AI that answers from your website, not from imagination

A grounded AI answer is one generated only from material you supplied — your crawled website pages and your approved question-and-answer pairs — rather than from the model's general knowledge. The distinction is the entire safety story of putting AI in front of customers.

An ungrounded model asked "what's your returns window?" will produce a confident, plausible, invented answer. A grounded one either answers from your returns page or admits it does not know. For a shop, the second behaviour is worth far more than the first.

How the step works

The Answer from Knowledge Base step searches your own material for the customer's question, sends an answer built from what it finds, and then branches two ways:

• answered — it found relevant material and replied.
• unanswered — your knowledge base does not cover it.

That second branch is the important one. It is not an error path; it is the normal, expected outcome for anything you have not written down, and it is where you wire a human.

Build it

1. Go to Automations → Knowledge Base and create a collection.
2. Add your website URL and let it crawl. Point it at the pages that answer questions — shipping, returns, sizing, FAQs — not your whole shop.
3. Add question-and-answer pairs by hand for anything not on your site.
4. In a new automation, trigger on Customer Message Received.
5. Add an If Customer Sent condition limited to text, so images and voice notes go to a person instead.
6. Add Answer from Knowledge Base and pick your collection.
7. Wire the unanswered branch to Assign Chat.

What it will and will not know

Question	Can it answer?	Why
"What's your returns policy?"	Yes	It is on a page you crawled
"Do you ship to Ireland?"	Yes	If your shipping page says so
"How do I wash this?"	Yes	If care instructions are on the product page
"Is the blue one in stock in medium?"	No	Nothing connects it to live inventory
"Where is my order?"	No	That is customer-specific data, not published content
"Can you do 20% off?"	No	Not a question your website answers

Why grounding matters commercially, not just ethically

An invented answer is not a neutral failure. If a bot tells a customer they have 30 days to return something and your policy says 14, you have either a complaint or a refund you did not owe. Multiply by the volume automation exists to handle and the downside is larger than the labour saved.

There is a platform dimension too. Answers that annoy customers produce blocks and reports, which move your quality rating, which gates your messaging tier. A bot that confidently misleads people is a slow route to a capped account.

Getting good answers out of it

1. Crawl the pages that answer questions, not the whole site. A knowledge base full of product listings answers policy questions worse than one containing only your policy pages.
2. Write the FAQ pairs your website is missing. Anything you find yourself typing repeatedly in the inbox belongs here.
3. Mine your own chat history. The questions customers actually ask rarely match the ones your FAQ page anticipates.
4. Re-crawl when policies change. A stale crawl answers with last season's returns window, confidently.
5. Read the unanswered branch weekly. Every question that lands there is either a gap in your knowledge base or something that should route elsewhere. This is the single most useful feedback loop in the whole setup.

Where it sits relative to keywords

Keyword rules and knowledge-base answering are complements, not alternatives. Keywords are exact and predictable, so use them for the intents where you want a specific flow — returns, wholesale, order status. The knowledge base is the catch-all underneath, for the long tail of genuine questions.

The ordering that works: specific keyword flows first, knowledge base last, with no keyword condition on it. If you are maintaining an eighty-phrase keyword list for one intent, that intent belongs in the knowledge base instead — keyword triggers that actually match covers where the line falls.

Cost and the window

An AI answer sent to a customer who just messaged you is free text inside an open window, so Meta charges nothing for it under its pricing rules. What it does consume is AI usage on your plan, which is metered per answer. A knowledge base wired to answer every inbound message, including "thanks" and "ok", spends money on nothing — which is another reason for the text-only condition and for putting specific keyword flows above it.

Always leave the door open

Two exits to a human, minimum: the unanswered branch, and an explicit option on any follow-up you send. Customers tolerate a bot that cannot answer. They do not tolerate one they cannot get past. Handover mechanics are in handing a chat from bot to human.

What to put in the knowledge base

The instinct is to crawl the whole site. Resist it — a knowledge base full of product listings answers policy questions worse than one containing only the pages that answer questions.

Include	Leave out
Shipping and delivery page	Individual product listings
Returns and exchanges policy	Blog posts and news
Sizing and fit guides	Checkout and account pages
Care instructions	Anything with prices that change
FAQ page	Terms and privacy boilerplate
Hand-written Q&A for gaps	Restricted product lines

That last row matters if any part of your range is sensitive. A crawl of your whole site lets the bot describe products you deliberately kept out of your catalog — see what you cannot sell on WhatsApp.

Writing Q&A pairs that work

The hand-written pairs are where most of the value is, because they cover what your website never says out loud. Three rules:

1. Use the customer's words in the question, not yours. "Do you deliver to my area?" beats "Delivery coverage".
2. Answer completely in the pair. The answer is what gets sent, so "see our shipping page" is a wasted entry.
3. Write one pair per question, not one per topic. A pair answering five things answers none of them cleanly.

The best source is your own inbox. Anything you have typed more than three times this month belongs here.

The unanswered branch is your backlog

Read it weekly. Every question that lands there is one of three things, and each has a different action:

• A gap in your material — write the Q&A pair.
• Something that should route elsewhere — an order-status question needs a lookup, not a knowledge base. Put a keyword flow above it.
• Something only a person can answer — which is what the handover is for, and that is the system working correctly.

This is the single most useful feedback loop in the whole setup, and it is usually ignored because nothing forces you to look at it.

Keeping the cost sensible

The message itself is free — the customer wrote to you, so it is free text inside an open window under Meta's pricing rules. What is metered is the AI usage, per answer.

So a knowledge base wired to respond to every inbound message spends money answering "thanks" and "ok". Two guards fix it: restrict the flow to text messages with an If Customer Sent condition, and put your specific keyword flows above it so the catch-all only sees what the others did not handle.

When it is the wrong tool

Be clear about the ceiling. A grounded knowledge base answers questions your published material already answers. It cannot see stock, prices that change, a specific customer's order, or anything happening in your business today. Those need a lookup or a person.

Used inside that boundary it is genuinely good, and it degrades honestly — the unanswered branch fires instead of an invented answer, which is exactly the behaviour you want in front of customers.

Never let AI produce a number a customer is charged

One hard rule, and it applies to the AI step generally rather than to the knowledge base: do not use a language model to calculate a total, a deposit, a discount, a day count or a subtotal. Models are unreliable at arithmetic in a way that is easy to miss, because wrong answers look exactly as confident as right ones.

The measurement behind that rule is worth quoting: tested against its own stated pricing rules, a live pricing prompt scored 10 out of 27, and asked the same question three times in a row it answered 440, 467 and 440. Not merely wrong — differently wrong each time, which is the property that makes it impossible to test your way out of.

Use the Work Out a Number step for any arithmetic instead. It is exact and repeatable. The sensible division of labour is to let the AI extract the inputs from what the customer wrote — how many nights, which size, how many units — and let the calculation step do the sum.

Tuning the match

Two settings control how readily the knowledge base answers, and the defaults are reasonable, but they are worth knowing:

• How many passages it considers — a higher number gives it more material to work from, at the cost of pulling in less relevant text.
• The minimum match score — raise it and the bot answers less often but more accurately; lower it and it attempts more questions it should have handed over.

If customers complain the answers are vague, the fix is usually better source material rather than looser settings. If the unanswered branch almost never fires, the threshold is probably too low and the bot is guessing.