Build a source inventory before importing documentation

Create one row per authoritative document

Record the canonical URL, title, product version, owner, last reviewed date and the reader task it supports. Add an inclusion decision and reason. This simple inventory makes accidental scope visible: an old tutorial, a marketing promise and a current help article should not silently carry the same authority. Start with the smallest collection that answers the pilot’s questions, not every crawlable page on the domain.

Separate unknown from current

A page without a review date is not necessarily wrong, but it is not verified current merely because the URL still loads. Ask the responsible owner or compare it with a known current specification. Mark unresolved rows for review rather than inventing dates. The inventory is an operational record you maintain; it is not a certification that a crawler or chatbot will import every passage correctly.

Use a concrete exclusion rule

In a fictional manual, include the current export guide and documented limits, exclude the retired beta tutorial, and hold an unreviewed blog post that contradicts the guide. Preserve each reason so another editor can revisit it. Exclusion is not deletion: the original page may remain publicly available for historical users. Your chatbot collection can have a narrower boundary than the website itself.

Check the inventory against the actual import

After an authorized import, compare the product’s visible source list with your approved rows. Look for missing documents, unexpected domains and duplicate versions. Keep a dated snapshot before testing answers so failures can be connected to a particular source collection. Do not start by changing prompts when the relevant document was never included. The next step is a small answer test, not a larger indiscriminate crawl.

• A practical row contains URL, version, owner, reader task, inclusion decision, review date and the reason for any unresolved hold.