Claude for your website
6 min read · updated August 27, 2026
Putting Claude on your website means giving it your pages at answer time and constraining it to those pages. The interesting part is not generation quality — it is what the system does when your pages do not contain the answer.
Recall is the wrong mechanism for business facts
Anything a model states from memory is a claim about the world as of its training data. Your delivery charge, your consultation fee, your opening hours and your returns window are not stable facts about the world; they are facts about you, and they change.
Retrieval fixes the mechanism rather than the symptom. The relevant passage of your own page is put in front of the model at the moment of answering, and the model is instructed to answer from it. When your page changes, the answer changes, with no retraining involved.
Why refusal behaviour decides the outcome
Every retrieval system will sometimes retrieve nothing useful. Someone asks about a service you do not offer, or phrases a question so unusually that nothing matches. What happens next is the whole ballgame.
A system that treats answering as mandatory will assemble something from whatever it retrieved, however thin the match. A system that treats not knowing as an acceptable outcome will say so and offer a handoff. For customer-facing support the second is the only defensible default, because a wrong answer is acted on and a handoff is not.
How to test refusal before you ship
This is testable in about ten minutes and almost nobody does it.
- Ask something your site genuinely does not cover — a service you do not offer, in the phrasing a confused customer would use.
- Ask something adjacent to a real page but materially different: if you publish a 30-day returns window for one product category, ask about a category you do not mention.
- Ask for a specific number your site does not publish, such as a fee or a lead time. This is where invented answers appear most readily.
- For each, check whether it declined, and whether the answer it gave cited a page that actually supports it.
What you still have to build
Model access is the smallest part. Around it sits crawling your site and re-crawling it on a schedule, splitting pages into passages, embedding and storing them, retrieving per question, streaming the reply, rate limiting, and a widget that does not collide with your site's styles.
That list is the honest reason hosted products exist. It is all buildable, and none of it is the interesting part of your business unless answering is your business.
Questions
- Can Claude read my website directly?
- Not as a standing capability for your visitors. A person can paste a page into a conversation. Answering customer questions on your site continuously means a system that fetches and indexes your pages and supplies the relevant ones per question.
- Does a bigger context window remove the need for retrieval?
- Not for a real site. Even where a whole site fits, sending all of it on every question is slow and expensive, and burying the relevant passage among thousands of irrelevant ones tends to make answers worse rather than better.
- How do I stop it answering outside my content?
- Instruction alone is not sufficient. It needs retrieval that returns nothing when nothing matches, and an answering step that treats an empty retrieval as a reason to decline rather than a prompt to improvise.
Related
tool
AI answer generator
Enter a website and a customer question, and get the answer as an AI representative trained on that site would give it — with the page it came from, or an honest "not covered".
tool
Website knowledge base generator
Reads your website and organises it into a knowledge base structure — topics, the pages that belong to each, and the topics customers ask about that your site doesn't cover. Exports as Markdown.
tool
Website support score
Scores a website on how well it could support customers — and how well an AI representative trained on it would perform. Five dimensions, a total out of 100, and the specific gaps behind the score.
term
What is grounding in AI?
A grounded system retrieves passages before it writes anything, puts them in front of the model, and instructs it to answer only from what it was given. The test of whether a system is grounded is simple and does not require access to its internals: ask it something, then ask it where the answer came from. A grounded system names a page. An ungrounded one describes its reasoning.
term
What is an AI hallucination?
Language models predict plausible text, not true text. Asked a question with no supporting material, a model does not fall silent; it produces the most likely-sounding answer, which is often a wrong price, a policy that does not exist, or a citation to a page that was never written. The behaviour is not a bug to be patched out of the model. It is what the model does when nothing constrains it.
term
What is a confidence score in AI answers?
Two very different things get called confidence. Retrieval confidence measures how close the best-matching passages were to the question — a property of the search step, computed before any answer is written. Model confidence is the model's own assessment of its answer, which it will happily report as high while inventing a price.
Want to skip to the working version?
RubyRep does everything in this guide from a single URL — reads your site, builds the knowledge, and gives you one line of code to install.
Train My RubyRep