Context Window
The context window is the total amount of text a language model like ChatGPT, Claude, or Gemini can hold in active memory during a single exchange. It is measured in tokens, and it covers everything at once: the system prompt, your input, any retrieved documents, the conversation so far, and the answer being generated. Once that budget is full, older content gets pushed out and the model can no longer refer back to it.
Why it matters for your AI visibility
When an AI search tool such as Perplexity, ChatGPT, or Google AI Overviews puts together an answer, it pulls a handful of source passages into its context window, not your whole page. If your key facts are buried under throat-clearing, spread across a long unstructured page, or repeated three different ways, the model may load an excerpt that misses your actual point, or skip your page for a competitor's tighter one. The fix is not writing "for AI" with hidden markup or special files: Google has said directly that no schema, no llms.txt, and no AI-specific content are required for AI Overviews or AI Mode, and warns against maintaining a separate version of your content for machines. What actually helps is ordinary clarity, your core claim, your name, and your differentiator stated plainly near the top, in a paragraph that stands on its own. That is what a retrieval step is most likely to lift into the window intact.
How it works technically
A model breaks text into tokens, roughly three-quarters of a word in English on average, sometimes a whole word, sometimes a fragment. The context window states how many tokens can occupy the model's working memory at once, commonly well over 100,000 in current systems. Every part of the exchange shares this same budget: the instructions the assistant is running under, your message, any documents or search results fed in, the back-and-forth history, and the reply itself. When the budget is exhausted, the oldest material is dropped to make room for new text, which is why a long chat can suddenly seem to forget something you said earlier, the model has not misunderstood, the tokens are simply no longer present. A bigger window lets a system reference longer documents, but it also costs more to run per request, and a larger window does not by itself make a model use its contents evenly.
Common mistakes
The most common misread is assuming a larger context window means better answers. Research on long-context models has repeatedly found a "lost in the middle" effect: models attend more closely to the start and end of a long input than to material buried in the center. A claim placed in the middle of a dense, unbroken block of text is genuinely more likely to be ignored, regardless of window size. A second mistake is over-stuffing the window with marginal or repetitive content on the theory that more context helps; padding a prompt or a page with tangential material dilutes the signal and raises the odds of a garbled or hallucinated answer. Columbia's Tow Center found that across eight AI search tools asked to identify a news article's source, headline, date, and URL, more than 60% of responses were wrong, ChatGPT alone misidentified 134 of 200 articles, a reminder that even well-resourced systems misattribute source material when the right passage was not cleanly available to pull. For content meant to be cited accurately, structure beats volume: short self-contained paragraphs, real headings, and a stated conclusion near the top of each section.
Relation to AI citations
Whether an AI system cites you depends on what actually made it into the context window at answer time. Under retrieval-augmented generation, the system retrieves a small number of passages from a search index or database and inserts only those into the window, the rest of your site is invisible to that particular answer. This is one reason AI citation behaves differently from search ranking: Ahrefs' analysis found only about 6 to 8% overlap between URLs ChatGPT cites and Google's top-10 for the same query, and roughly 80% of ChatGPT-cited URLs do not appear in Google's top 100 results at all. It also helps explain why third-party mentions matter more than most sites assume, Ahrefs found brand mention frequency across the web correlates with AI citation rate at about 0.664, roughly three times the correlation seen for backlinks alone. A paragraph that reads clearly in isolation, without depending on context from three sections earlier, is simply more likely to survive being lifted alone into a model's limited window and attributed to you correctly.
Example
Say a small tax firm in Leipzig uploads its 80-page service and pricing handbook as a PDF and asks an AI assistant about the fee for a specific filing type. If the whole document fits inside the context window, the assistant can locate the right line. If the file is too long for the window, only part of it gets loaded, and the exact figure being asked about may fall outside that portion. The assistant then either answers incompletely or guesses at a number that was never actually in the document. A shorter, clearly labeled overview with real subheadings makes it far more likely the relevant figure lands inside the window and gets quoted correctly, rather than invented.
Common questions
Is a bigger context window always better?
Not on its own. A larger window allows longer input, but models still attend less to content buried in the middle of a long text than to the start or end. Clear structure and a well-placed core claim matter more than raw length.
What does the context window have to do with tokens?
The window's size is stated in tokens, the word-fragment units models process text in. Your input, the conversation history, any retrieved documents, and the model's reply all draw from the same token budget. Once it's full, the oldest content is dropped to make room.