A context window is the maximum amount of text, measured in tokens, that a large language model can take into account at one time when it writes a response. Everything the model is working with, including instructions, the conversation, retrieved web pages and its own reply, has to fit inside it.
This page explains what a context window is, how tokens work, how big current windows are according to official documentation, and why the concept matters for how AI tools read, summarize and quote your pages.
What is a Context Window?
Think of the context window as a model's working memory for a single request. Anthropic's documentation on context windows describes it as all the text a model can reference while generating a response, including the response itself, and makes the point that it is separate from the data the model was trained on.
That distinction matters. Training data is what the model learned months or years ago. The context window is what it is looking at right now. When an AI assistant searches the web and reads your page, your content enters the context window for that conversation. It does not become part of the model's long-term knowledge.
Context windows are measured in tokens, not words. OpenAI's help article What are tokens? explains that tokens are pieces of words, and that in English one token averages about four characters, or roughly three quarters of a word. So 1,000 tokens is around 750 English words.
Why Context Window matters
For marketers, the context window explains a lot about how AI search treats content. When a tool like ChatGPT, Perplexity, Gemini or Claude answers a question with live web results, it typically pulls in several pages plus the user's conversation, system instructions and tool definitions. All of that competes for space and, more importantly, for the model's attention.
Windows today are large. But bigger is not the same as better. Anthropic's documentation warns that accuracy and recall can degrade as the amount of context grows, a problem it calls context rot. Google's long context guide for the Gemini API similarly notes that performance can drop when a model has to find several pieces of information at once in a long input.
Research backs this up. The 2023 paper Lost in the Middle, published in Transactions of the Association for Computational Linguistics, found that models did best when relevant information sat near the start or end of a long input and worse when it was buried in the middle. Models have improved since, but the lesson holds: clear, well-placed information is easier for a model to use than facts scattered through a long page.
The honest limit: you cannot see or control the context window of someone else's AI product. You do not know how many pages it retrieved, how much of your page it kept, or how it was trimmed. What you can control is how easy your content is to understand and quote once it gets in.
How Context Window works
What counts toward the window
Per Anthropic's docs, the window includes the system prompt, every message in the conversation (including tool results, images and documents), tool definitions and the output generated for that turn. Reasoning or "thinking" tokens count too. When an assistant runs a web search, the search results and fetched page text take up part of that budget.
Current sizes, from official docs
As of October 2026, the major providers document windows of roughly a million tokens for their flagship models:
- Anthropic: the Claude models overview lists a 1M-token context window and up to 128K output tokens for its current models. Some older models use 200K.
- OpenAI: the OpenAI API models page lists about 1.05M tokens for its current flagship models.
- Google: the Gemini API long context guide says many Gemini models have windows of 1 million tokens or more.
These figures apply to the developer APIs. Consumer apps may use different limits, trim retrieved content or summarize earlier parts of a conversation, and providers change them often. Always check the current docs before quoting a number.
What happens when it fills up
If the input is too large, the request fails or older material has to be dropped, summarized or cleared. Developer tools now offer features like compaction, which summarizes earlier context, and context caching, which lowers the cost of reusing the same large input. For AI search, the practical effect is that retrieval systems pick a handful of relevant passages rather than stuffing entire websites into the window. That is the core idea behind retrieval-augmented generation.
Context Window in practice: an example
Here is a hypothetical. A user asks an AI assistant, "What is the best CRM for a 10-person real estate team?" The assistant runs a few searches and pulls text from six pages: two review sites, a Reddit thread, two vendor pages and a comparison blog post.
One vendor's pricing page is long, heavy on scripts and splits its plan details across tabs that need JavaScript to load. Little of the useful information makes it into the context in readable form. The other vendor's page states plan names, prices, seat limits and the target customer in plain HTML near the top. When the model writes its answer, the second vendor's details are easy to find and quote, so that vendor is far more likely to be described accurately and cited.
Neither page "won" because of the window size. The second page simply made the most of whatever share of the window it got.
How to optimize content for Context Windows
- Put the answer first. State the key fact, definition or recommendation near the top of the page and of each section, so it survives truncation and sits where models read most reliably.
- Write self-contained sections. Each section should make sense on its own, with the subject named rather than implied. This is the idea behind content chunking.
- Serve important text in plain HTML. Some AI fetchers do not run JavaScript. Anthropic's web fetch tool documentation, for example, says it does not support pages rendered dynamically with JavaScript. Keep core content in the initial HTML.
- Cut filler. Long intros, repeated boilerplate and keyword padding spend tokens without adding meaning. Shorter, denser pages give models more signal per token.
- Use clear headings and lists. Descriptive headings and simple lists help both retrieval systems and models locate the right passage.
- Keep facts consistent. If your pricing or specs differ across pages, a model reading several of them in one context may pick the wrong one or hedge. Align them.
Common mistakes and misconceptions
"A bigger context window means the AI reads my whole site." No. AI search tools retrieve a small number of relevant pages or passages per question. A large window allows more material, but retrieval still decides what gets in.
"If the AI read my page, it will remember it." Content in the context window is used for that response only. It does not update the model's training. Future answers depend on fresh retrieval or future training runs.
"Tokens and words are the same." A token is a piece of a word. In English it averages about three quarters of a word, and other languages, code and unusual terms can use more tokens per word.
"Long-form content always performs better in AI." Length is not the goal. A long page that buries its answers can be harder for a model to use than a shorter page that states them clearly.
Context Window vs. Training Data
Training data is the large body of text a model learned from before release; it shapes what the model knows by default and has a cutoff date. The context window is the temporary space for a single request, where live information such as search results can be added. To influence what AI says about you today, you mostly work on what gets retrieved into context. Training data changes only when models are retrained.
How Butter approaches Context Windows
We treat the context window as a reason to write clearly, not a trick to exploit. Our focus is fundamentals: pages that load their key content in HTML, put answers up front and keep facts consistent across a site.
Through our GEO plans, which start at $399/month, clients get monthly AI prompt testing, indexing strategy, schema markup guidance, knowledge graph submissions and on-page recommendations. On-page recommendations often include restructuring sections so they stand alone and moving buried facts higher on the page.
We also test how AI tools actually describe a brand, because that is the only reliable way to see whether your content is making it into answers in usable form.
Frequently asked questions
How big is ChatGPT's context window?
As of October 2026, OpenAI's API documentation lists about 1.05 million tokens for its current flagship models. The ChatGPT app may apply different limits depending on plan and feature, so check OpenAI's current documentation for exact figures.
How many words are in 1 million tokens?
Using OpenAI's rule of thumb of about three quarters of a word per token in English, 1 million tokens is roughly 750,000 words. The real number depends on the language and the kind of text.
Does the context window affect SEO?
Not traditional Google rankings directly. It does affect how AI tools that read your pages extract and quote information, which matters for AI visibility and citations.
What happens when a conversation exceeds the context window?
Depending on the product, the request may fail, older messages may be dropped, or earlier parts of the conversation may be summarized to make room. Details lost this way are no longer available to the model.