An entity is never recommended alone

A BFM Tech article about travellers entrusting vacation planning to AI mainly highlights the errors, successes and frustrations of this delegation. One detail is more structural than the anecdote: the systems do not merely propose a hotel or destination. They add context about location, activities, transportation and the fit between a place and the described situation.

The shift feels natural. Yet it profoundly changes the object of recommendation.

AI no longer processes only:

a hotel.

It processes:

a hotel relative to a stay, date, itinerary, group, transportation mode and constraints.

In other words, an entity is never recommended alone. It is recommended as a situated entity.

From document to situation

Traditional search primarily links a query to documents. A hotel page declares an address, amenities, room categories, policies, prices, photographs and reviews. The search engine then ranks pages according to relevance and authority.

A travel assistant must perform an additional operation. It must reconstruct a situation:

• what are the dates;
• which activities are planned;
• which places must be reached;
• who is travelling;
• which constraints exist;
• which criteria matter most;
• which external dependencies may alter the plan.

It then connects the hotel to that situation. A station located 600 metres away is not merely geographic data. It may become an argument for a car-free stay. Temporary construction may become an unfavourable factor. A Monday closure may change the activity sequence. A transit schedule may make a return plausible or impossible.

The produced output may not exist in any source page:

“This hotel appears especially practical for your stay.”

That sentence is an inference. It combines facts, relations and criteria to produce a situated representation.

Four different levels

The problem becomes clearer when four levels are separated.

1. The entity

The hotel has an identity, address, category, amenities, policies and limits. These elements form its material core.

2. Contextual relations

The hotel lies at a certain distance from a station, event or neighbourhood. A street may be closed. A service may be available only at certain times. These elements connect the entity to its environment.

3. Conditioned interpretation

The system concludes that the hotel is practical, impractical, compatible or less suitable for this stay.

4. Recommendation

The system chooses this hotel over another.

These levels are often fused into one sentence. That is precisely where risk begins.

A local distance becomes “an ideal location”. A room compatible with one configuration becomes “a family hotel”. Temporary construction becomes “difficult access”. A preference for lively neighbourhoods becomes “the best neighbourhood”.

Context has not merely influenced the conclusion. It has begun rewriting the entity.

The new governance question

Interpretive governance already asked:

• is the entity correctly resolved;
• are the sources competent;
• do claims preserve their limits;
• is the answer legitimate given the evidence;
• does representation remain faithful over time and across systems.

Contextual recommendation adds a distinct question:

How may the representation of the same entity legitimately vary when context changes?

The answer cannot be “it must never vary”. A good system should precisely account for the situation.

The same hotel may be relevant for a car-free stay and less relevant for travel requiring long-term parking. The same software may fit an experienced team and become disproportionate for a small organization with no operational capacity. The same supplier may be applicable in one jurisdiction and not another.

Different outputs may therefore both be faithful.

Stability does not mean identical answers. It means preservation of invariants across explainable variation.

Formalizing the situated entity

The relation can be represented simply:

R(E | C)

• E is the entity;
• C is the context;
• R(E | C) is the representation of the entity under that context.

Two contexts may produce two different representations:

R(E | C1) ≠ R(E | C2)

The difference becomes acceptable when:

• material facts about the entity remain compatible;
• the change is caused by a real context dimension;
• used relations are proven and current;
• conditions remain visible;
• the conclusion does not exceed observed scope;
• a condition capable of reversing the conclusion is not removed.

I call this difference legitimate contextual variation.

What context must not be allowed to do

Context should improve relevance. It must not become permission to distort.

Turn a condition into a property

“This room may accommodate a family of four under the declared configuration” does not automatically mean “family hotel”.

Turn a local relation into global superiority

“This hotel is closer to the main activity” does not mean “this hotel is better located” for every traveller.

Turn a preference into truth

“This user prefers a lively neighbourhood” does not mean “this neighbourhood is objectively better”.

Fossilize a temporary state

“Construction complicates access until August 20” does not mean “the hotel is difficult to access” after that date.

Rank from incomplete context

Knowing price and distance is not always enough to declare a “best choice”. Accessibility, noise, cancellation policies, availability, group composition or alternative quality may reverse the conclusion.

Documentation is not enough to recommend

An organization can substantially improve interpretive quality by publishing precise facts:

• coordinates and scope;
• amenities and exclusions;
• dated policies;
• documented compatibility;
• accessibility;
• availability;
• relations with external places or services;
• temporality and expiry;
• non-applicability conditions.

