Aiso framework · Demand → Retrieval → Evidence → Selection
The Aiso Recommendation Chain
First published .
Demand
Is the buyer actually asking a question where your brand could be chosen?
Retrieval
Can the assistant or its search layer find the right facts about you?
Evidence
Do the pages it finds support the claim strongly enough to ground an answer?
Selection
Given the user's constraints, does the model have a reason to pick you over the alternatives?
Why the order matters
Teams often jump straight to Selection: “How do we make ChatGPT recommend us?” But if the demand is not there, there is no recommendation to win. If retrieval is broken, persuasion copy cannot be seen. If the evidence is weak, the model has nothing safe to ground. Only after those stages work does the final competitive choice become the main problem.
A concrete example
Imagine a traveler asks for a quiet two-bedroom apartment near a specific neighborhood with air conditioning and a balcony. A property manager can lose at four different points.
- Demand: the team tracks “vacation rental” but never measures that constrained trip-planning prompt.
- Retrieval: the balcony or neighborhood fact only lives inside a booking widget the search layer cannot reliably parse.
- Evidence: the direct site says “quiet,” while OTA reviews repeatedly complain about street noise.
- Selection: all facts are understood, but another property is cheaper or closer and therefore wins the user's actual constraint set.
Those require four different fixes. Calling all of them “GEO” hides the useful part.
The metric for each stage is different
Demand needs prompt/session evidence. Retrieval needs search paths, bot access and fan-outs. Evidence needs citations and source consistency. Selection needs repeated recommendation outcomes under controlled buyer constraints.
This also explains why a single visibility score is hard to act on. It collapses four failure modes into one number.
How I use the chain in an audit
Take one commercially important prompt where a competitor wins. Work left to right. Do we see enough real demand to care? What did the assistant retrieve or search? Which sources support the competitor? Which criterion in the final answer actually separated them from us?
Stop at the first stage with a clear failure. That is usually the highest-leverage next action.
