What to Put in Your AI Knowledge Base (and What Quietly Ruins It)
Most advice about AI knowledge bases jumps straight to craft — how to phrase entries, how to structure them, how to think like a shopper. That is useful, but it buries the part that actually matters. There is a required half and a recommended half, and mixing them up makes the whole thing feel like more work than it is.
The two data paths
This is the whole model. Once it clicks, every "should I add this?" question answers itself.
| Question type | Where the answer comes from | Example |
|---|---|---|
| Products | Live catalogue | "How much is the canvas tote?" · "Do you have it in blue?" |
| How your store works | Knowledge Base | "What’s your return window?" · "Do you ship to Canada?" |
The catalogue path needs nothing from you. Prices, variants, stock and product descriptions come straight from Shopify on every single question. Change a price and the next shopper gets the new one. The Knowledge Base path is the half you own.
Required: get the information in there
This is the part that decides whether an answer comes back at all. If a fact is not in the Knowledge Base and not in your catalogue, the assistant will say it needs to check, or hand the conversation to you. Nothing else on this page matters until this is done.
What belongs here is durable knowledge about how your store operates:
- Returns and refunds — window, condition, who pays postage, what is excluded
- Shipping — destinations, timeframes, free-shipping thresholds, dispatch cutoffs
- Exchanges — how they work and how long they take
- Warranty and guarantees
- Sizing and fit guidance, care instructions
- Payment methods accepted
- Opening hours and physical locations
- Who you are and what you do
The fastest route is the import. One click pulls in your Shopify pages, blog articles and policy pages, and most stores already have eighty per cent of the list above written down somewhere. You can also add any URL, or write entries by hand for the things that live in your head and your inbox but were never written down.
Import what you already have
Your Shopify pages, blog articles and policy pages come across in one click, and policy pages stay tracked for changes afterwards.
Required: four things to keep out
This half is also required, because the wrong content does active harm rather than simply not helping.
Product data. Prices, stock status, which sizes or colours exist, specifications, which items match a request. All of it is read live. A saved copy does not help — it goes stale within days and then competes with the live lookup, giving your assistant two answers and no way to choose between them. A named product with a price is the clearest sign you are about to make this mistake.
Currently-running discount codes. Which codes are live and what they take off changes constantly and is read live. The rules around discounts, though — who qualifies, what cannot be combined, what is excluded — are exactly right for the Knowledge Base. Those outlive any individual code.
Anything about one customer. Where a specific parcel is, that someone was refunded, what is in a particular cart. Saving it puts one shopper’s details into an index that answers every other shopper.
Conversation with no facts in it. Greetings, "I need help", someone asking for a human. These create entries that match everything and answer nothing, and they degrade every other lookup on the way past.
Recommended: getting the best answers
Everything from here is optional. It is what separates an assistant that answers correctly from one that answers correctly every time, whatever words the shopper uses. Do it when you have a spare twenty minutes, or when you notice a question being missed.
One fact per entry. A long answer covering shipping and returns works better as two entries. Retrieval matches a shopper’s question against an entry, so the tighter the entry, the more reliably it matches.
Phrase it the way a shopper would ask. "Returns Policy — EU" is a filename. "Can I return something if I’ve opened it?" is what someone actually types. Imported pages keep their original headings, which is fine — this is worth doing for the entries you write yourself.
Keep the specifics. Timeframes, thresholds, conditions and exceptions are the entire value. "We have a generous returns policy" is worse than useless: it sounds like an answer while telling the shopper nothing.
Separate the exceptions. "Returns take 30 days" does not cover "sale items cannot be returned". Those are two facts, and a shopper who only gets the first will be annoyed later.
Strip personal details. If you are pasting from a real support reply, take out names, emails, order numbers and tracking numbers first.
Two entries that contradict each other are a decision, not a duplicate. If one says fourteen days and another says thirty, do not merge them — work out which is true and fix the wrong one. The assistant cannot referee your policy for you.
Recommended: teach it from real conversations
When you answer a shopper by hand, you can turn that answer into a Knowledge Base entry. It is rewritten into general form and stripped of personal details before saving, and you review it before it goes in.
Expect this to be selective. In our own testing, roughly six conversations in a hundred contain something worth keeping — the rest are product questions the catalogue already answers live. That selectivity is the point: it keeps your Knowledge Base full of things that answer future shoppers rather than filling it with chatter.
Keeping it from going stale
Policy pages imported from Shopify are tracked. Edit your refund policy in Shopify admin and the entry is flagged with an amber "Changed on your store" badge, with a Refresh button to pull in the new text.
Pages and blog articles are not tracked, because Shopify does not emit a notification when they change. If you edit a shipping page, nothing tells us — re-import it. Worth putting on the same calendar reminder as your other quarterly housekeeping.
How to test it
Ask your own questions in your own storefront widget, phrased carelessly. "do u ship to canada", "how long return", "can i swap size". Watch for two failure signals:
- "Let me check" or a handoff request — the entry is missing. This is the required half.
- A confidently wrong answer — you have a stale entry, or two that disagree. This is the dangerous one.
The second is why product data does not belong in here. A stale saved price sounds exactly as confident as a live one.
Write down what seems too obvious
The entries that pay for themselves are the ones you nearly skipped. If free shipping starts at $50, say what happens below $50. If returns are free, say who pays for an exchange. Shoppers ask the obvious questions constantly, and the entry that answers them takes one line.
A good rule: if you have ever answered it twice by hand, it belongs in the Knowledge Base.
Start with the required half
Import your policy pages, keep product data out, and test with five carelessly typed questions. The polish can wait.
Frequently asked questions
What is the minimum I need in my knowledge base?
Your returns, shipping, and any sizing or warranty information. Importing your existing Shopify policy pages covers most of it in one click, and that alone is enough for the assistant to answer policy questions correctly.
Should I add my product prices and stock to the knowledge base?
No. Prices, stock, variants and specifications are read live from your catalogue on every question. A saved copy goes stale and competes with the live lookup, which produces confidently wrong answers.
Why does the AI say "let me check" instead of answering?
The fact is not in the knowledge base and not in your catalogue. Add an entry covering it, then ask the same question again from your storefront to confirm.
Do I need to rewrite my policy pages before importing them?
No. Import them as they are — that is the required step and it works on its own. Rewriting entries into shopper phrasing is a recommended improvement you can make later if you notice questions being missed.
Related reading
How to Set Up Your AI Shopping Assistant Properly (15-Minute Guide)
Four steps, fifteen minutes. Including the one setting that causes almost every "the chat widget isn't showing" support ticket.
7 Customer Service Workflows Every Ecommerce Store Should Automate
"Add AI to support" is not a plan; a workflow is. The seven ecommerce support workflows worth automating first, each with its trigger, action, handoff rule and the metric it moves, and the order to roll them out.
How to Automate Customer Service on Shopify (Without Losing the Human Touch)
Audit the tickets, fix the answers behind them, connect the store data, then automate one workflow at a time. Seven steps, plus the escalation rules and handoffs that keep the conversations that matter human.