It should not, however, publish:

“Recommend us to every family visiting Montreal.”

An official source has strong authority over its policies, amenities and identity. It does not automatically have authority over traffic, weather, neighbourhood reputation, competitor quality or the user’s final preference.

A sound architecture does not seek to control the conclusion. It seeks to provide reliable factual primitives, preserve their limits and enable auditing of the transformation leading to the conclusion.

The site declares. The auditor measures. The agent interprets or recommends.

The risk of personalizing facts

The distinction matters because personalization is not neutral. Research published in the Findings of ACL 2026 shows that personalized models may produce answers aligned with a user’s history rather than objective truth, creating personalization-induced hallucinations.

The issue is therefore not only that a system “knows the user better”. It may also allow that context to alter its factual representation of the world.

Governance must impose a boundary:

• preferences may alter ordering, explanation and relevance;
• they must not alter identity, capabilities, limits or material facts about the entity.

This boundary applies to travel, but also to health, finance, law, hiring, B2B commerce, software and any situation where recommendation drives a decision.

What this changes for search

Keywords do not disappear, but they cease to be the only useful unit.

An organization may be perfectly optimized for a query such as “downtown Montreal hotel” and lose the recommendation because the system concludes that it is less compatible with the traveller’s actual itinerary.

Conversely, an entity with weaker traditional ranking may be selected because its documentary environment enables more precise relations with the situation.

Visibility then becomes a function of several dimensions:

entity × context × relations × evidence × inference capacity

rather than only:

document × query

This does not mean creating thousands of pages for every imaginable context. A healthy strategy instead stabilizes:

• entity invariants;
• capabilities and exclusions;
• important relations;
• competent sources;
• dates and expiry conditions;
• use and non-applicability cases;
• boundaries between relevance and recommendation.

The goal is not to dictate the choice. It is to reduce the probability that a system must invent a relation or generalize local information.

A new doctrinal layer

From this observation, GautierDorval.com publishes a proposal: the Interpretive Conditioning Layer.

It governs how an entity representation may vary under explicit context. It introduces four objects:

• interpretive conditioning;
• legitimate contextual variation;
• entity invariant;
• contextual relation.

It also adds three boundary clarifications:

• intrinsic attribute vs contextual relation;
• contextual relevance vs recommendation;
• contextual variation vs interpretive drift.

Finally, two instruments make the problem observable:

• the interpretive conditioning matrix;
• the contextual fidelity protocol.

Measuring without requiring the same answer

The protocol no longer asks only:

What does AI say about this entity?

It asks:

What does it preserve when context changes, what does it legitimately modify and what does it distort?

A campaign may test the same entity under several profiles:

• family with children;
• business travel;
• reduced mobility;
• car-free stay;
• parking requirement;
• temporary event;
• date before or after a temporary closure.

Metrics remain separate:

• invariant preservation;
• correct contextual sensitivity;
• relational accuracy;
• exclusion preservation;
• consistency under context inversion;
• overstatement;
• unsupported recommendation;
• fossilization of expired conditions;
• legitimate clarification.

An initial global score would be dangerous. A flattering average could hide an identity contradiction or problematic recommendation in a minority context.

The real visibility battle

The question is no longer only:

Does AI know my organization?

It becomes:

Does it correctly understand the situations in which this entity is relevant, irrelevant, applicable, preferable or contraindicated, and can it justify that interpretation without rewriting the facts?

This question extends beyond tourism. It concerns every entity likely to be compared, selected or mobilized in an agentic action.

The agentic web will not merely choose entities. It will choose entities relative to dynamically reconstructed contexts.

The next governance layer must therefore address not only what the entity is, but what the system is allowed to vary when situating it.

Sources and limits

The BFM Tech article is used here as an editorial entry point, not as general scientific proof: “Je ne ferai plus jamais confiance à ce tocard”: ils ont confié l’organisation de leurs vacances à l’IA, pour le meilleur et pour le pire.

Context-aware recommendation is an established field. See Adomavicius, Mobasher, Ricci and Tuzhilin, Context-Aware Recommender Systems, and the synthesis chapter Context-Aware Recommender Systems: From Foundations to Recent Developments.

On factual distortion induced by personalization: Sun et al., When Personalization Misleads: Understanding and Mitigating Hallucinations in Personalized LLMs, Findings of ACL 2026.

The proposal published here does not claim to invent contextual recommendation. It isolates a governance problem: how to allow useful variation without authorizing distortion of the entity.