A business owner adds an assistant to a WordPress site with forty or so pages built over six years.
It answers well on the main services and confidently wrong on two things, because a page from four years ago describes a service that no longer exists and a pricing page was superseded without being removed.
Neither page had been seen by a human visitor in years. Sitting unvisited they were harmless, and as training material they were not.
Step 1: list what is actually published
Including the pages nobody remembers, the old landing pages and anything left from a previous version of the site.
Step 2: remove or correct what is out of date
The single most important step. Old prices, discontinued services and superseded policies should be fixed before anything reads them.
Step 3: check the pages nobody visits
Low traffic pages are where stale information survives, precisely because nobody looks at them and nobody notices they are wrong.
Step 4: write down the answers that exist nowhere
Price ranges, coverage area, lead times and process, which owners answer by email constantly and have never put on the site.
Step 5: include the sales and support emails
The real answers, in the real words, which are frequently better written than anything on the website.
Step 6: state the boundaries explicitly
What it must not answer, what must escalate, and any claim it must never make, which is not something a website can imply.
Step 7: decide what to exclude from training
Blog posts making time bound claims, old campaign pages and anything aspirational rather than factual.
Step 8: handle the WooCommerce content separately
Product details, stock and shipping behave differently from page content, which is the ground in how to train a website assistant on product docs.
Step 9: test with the questions customers actually ask
From the inbox rather than from imagination, since the two lists differ considerably.
Step 10: check what it says about price
The most likely place for a wrong answer and the most damaging one, so test it specifically and repeatedly.
Step 11: watch the first fortnight closely
Reading conversations is where most of the improvement happens, and where the remaining stale content reveals itself.
Step 12: act on what it could not answer
Each gap becomes both a new answer and a page the website should have, which improves both at once.
Step 13: set a refresh routine
When prices, services or hours change, the site and the answers change together, or the whole thing drifts within a year.
Step 14: check what it says about anything regulated
Guarantees, safety, eligibility or anything with legal weight. A site can imply something loosely across three pages that becomes a single confident sentence when something answers from it.
Step 15: train it on a tidy site rather than a large one
An assistant trained on twenty accurate pages performs considerably better than one trained on forty containing six that are wrong, which makes the tidying the highest value part of the setup. Set it up through install the widget and train your bot, connect the store through integrations, and the general method sits in how to train your chatbot on your website content. The installation concerns are covered in adding an assistant without breaking the theme and page speed questions before adding anything. The sector view sits under small business.