All posts
Research
8 min read
The anatomy of an AI answer
Before a model recommends a brand it retrieves, ranks and compresses. Here is what happens in each step, and where you can influence it.
Maya Okafor
Head of Research
Ask an assistant for a recommendation and it looks effortless: one confident paragraph, three names, a short reason for each. Underneath, a surprising amount of machinery decides which three names make it.
Retrieval comes first
Most assistants now search before they answer. The query they run is rarely the question you typed. It is rewritten, expanded and often split into several narrower searches. Being findable for those rewritten queries matters more than ranking for the original.
Sources are weighed, not counted
A single well-structured comparison on a trusted site can outweigh dozens of passing mentions. Models favour pages that state facts plainly, keep them current and agree with other sources.
The answer is a compression of what the sources agree on. If nobody agrees on what you do, you get compressed out.
What to do with this
Publish the plain facts: pricing, limits, integrations, who it is for.
Keep third-party listings as current as your own site.
Track the rewritten queries, not just the prompt.
None of this is exotic. It is the unglamorous work of being easy to describe correctly.
Keep reading.
All posts