What the documentation says
Anthropic's platform documentation defines it in two sentences, and the second one is the part that gets lost:
"The 'context window' refers to all the text a language model can reference when generating a response, including the response itself. This is different from the large corpus of data the language model was trained on, and instead represents a 'working memory' for the model."
Two things follow immediately.
Training and context are separate stores. What a model learned during training is not in the context window. What is in the window is what was put there for this one exchange: the question, any documents retrieved, earlier turns of the conversation, and the growing answer.
The answer competes for the same space. A long answer eats the room available for the material it is answering from.
More context is not better
The same documentation says this outright, and it runs against the way capacity gets marketed:
"A larger context window allows the model to handle more complex and lengthy prompts, but more context isn't automatically better. As token count grows, accuracy and recall degrade, a phenomenon known as context rot. This makes curating what's in context just as important as the window size."
Accuracy and recall degrade. Not throughput, not cost. The model becomes worse at finding and using what is in front of it as more is placed there.
For anyone whose pages are the material being placed there, that reverses an intuition. The risk is not only that a long page gets cut off before the relevant part. It is that a page which arrives whole, inside a window already carrying a conversation and several other sources, is competing for attention that measurably thins out.
More context is not automatically better.
Accuracy and recall ↑
Why this shows up as "the model ignored my page"
Site owners report a pattern: a page is clearly retrieved, clearly relevant, and the answer does not use it. Context rot is one plausible mechanism, and it is documented by a model provider rather than inferred.
It also explains why the position of an answer inside a page matters more than page length. A page that states its answer in the first screen puts that answer near the start of whatever gets placed in the window. A page that builds to it over two thousand words puts the answer where the degradation is worst, assuming the whole page even fits.
What to do with it
Put the answer where it survives truncation. Near the top, complete, in one place. Not assembled from three sections a reader would have to combine.
Write the passage so it stands alone. Anything that depends on a sentence eight paragraphs earlier loses its meaning when only part of the page is in the window.
Stop treating length as depth. More words means more tokens competing with the answer being generated, and the documentation says what happens then.
Do not quote a window size as a fixed fact. Capacities differ per model and change with releases. What holds across all of them is the shape: the window is working memory, the answer is inside it, and filling it costs accuracy.
Source
- Anthropic, "Context windows", Claude Platform documentation (retrieved 3 September 2026): https://platform.claude.com/docs/en/build-with-claude/context-windows — source of the quoted definition including "including the response itself" and the distinction from the training corpus, and of the quoted passage on more context not being automatically better, on accuracy and recall degrading as token count grows, and on the term "context rot".
- The documentation is one provider's, and the definition it gives is general rather than product-specific. Where this entry describes consequences for pages rather than for prompts, that reading is ours and is marked as such in the text.
Frequently asked questions
Is the context window the same as what the model knows?
No. The documentation separates the two explicitly: the training corpus is not the context window, which is described as "working memory" for one exchange.
Does a bigger context window solve the problem of long pages?
Not on its own. The same source says more context is not automatically better and that accuracy and recall degrade as the token count grows.
Does the answer itself take up space in the window?
Yes. The definition says "including the response itself".
How many tokens is a context window?
It depends on the model and changes with each release, so a number here would age badly. The behaviour that does not change is that the window holds question, sources and answer together.