LLM perception drift and AI perception drift
LLM perception drift names a problem that goes beyond AI visibility. An entity may appear in a generative answer, be cited by an answer engine, be mentioned in a summary, and still be reconstructed in a way that progressively differs from its canon.
This page is the conceptual hub for the AI perception drift cluster on gautierdorval.com. It connects the emerging market vocabulary with concepts already stabilized in this corpus: interpretive drift, AI brand representation, canon-output gap, proof of fidelity, framing stability, drift detection, and interpretive risk.
Working definition
LLM perception drift is the observable change in how large language models describe, classify, compare, recommend, or prioritize an entity over time.
The broader expression AI perception drift includes not only textual LLMs, but also answer engines, assistants, RAG systems, generative summaries, agents, and recommendation surfaces that produce a representation usable by a human or another machine.
The issue is not simply whether a brand is visible. The issue is whether the produced representation remains faithful, current, bounded, and differentiating.
Why this cluster exists
Market vocabulary tends to reduce perception drift to a tracking metric: citation count, generative share of voice, sentiment, average answer rank, or recommendation frequency.
Those measures are useful, but insufficient. They observe the effect. They do not always explain why the representation changes, which source dominates, which category absorbs the entity, which evidence is missing, which competitor contaminates the semantic neighborhood, or which older version continues to structure the answer.
The role of this cluster is to state a stronger thesis:
AI visibility is no longer enough. What matters now is AI perception stability.
That stability requires a legible canon, bounded definitions, source hierarchy, accessible evidence, exclusions, disambiguation links, and observation of gaps between what is declared and what is reconstructed.
Main forms of drift
AI perception drift is not a single phenomenon. It can take several forms.
Category drift
The entity is placed in the wrong market, an overly broad category, or a competitive neighborhood that does not match its actual role. A digital readability firm may be reduced to an SEO agency. A doctrine of interpretive governance may be reduced to a simple set of bot-facing files.
Representation drift
The portrait produced by systems changes: services, audiences, differentiators, limits, evidence, author, perimeter, or status. The representation becomes plausible, but less exact.
Recommendability drift
The brand or concept is not necessarily absent. It becomes less spontaneously recommended, recommended for the wrong reasons, or positioned behind actors whose role is different.
Cross-model drift
Several systems stabilize incompatible versions of the same entity. ChatGPT, Gemini, Perplexity, Claude, Copilot, or an internal RAG system may not converge on the same portrait.
Temporal drift
An older version of the entity continues to dominate. The system reconstructs what was true before a redesign, repositioning, doctrine publication, or correction.
Difference from hallucination
A hallucination is often understood as a visible error. AI perception drift is subtler. It may contain true elements, but organize them through a frame that changes the overall perception.
For example, an answer may say true things about a person while placing them in a professional category that is too narrow. It may cite a real offer while erasing the doctrine that makes it different. It may recommend a brand while associating it with a commercial intent that is not its main axis.
The risk comes from plausibility. A weakly false representation can become more durable than an obvious hallucination.
Governance condition: the canon
Perception drift cannot be measured without a reference state. The canon may take several forms: definitions, doctrine pages, service pages, evidence, entity graphs, machine-first corpus, policies, version history, or relationship maps.
The canon does not merely say “this is the truth”. It makes the distance between declared source and reconstructed output measurable. That is the role of the canon-output gap.
When a gap is observed only once, it may be an incident. When it repeats, propagates, worsens, or stabilizes across several systems, it becomes a drift phenomenon.
Recommended reading path
Start with the definitions:
- AI perception drift
- LLM perception drift
- AI perception stability
- AI perception baseline
- Cross-model drift
- Category drift
- Recommendability drift
- AI representation drift
- LLM perception drift audit
Then read the clarifications:
- AI perception drift vs hallucination
- LLM perception drift vs AI visibility
- AI perception drift vs interpretive drift
Then the methods:
Relation to interpretive risk
Perception drift becomes an interpretive risk when the produced representation is used to decide, compare, recommend, exclude, buy, delegate, or automate an action.
The issue is therefore not only editorial. It becomes operational. A wrong category can influence a vendor list. An older version of a company can influence a due diligence summary. An overly generic representation can reduce offer recommendability. An authority confusion can make a secondary source replace the canon.
That is why this cluster must remain connected to the Interpretive risk page.
Role of this site
gautierdorval.com treats this theme as a doctrinal object. The site does not only monitor AI citations. It names the conditions under which an entity remains interpretable, faithful, current, and governable in a response web.
The LLM perception drift cluster therefore acts as a bridge between an emerging market term and a broader architecture: canon, output, gap, drift, risk, correction, resorption.