

The audit separates AI perception drift, which requires a qualified gap against a baseline, from interpretive variability, which first observes the dispersion of answers across systems, formulations, sources, moments and test contexts.

This distinction matters: an answer can vary without being false, but repeated variation can reveal an unstable category, weak proof activation, fragile recommendability or excessive dependency on certain sources.

In a Pagup engagement, this dimension helps identify:

• gaps against the organization’s canon;
• variation across AI systems;
• variation across query formulations;
• observable temporal variation;
• recommendability variation;
• variation in cited or activated sources.

Pagup does not infer global AI performance from one isolated answer. An actionable conclusion requires a baseline, preserved outputs, documented contexts and qualified gaps.

The audit now adds a further guardrail: distinguish the model’s native reconstruction from the delivered reconstruction served by an application. A repeated answer may reflect real stability, but it may also come from cache, routing, an approved answer store, or non-visible orchestration. This boundary is treated as delivery-layer fixation, not as automatic proof that a named provider uses caching.

Why this is not just an AI citation audit

An AI citation audit mainly answers the question: “Are we mentioned?” That question is useful, but too narrow. An organisation can be mentioned for the wrong reasons. It can be cited in a secondary category. It can appear in an answer without the proof that supports its authority. It can be named, then described in a way that weakens conversion.

An LLM perception drift audit answers a more strategic question: do AI systems still understand the organisation as it needs to be understood today?

The distinction matters. Citation measures presence. Drift measures a change in representation.

Method

The audit begins with a canonical formulation of the organisation: what it is, what it does, who it serves, which proof supports it, what its limits are and what differentiates it. Pagup does not invent that formulation. It must be extracted from legitimate sources or rebuilt with leadership when the existing sources contradict each other.

We then test several query families:

1. identity queries: “What does this company do?”;
2. category queries: “Which providers or experts exist in this field?”;
3. comparative queries: “Compare this organisation with other options”;
4. recommendation queries: “Who would you recommend for this problem?”;
5. temporal queries: “What is the current version of the offer?”;
6. proof queries: “What supports this expertise?”

The reading is compared with the AI perception baseline. Gaps are then classified by severity: lack of precision, dilution, confusion, remanence, category drift, recommendability drift or deeper contradiction.

Typical deliverables

Depending on the context, the audit may produce:

• an AI perception baseline, usable as a point zero for future comparison;
• a representation gap map, between the canon and observed responses;
• a cross-model drift grid, distinguishing isolated incidents from transversal patterns;
• a category and semantic neighbourhood analysis, to see whether the organisation is placed in the right market;
• a recommendability reading, showing when and why the organisation is proposed or excluded;
• a priority correction list, tied to pages, proof, content and signals to reinforce;
• a stabilisation trajectory, indicating whether the issue calls for targeted correction, brand disambiguation or broader architecture.

What the audit helps decide

A strong audit should reduce decision ambiguity. It helps determine whether the problem comes from content, brand, structure, proof, governance or a combination of these layers.

It may show that simply adding content will not be enough. It may also show that a full rebuild would be excessive if the main cause sits in a few poorly prioritised surfaces. In some cases, it reveals that the brand is visible enough, but its representation is too thin to support a buying decision.

That is where the audit creates value: it prevents AI visibility from being confused with AI understanding.

Conceptual reference

The doctrinal definition of LLM perception drift is available on gautierdorval.com. Pagup treats the phenomenon operationally: diagnosis, correction, stabilisation and monitoring across an organisation’s digital assets.