Feed your own assistant

Sooner or later you put an assistant on your own site, or a Custom GPT in front of your team, or a Copilot agent over your documents. Then it confidently tells a customer something that stopped being true in 2023.

That is not a model problem. It is a corpus problem, and the corpus is yours.

Why your own chatbot gets your own facts wrong

Almost always one of five things:

1. Two pages disagree. Your pricing page says one thing, an old blog post says another, and nothing marks which is current. A retrieval system has no way to prefer one, so it picks whichever matched better.
2. The answer is in a PDF. Parsed badly, or not at all.
3. The chunk has no context. A paragraph beginning "this includes" is meaningless once it is lifted out of the page.
4. Nothing is dated. Freshness cannot be judged, so stale content competes on equal terms with current content.
5. The terminology drifts. You call it a "site visit" on one page and an "inspection" on another. Someone asking about one will not retrieve the other.

Every one of those is fixable, by you, without anyone's permission.

The overlap that makes this cheap

Here is the useful part: the work is nearly identical to the first pillar.

Clear headings create retrievable chunks. Answers stated in the first sentence survive extraction. Dates let freshness be judged. Consistent naming lets a query match. Content in HTML rather than images is parseable.

Do it once and it serves both. That is the strongest argument for structural work on this whole site: the vendor-facing benefit is probabilistic, but the benefit to your own assistant is direct and measurable.

What is different about the second pillar

A handful of things matter here that do not matter for vendor visibility:

• One canonical answer per question. For a public search engine, three pages on a topic is depth. For your own retrieval system, it is three competing answers.
• A llms-full.txt or markdown export. A clean, single-file version of your content is far easier to load into a Custom GPT or Project than asking it to crawl your website.
• Deliberate deletion. Old content that ranks is an asset. Old content in your assistant's corpus is a liability.
• Terminology discipline. Pick a word for each thing and use only that word.
• A review cadence. Retrieval corpora go stale silently. Nobody complains — the assistant just answers wrongly, politely.

Where this is going

Guides on chunking for retrieval, writing canonical answers, auditing your corpus for contradictions, exporting content for Custom GPTs and Projects, and keeping a knowledge base honest over time.

If you want to be told when they publish, or you have a specific problem with an assistant that is answering wrongly, tell us — the early guides will be shaped by what people actually ask.