AI recommendation stability, 2026: how much an engine changes its answer week to week

Stated limits

• Churn is an UPPER bound: aliases the canonicalisation does not merge count as a change. Two unambiguous same-product spellings (ChatGPT Enterprise with OpenAI ChatGPT Enterprise, Dialogflow with Google Dialogflow) are merged for this series only; other look-alikes that may be different products are left apart. Of 166 top-5 departures across adjacent weeks, only 18 had a same-looking arrival, so most churn is real replacement and not spelling.
• One capture per question per week. Part of any week-to-week change is the engine's own run-to-run noise; the tracker's cited noise floor (a second run of the same week) is reported in the tracker snapshot, not re-measured here.
• Only engines captured in both weeks of a pair are compared; a skipped week is recorded as a gap and never bridged.
• Model-memory engines only for most weeks: these are what a model says from training, not live web results, unless the engine label says grounded.
• These are one estate's categories and one capture per question per week. It describes how this index behaves, not how every buyer's query behaves.

Conflict of interest. Lucreya is part of the Lattice network, which sells disclosed placements on pages that AI engines cite. That is a commercial interest in this subject. These files are published so the findings can be checked against us.