Retrieval
Retrieval is the step where an AI platform fetches web pages or passages to answer a question, on top of what its model learned in training.
Retrieval is the step where an AI platform goes and gets information to answer a question. Usually that means searching the web, reading the pages it finds, and pulling out the passages that seem to answer what was asked. The model then writes its answer from those passages and from what it learned in training.
For marketers, retrieval is the part you can influence this week. You can't change what a model learned last year. You can change the pages it reads today.
Retrieval vs training
A model learns from a huge amount of text before it's released. That knowledge is fixed until it's trained again, and it goes out of date. Retrieval fills the gap: the platform looks things up at the moment someone asks.
The catch is that the model only works with the pages it fetched. If your page isn't found, can't be read, or loses out to other pages, it plays no part in that answer, however good it is.
What happens during retrieval
The details differ between platforms, but the steps are similar. The platform turns the question into one or more searches. It gets back a list of pages. It reads them, picks the passages that match best, and gives those to the model.
For example, someone asks "Which UK coffee roasters do plastic-free bags?" The platform searches, gets 20 pages, and picks passages from five of them. Kestrel Roasters says "plastic-free" in a heading on its packaging page. Fernhill Coffee only mentions it in an image caption. Kestrel's passage gets picked.
Retrieval in AnswerOrbit
Whether a platform searches depends on the platform. ChatGPT and Perplexity always search when we ask, Gemini and Copilot decide for themselves, and AI Overviews and AI Mode are part of Google Search. How we count has the table. Every answer we keep comes with the pages it used, and the Searches tab in Sources shows how often each platform searched and how often you were named when it did.