In August 2026 we asked four AI engines about two properties in Anse Royale, Mahé: a large branded resort and a family-run guesthouse a few hundred metres away. The point was not to rank them. It was to see what an assistant believes before it answers a traveller.
The branded resort: present, and wrong
The resort appears in every recommendation query. The surface the engines assemble for it is also contradictory:
- a golf course that does not exist — an aggregator artefact, repeated;
- “steps from the beach”, where the real walk is ~250 m along a road without a pavement, or through a neighbouring café;
- a cuisine label (“Cajun Creole”) that no one at the property would use;
- a star rating that flips between 4 and 4.5 depending on which copy was read;
- a look-alike domain (
*.com-seychelles.com) presented as official.
None of this is malicious. It is what happens when many sources copy each other and a model weighs that agreement as evidence.
The guesthouse: absent
The guesthouse is not in the model’s memory and rarely in its retrieval. Where it does surface, the inventory is stale and inconsistent (12 vs 16 units), and the airport transfer time is given three different ways by three sources (15 / 20 / 25 min). The owner’s verdict, verbatim: “there is no correct information.”
In recommendation-style queries (“guesthouse near Anse Royale, two adults, four nights”) it did not appear at all. The resort appeared in every one.
What this tells us
Cross-source consistency is not accuracy. Aggregators copy aggregators; engines treat the resulting consensus as a signal. The only thing that breaks a copy chain is a statement from the party that actually knows — the owner — confirmed by someone independent — a guest, a licence registry, a destination authority.
That is the whole reason the claim-audit layer exists. A profile is not “verified” because eight sources agree. It is verified when the claim carries who stated it, when, and who checked it — and when that record can be replayed later.
Method notes
- Engines: four, queried with retrieval enabled; each property asked in a fixed set of factual and recommendation prompts.
- Ground truth: owner interview and on-site visit for the guesthouse; public licence record and property website for the resort.
- Numbers here are descriptive. The measurement protocol that turns this into metrics is documented in the IK1 protocol and is being run separately.