--- title: "How to Train a Site Assistant on Your Own Articles" description: "Learn how to train a chatbot on website content you already published: limit the index to QA'd articles, cite canonical URLs, and hold stale or duplicate pages." url: "https://articles.flowcrews.com/how-to-train-a-site-assistant-on-your-own-articles" category: "Concierge" date_published: 2026-10-01 reading_time_minutes: 7 word_count: 1690 --- # How to Train a Site Assistant on Your Own Articles *A fail-closed intake contract: which published articles may be indexed, how a section becomes a citable chunk, and when to hold the rest.* ![How to Train a Site Assistant on Your Own Articles](https://pub-07fb5e4955ba485b822d6b388be96d9a.r2.dev/2e474d51-a072-472e-ab2f-1a43a8990566/how-to-train-a-site-assistant-on-your-own-articles/hero-ba36830e-c508-464e-8c4b-04243c0cf01b.jpg) **TL;DR:** - Training here means retrieval over indexed articles, not fine-tuning the model. - Index only owned, current, canonical pages; hold carts, archives, print views, and duplicate or conflicting policies. - Write so one heading is one answer a chunk can stand on by itself. - A small curated set beats a dump of the whole site, and a stronger model cannot fix missing or contradictory pages. - Citations must be an allowlist of indexed canonical URLs, and off-corpus or unstated prices should be refused. ## Training Means Retrieval, Not a New Model ## Admit Only Owned, Current Articles ## Write Each Section So a Retrieved Chunk Can Stand Alone ## Why stale and contradictory pages stay out of the index ## Prove Retrieval Before You Widen the Corpus ## Conclusion - Retrieval, not fine-tuning — Train the assistant by indexing owned pages and supplying retrieved passages as evidence; fine-tuning blends facts into the model and multiplies errors, while a named assistant without a site knowledge base remains a generic model. - Owned and current only — Admit a limited, QA-cleared crawl of owned pages at depth 0, shaped by real visitor logs, and keep carts, tag archives, print views, translations, and conflicting policy copies out; a focused set beats a large, poorly organized corpus. - Self-contained chunks — Write one topic per section, lead with the fact, and keep the product or policy, condition, and supporting detail in that same section so a retrieved passage never depends on “as noted above.” - Metadata on every chunk — Attach title, canonical URL, updated date, and section label; enrichment alone can lift precision without changing the retriever, embeddings, or chunking. - Hold, don’t index both — Keep stale, duplicate, or contradictory pages out of the index, never quote a price the current page does not state, and re-index changed pages instead of training a new model for a content edit. - Prove before you widen — Citations must be an allowlist of known chunk IDs on indexed canonical URLs; test a fixed set of real visitor questions and unique single-page facts, and treat a miss as a hold rather than a reason to add sources or switch models. Start with a small, owned, fail-closed index and a fixed question set, then let Flowcrews help you keep that site assistant tied to pages you can actually cite. ## Frequently Asked Questions ### Does training a site assistant mean fine-tuning it on my articles? No. In this setup, training means indexing pages so the assistant can retrieve passages as evidence. WebChatAgent states that it does not retrain or fine-tune the underlying model. Current practice is RAG for facts and fine-tuning, if used at all, for style. ### Should I index the whole website? No. Start with pages you own, at a tight crawl scope, and exclude carts, tag archives, print views, and duplicate URLs. Conferbot’s guidance is that a focused set of 30–50 high-quality articles typically beats about 500 poorly organized documents. Entexis runs a production assistant on 63 sources and 271 chunks. ### Why is a stale article worse than leaving a topic out? A missing page forces a refusal or a gap. An outdated page can override a correct answer. Stale-document research found outdated retrieval flipped 30% of Llama and 37% of Qwen answers even without instructions to trust the document, while matched current evidence was followed in 97–100% of trials. ### Can the assistant invent links to my articles? It should not. Citations need to be an allowlist of indexed canonical URLs and known chunk IDs. Ahrefs found AI assistants sent visitors to 404 pages 2.87 times more often than Google Search, and 2.38% of ChatGPT-cited URLs in a crawler check returned a 404. ### How do I know the right article was retrieved? Keep a fixed set of real visitor questions and the source fact each one should hit, then rerun it after content or model changes. WebChatAgent recommends collecting 10 to 20 real customer questions with expected source facts, including facts that exist on only one page. ### Should the assistant quote prices from old pages? Only if the current indexed page states that price, and many operators still refuse specific dollar amounts. Entexis barred its production agent from quoting specific dollar amounts after an unsanitized knowledge-base price set the wrong expectation on a later sales call. ## Sources - [https://webchatagent.com/tutorials/train-chatbot-on-website](https://webchatagent.com/tutorials/train-chatbot-on-website) - [https://entexis.in/how-we-built-ai-agent-entexis-website-chatbot-case-study](https://entexis.in/how-we-built-ai-agent-entexis-website-chatbot-case-study) - [https://botnation.ai/en/train-chatbot-own-data/](https://botnation.ai/en/train-chatbot-own-data/) - [https://www.conferbot.com/blog/train-chatbot-on-knowledge-base](https://www.conferbot.com/blog/train-chatbot-on-knowledge-base) - [https://atlan.com/know/llm-knowledge-base-data-quality/](https://atlan.com/know/llm-knowledge-base-data-quality/) - [https://arxiv.org/abs/2609.31342](https://arxiv.org/abs/2609.31342) - [https://ahrefs.com/blog/how-often-do-ai-assistants-hallucinate-links/](https://ahrefs.com/blog/how-often-do-ai-assistants-hallucinate-links/) - [https://keganquimby.com/blog/how-to-build-rag-knowledge-base-from-sitemap](https://keganquimby.com/blog/how-to-build-rag-knowledge-base-from-sitemap)