Entity Drift: When Models Describe You Wrong
Entity drift — the divergence between how an organization defines itself and how models reconstruct it — is detected through prompt monitoring and corrected only through canonical consistency across every surface the engines read.
Models reconstruct entities from whatever evidence they retrieve; when that evidence is stale or contradictory, the reconstruction drifts from the canonical truth. Drift is detected in answers and corrected on surfaces.
What Entity Drift Is
A model does not store a company's About page. It stores a compressed reconstruction — a blended description assembled from every mention the training data and retrieval layers carry. Entity drift is the divergence between that reconstruction and the entity's canonical self-description: the model places the product in the wrong category, credits it with capabilities it does not have, omits the ones that matter, or keeps repeating a positioning the company retired.
Drift is mechanical, not malicious. The model reconstructs the entity from whatever evidence it retrieves, and when the evidence is stale, sparse, or contradictory, the reconstruction inherits the defect. The wrong description is a symptom; the evidence graph is the cause.
It also compounds through a feedback path. A wrong description, repeated often enough in answers, gets quoted by third parties and re-ingested as independent corroboration — the model's error becomes the web's record, and the web's record becomes the model's evidence. Left unmeasured, drift is self-sealing: each repetition makes the next repetition more defensible.
Where Drift Comes From
Four sources account for most of it.
- Stale third-party surfaces. Old profiles, archived articles, and unmaintained directories outnumber or outrank the current first-party record. The model corroborates the past because the past is what the retrieval set contains.
- Self-contradiction. Pages that describe the entity differently — an old tagline on one route, a new one on another, inconsistent status labels — hand the model conflicting evidence, and conflicting evidence resolves to vague synthesis.
- Evidence sparsity. Where little can be retrieved, the model fills gaps with inference from category neighbors. Inference is where wrong capabilities and wrong categories come from.
- Change outrunning the record. Renames, launches, and repositioning alter the canonical truth instantly; the public evidence graph updates slowly. Drift is the lag made visible.
None of these require an adversary. They accrue from ordinary publishing — pages added, pages forgotten, third parties repeating what was true last year.
Detection: Watching the Answers
Drift is invisible to conventional monitoring because it lives in answers, not rankings. A fixed prompt set — the questions where the entity should appear, run repeatedly across engines — makes it measurable. Each run captures how the entity is described: its category, its capabilities, its status, and which sources get cited for those claims.
The detection work is classification. Errors sort into a small taxonomy: wrong category, wrong capability, wrong status, conflated with another entity, or absent where presence was expected. Classification turns "the model gets us wrong" into a specific defect list, and each defect maps back to the evidence that produced it — the stale source, the contradictory page, the missing surface.
This is the measurement layer Geolify's Prompt Monitor provides between audits: a defined prompt set tracked over time, with the brand mentions inside each response recorded, so changes in how the engines describe the entity surface as data rather than anecdote.
Correction: Canonical Consistency
The model cannot be edited; the evidence it reads can be. Correction means making every controlled surface state the same entity facts in the same vocabulary: the site's pages, its structured data, its declaration files, the third-party profiles within the owner's control. Consistency does the work volume used to — corroborating surfaces raise the model's confidence in the canonical description, and each corrected surface removes one piece of evidence for the drifted one.
What correction cannot do is argue. Publishing a rebuttal the model never retrieves changes nothing; the unit of correction is the surface, not the argument — another page that states the right facts in citable form, with structured data to match, is worth more than any post asserting the record is wrong.
Correction is gradual by mechanism, not by preference. Engines re-derive descriptions as the retrieval record changes; a fixed page does not rewrite an answer overnight. The operating loop is monitor, correct the surfaces, then monitor again — watching the reconstructed description converge toward the canonical one run by run.
A Standing Discipline
Drift is not a defect that gets fixed once. Every product change, every new surface, every stale page left standing re-opens the gap between self-description and reconstruction. The entities that hold their description are the ones treating canonical consistency as maintenance — a weekly loop of auditing the gaps, fixing what matters, and tracking whether the answers moved. Drift never stops accruing. The only question is whether anyone is measuring it.