Context window

How much material goes into one answer?

More than the answer can use. In the 30 days to 4 October 2026, read over the Attensira MCP, ChatGPT attached 78,667 citations to 2,349 answers to the questions our workspace tracks: about 33.5 per answer. Perplexity attached 21,713 to 2,175 answers, about 10.0 each. A citation here is one URL the engine attached to its answer.

We cannot see inside either engine's context window. From the outside, though, the shape is clear. A ChatGPT answer that carries thirty-odd sources is a few paragraphs long, so most of those pages contribute a sentence at most. Being retrieved into the window is not the same as being used in the answer.

The two engines also differ by a factor of about three on the same questions. A page competing for a ChatGPT answer shares the window with far more material than the same page in a Perplexity answer. These readings cover ChatGPT and Perplexity only; Claude and Gemini answers are not read in this workspace.

What is a context window?

It is the maximum amount of text a language model can process in one request, measured in tokens. Everything in the request shares it:

• the system instructions;
• the conversation so far, in conversational search;
• documents retrieved for this answer, in retrieval-augmented generation;
• the answer the model writes.

Anything that does not fit is not seen. The model has no other view of the world in that moment except what it learned in training.

How large are context windows now?

Large. As read on 4 October 2026, Anthropic's models overview lists:

Anthropic notes 1M tokens is roughly 555k words on its current tokenizer. Google says many Gemini models accept 1 million or more tokens. These numbers change with every release, so treat any table, including this one, as dated.

If windows are so big, why do answer engines still select?

Because the web is bigger. A million tokens holds a long book or a mid-sized documentation site. It does not hold the web, or even the top results for one question once you include every page's navigation, scripts and boilerplate. So answer engines still search, chunk, score and select a few passages per question.

Two more reasons selection persists:

• Cost and speed. Every token in the window costs compute and time on every answer.
• Attention is not uniform. Anthropic's retrieval research found that passing the top 20 chunks beat passing 5 or 10, but more is not endlessly better: extra material can distract a model. Engines tune how much to include rather than including everything.

Anthropic's own rule of thumb: a knowledge base under about 200,000 tokens can go in the prompt whole. Above that, retrieve.

What does this mean for your pages?

Your page enters the window in pieces, if at all, and competes with pieces of other pages:

• One question per section. A section that answers one thing is a clean candidate.
• Self-contained passages. The subject, the claim and its qualifier in the same few sentences, because the paragraph above may not be in the window.
• Your name in the sentence. If the selected passage does not name you, the answer can use your fact without crediting you. See AI brand mentions.
• Less boilerplate. Content that renders in the HTML, without walls of navigation before it, is cheaper to read and easier to extract.

What are the common misunderstandings about context windows?

"A 1M window means the model reads my whole site." It reads what the application chooses to send, usually a few passages.

"The window is memory." It is per request. Memory features store notes and resend them.

"Bigger windows make SEO irrelevant." They make selection less tight, not unnecessary. Retrieval still decides what is read.

"Fill the page with everything, the model will sort it." Extra material can dilute the passage that matters.

How does Attensira use this?

Attensira records which of your pages each engine cites and the sentence of the answer each link supports, so you can see which passages made it from the window into the answer. The crew rewrites the passages that are read but not used, ships each change after your approval, and measures named and linked rates before and after. The goal is to be one of the few sources the answer actually uses, not one of the thirty it attaches. Citation tracking shows how each link is tied to the sentence it supports.