Context windows, passages and the limits of long documents

A context window bounds information a model can process in a request. Instructions, conversation history, tool results and supplied materials consume input capacity; output limits also depend on the model. The model’s published capacity does not establish how many pages a search product fetches or how much of each page is supplied.

Token counts also differ from word counts. Unusual names, punctuation and languages can tokenize differently. Use actual token-counting tools in your own application, not a universal page-length rule for GEO.

A source-informed answer can use excerpts or other selected material rather than a complete website. The surrounding navigation, introduction and distant footnotes may not be part of that material. This is a useful conceptual model for editing, not a claim that every assistant uses the same chunk size.

A passage should identify the subject, explain the relevant fact and retain important exceptions. Avoid pronouns with unclear antecedents and claims whose meaning changes when a distant qualification is removed. Preserve the full explanation on the page as well; self-contained passages are not an argument for fragmenting a guide into dozens of thin pages.

“Lost in the Middle” studies multi-document question answering and key-value retrieval. It reports sensitivity to where relevant information appears in the tested contexts, including degradation for information in the middle.

That is evidence about specified tasks and tested models, not a universal law that current assistants ignore the middle of every page. A page’s visual position is also not necessarily its position in model input. Do not convert the study into an unsupported instruction to move every keyword to the top or bottom of a website.

Illustrative scenario: an export guide begins “Reports can be exported,” while much later it says only administrators on a particular plan have that permission. An excerpt of the opening could support an overbroad answer.

A clearer opening states the feature, permission and plan requirement together, followed by the detailed procedure. This is an improvement in the completeness of the evidence. It does not prove that an assistant will retrieve the passage or preserve its restriction. Evaluate observed answers after editing rather than treating paragraph placement as a citation guarantee.

Inspect the exact text supplied to the model and whether it was truncated. Define passage boundaries that retain essential conditions. Retain document identifiers, update dates and source links so claims can be traced back to evidence. Test difficult cases with conflicting versions or prerequisites.

Measure retrieval and answer quality separately. A relevant passage appearing in context does not mean the answer used it correctly. A failure to mention a condition does not automatically mean the token budget was exhausted.

• Record prompt, mode, conversation conditions and date.
• Inspect citations and compare claims with source passages.
• Label omitted limitations separately from incorrect statements.
• Repeat comparable prompts rather than drawing conclusions from one response.
• State that the underlying context and selection process were not visible.

OpenAI’s search documentation warns that citations and results can be incomplete or wrong. Visible omissions are findings about the answer; they are not direct measurements of the assistant’s hidden context window.