How AI Chooses Who to Recommend
In Wave 1, the same recommendation question did not have one stable machine answer. The study tested 72 web-grounded API observations across four model families, six prompt contexts and three repeated runs.
The subject is still the machine.
This is not an agent ranking. Study 001 measures how recommendation systems disagree, change their own answers, choose visible sources and react to small changes in context.
Across the six prompt families, mean cross-model recommendation overlap was only 0.0665 Jaccard.
The same exact prompt often produced a materially different shortlist on the next run.
108 citation domains appeared, with portal and directory sources making up the largest coded class.
Buyer, seller, luxury, relocation and neighborhood wording changed the recommendation pool again.
Now test the product people actually use.
Wave 1 used frozen API models. The next replication should run the original consumer-interface protocol literally, including the separate evidence follow-up, then compare product behavior with the API results.