Claude

What does Claude read from a website?

Two different things, depending on when you ask. Claude answers from training up to the model's knowledge cutoff, and from pages Anthropic's agents fetch at answer time. Each agent leaves a different trace in your log and answers to a different robots.txt rule:

• Claude-User

  What it is for

  a person asked Claude something and it opened your page

  What a visit tells you

  Claude used your page in a live answer, there and then

• Claude-SearchBot

  What it is for

  building Claude's search results

  What a visit tells you

  your page can be found when Claude searches

• ClaudeBot

  What it is for

  collecting content that may contribute to training

  What a visit tells you

  your page may shape a future model, not today's answer

Claude-User is the row to watch. It is the most direct evidence a log can give that Claude read you for a real question. Our own totals for attensira.com, including the share that was Claude-User, are on Anthropic; further down this page are the pages it opened.

Our workspace reads ChatGPT and Perplexity answers, not Claude's, so this page quotes no Claude answer rate; a number we did not measure is not zero. The answer at the top of this page is Perplexity's, to a question about tracking Claude. It describes the method fairly: run the same questions on each assistant many times and record what comes back.

What is Claude?

Claude is Anthropic's family of large language models and the assistant built on them. Anthropic's models overview calls it a family, and the plural matters. As read on 4 October 2026:

• Claude Fable 5.1

  Context window

  1M tokens

  Reliable knowledge cutoff

  Jun 2026

• Claude Opus 5.5

  Context window

  1M tokens

  Reliable knowledge cutoff

  Jun 2026

• Claude Sonnet 5.5

  Context window

  1M tokens

  Reliable knowledge cutoff

  Jun 2026

• Claude Haiku 4.5

  Context window

  200K tokens

  Reliable knowledge cutoff

  Feb 2025

Older models remain available too. Anthropic notes 1M tokens is roughly 555k words on its current tokenizer. This lineup changes; the page you are reading had a different one a month earlier.

Two of these properties change what Claude can say about you:

• Knowledge cutoff. What a model knows without searching stops at its cutoff. A rebrand, a new product or a price change after that date does not exist in its memory.
• Context window. How much text it can read at once. Even a million tokens is finite against the web, so something still chooses which passages enter. See context window.

How does Claude use the web?

When Claude searches, it does retrieval-augmented generation: it fetches pages, puts selected passages into its context and writes from them, citing what it used. That is grounding, and it is the only way anything published after a model's cutoff reaches an answer.

It depends on Anthropic's agents being able to fetch you. Anthropic documents three, each controlled separately in robots.txt: ClaudeBot for training, Claude-SearchBot for search, and Claude-User for fetches a person triggers. Anthropic publishes the IP ranges they use at claude.com/crawling/bots.json, so a log entry can be verified rather than trusted.

Why do Claude's answers about you vary?

Because "what does Claude say about us" is several questions at once:

• Which model answered. A smaller, older model and a newer one have different cutoffs and different knowledge of you.
• Whether it searched. An answer from memory and an answer grounded in your current pages can disagree.
• Run to run. Language models are not deterministic; the same question asked twice can be answered differently.

So one Claude answer is one observation. Read many before concluding anything.

What are the common mistakes about Claude?

A blanket block. A rule naming every Anthropic agent, or a broad wildcard, can reduce your visibility in Claude's answers as a side effect of a training decision.

Reading a stale answer as a hallucination. A model that never saw your rebrand will describe the old brand confidently. It is correctly recalling an old fact. The fix is making the new fact fetchable. See AI hallucination.

Assuming a citation validates the sentence. Retrieval and writing fail separately; a correct link can sit under a wrong summary.

Writing pages, not passages. Claude selects sections. A claim whose qualifier lives three paragraphs away arrives without it.

How does Attensira track Claude?

Claude is one of the engines Attensira can read every day: it asks your buyers' questions, stores Claude's answers verbatim, and records whether you are named and linked, with the sample size. Its crawler logs show which Anthropic agents fetched which pages, and model tracking explains which assistants a plan reads. The crew then fixes access and passages and ships each change after your approval. The free audit asks Claude eight questions about your category and shows whether it named you.