The context window is a language model's working memory for a single request. It holds the system prompt, the conversation so far, any documents or search results passed in, and the answer the model is writing. Anything outside the window is invisible to the model for that request. Window sizes are measured in tokens and have grown a lot in recent model generations, but they are still finite and still shared between everything in the request.
What fills the window
In an AI search product, the window is shared by several things: instructions from the provider, the user's question and earlier turns, the retrieved passages, and the reply. The provider decides how much space retrieval gets, and that decision is not usually published. A long window does not mean a whole website gets read; it means more room for whatever the retrieval step selected.
- System prompt and tool instructions
- Conversation history
- Retrieved passages or uploaded files
- The model's own output
Bigger is not the same as better
Research and vendor guidance both note that models can use information unevenly across a long context: details buried in the middle of a large input are sometimes used less reliably than details near the start or end. Longer inputs also cost more and take longer to process. For these reasons, AI search systems typically retrieve a focused set of passages rather than stuffing in everything that might be relevant.
What it means for your content
Since only selected passages make it into the window, the unit that competes for space is the passage, not the page. A section that answers one question completely, with its key fact near the top, is easier to select and easier for the model to use once it is there.
Example
A pricing page that gets cut off
Graphwick, a made-up analytics product, publishes one very long pricing page with a feature matrix, legal notes and a long FAQ. When asked about Graphwick's free plan, an assistant cites an old review instead. One plausible reason: the free plan details sit deep in the page, after thousands of tokens of table markup. Moving a short plain-text summary of each plan to the top gives retrieval a compact passage to pick up.
Common mistakes
Assuming a large window means full-page reading
AI search products choose what to include. Your page competes passage by passage for a slice of the window.
Burying key facts
Put plan names, prices, limits and definitions where they are easy to lift out, not at the end of long sections.
Questions
Only what the application puts back into it. Memory features work by storing information and reinserting it into the context of later requests.
It varies by model and changes with each release. Check the vendor's model documentation for current limits rather than relying on a fixed number.
