Glossary · Beginner
What Is a Context Window? Definition and Example
A context window is all the text a language model can reference when generating a response, including the response itself. That's Anthropic's definition, which also describes it as "a 'working memory' for the model", separate from "the large corpus of data the language model was trained on" (Anthropic).
On this page
- How does a context window work?
- What's an example?
- Why does it matter for GEO?
- What are the common misconceptions?
- From MentionX data: how often answers show their sources
- Worked example: why Ledgerline's pricing wasn't quoted
- What fills the context window in AI search?
- What can marketers do about it?
- Common questions
- Related
- Sources
Checked on 31 October 2026. "Our view" marks MentionX opinion. Ledgerline is a fictional company used for illustration.
How does a context window work?
Everything in a conversation counts: the system instructions, your messages, any retrieved web pages or files, and the model's replies. Anthropic's docs explain that as a conversation advances, "each user message and assistant response accumulates within the context window" (Anthropic). Windows are measured in tokens. Anthropic shows Claude windows of "1M (or 200k) tokens" depending on the model, and notes that chat interfaces can manage the window on a rolling "first in, first out" basis (Anthropic).
What's an example?
Fictional example: an AI search engine retrieves eight web passages for a question and places them in the context window alongside the question. The model writes its answer from those passages. A page that wasn't retrieved isn't in the window, so the model can't quote it in that answer.
Why does it matter for GEO?
When an engine grounds an answer in web results, only the retrieved text sits in the context window. Our view: that's why concise, self-contained passages help. If the key fact is buried deep in a long page, it may be cut or left out when the engine chooses what to include.
What are the common misconceptions?
- "A bigger window means the model reads my whole site." It reads only what the system puts in.
- "The model remembers past chats." Not unless the product adds them to the window or uses a separate memory feature.
- "Context equals training." Training shapes the model. Context is temporary input for one response.
From MentionX data: how often answers show their sources
Method. Read-only aggregate queries over MentionX scans from 22 August to 29 October 2026 (UTC): 2,217 engine runs (2,154 completed) across 18 anonymised projects and six engines.
We can't see what's inside an engine's context window. We can see whether answers cite sources, which is the visible trace of retrieved material:
| Engine | Brand mentioned | Answers with no citations |
|---|---|---|
| Gemini | 47.3% | 50.7% |
| OpenAI | 42.9% | 30.2% |
| Perplexity | 31.5% | 4.1% |
| AI Mode | 31.3% | 12.6% |
| Copilot | 26.0% | 47.0% |
| DeepSeek | 24.0% | 99.0% |
Engines that rarely cite may still use retrieved text, or may answer mostly from training. From the outside, you can't always tell which.
Our view: for engines that cite, the cited pages are your best clue to what went into the window. Track them, and look at what those pages say that yours doesn't.
Limits. A convenience sample of mostly B2B software projects, run through engine APIs, which can behave differently from the consumer apps. Perplexity, Copilot and DeepSeek figures come from a single day (26 October 2026), so treat them as a snapshot, not a trend.
Worked example: why Ledgerline's pricing wasn't quoted
Ledgerline is a fictional accounts-payable tool. Figures are illustrative.
A buyer asks an AI search engine, "How much does Ledgerline cost for a team of ten?" The engine retrieves a handful of passages, including a review site's summary but not Ledgerline's pricing page, which shows prices only inside an interactive calculator.
Because the pricing page's text never entered the context window, the model answered from the review site's outdated figure. Ledgerline adds a plain-text pricing table to the page so a future retrieval can include the current prices directly.
What fills the context window in AI search?
Typically: the product's own instructions, the user's question and any earlier turns in the conversation, the passages retrieved from the web and the model's reply as it's written. Each takes up space measured in tokens. Products decide how many passages to include and how long each can be; they don't publish those limits for consumer search.
What can marketers do about it?
You can't control the window, but you can make your passages worth including. Answer the question early in each section, keep key facts in text rather than images or scripts and avoid long preambles before the point. Those habits help human readers in exactly the same way.
Common questions
Is the context window the same as memory? No. Memory features store information between chats; the window is what the model can use in one response.
Does a longer window make answers more accurate? Not by itself. Accuracy depends on what's put in the window and how the model uses it.
Related
Sources
- Anthropic, Context windows (accessed 31 Oct 2026)
Questions
Is the context window the same as training data?
No. Anthropic's docs describe it as working memory, separate from the large corpus the model was trained on.
How big are context windows?
It varies by model. Anthropic's documentation shows Claude context windows of 1M or 200k tokens, depending on the model.