Shopify's knowledge base now sits in your admin under Sales channels > Agentic > Knowledge base. It moved there on 26 October, and the standalone Knowledge Base app that some stores installed earlier is being retired in January 2027. So the setting is easy to find. The hard part is what you put in it, because those answers are what an AI shopping agent repeats to a customer who never visits your site, and a wrong shipping cutoff or a vague fit note costs real money.
Most of the stores we look at treat this as a settings task. Fifteen minutes, paste the FAQ, done. That produces a knowledge base that answers nothing an agent couldn't already guess, and hallucinated answers where it stays quiet.
Where the knowledge base lives in Shopify admin
Open your admin, go to Sales channels in the left nav, then Agentic. Knowledge base is inside it, alongside the other controls for how your store presents itself to AI shopping surfaces. If you don't see Agentic listed, check that your store is on a plan and region where it has rolled out, and that your account has the right permissions. Staff accounts with limited channel access often can't see it at all, which is usually why one person on the team swears it isn't there.
Entries are short question-and-answer pairs. You write the question the way a customer would ask it, and the answer the way you'd want a support rep to say it out loud.
Migrating from the standalone Knowledge Base app before it's removed
If you installed the separate app, your content isn't going to migrate itself the day the app disappears. Do this now rather than in December 2026:
- Export or copy every entry out of the app into a spreadsheet, with a column for the question, the answer, and the policy page it came from.
- Delete anything that's stale. Most stores have an answer about a 2020 shipping delay still sitting in there.
- Rewrite what survives against the structure further down this post, then paste into the admin knowledge base.
- Uninstall the app only after you've confirmed the entries render in the new location.
January 2027 sounds far away. It isn't, if your plan is to deal with it after the next two Black Fridays.
What information AI shopping agents pull from a store
Three sources, roughly in order of how much they trust each one.
Product data first: title, description, variant options, price, availability, images, and whatever structured attributes you've filled in. This is the spine. An agent comparing two pairs of running shoes is reading fields, not prose.
Then your knowledge base, for everything the product record can't carry. Shipping speed, returns window, exchange mechanics, warranty, care instructions, whether you ship to a PO box.
Then your public pages, if they're crawlable: policy pages, size guides, collection copy. The weakest signal, because an agent has to infer structure from a page built for human eyes. If your returns policy is a 900-word block with the exchange window buried in paragraph six, expect it to be summarised badly or skipped.
The practical consequence: anything you want said accurately should exist as a short, explicit answer somewhere an agent can read without interpreting a layout.
Knowledge base vs FAQ page: you need both, for different readers
An FAQ page is a page. It ranks, it absorbs long-tail search traffic, it lets a human skim six questions and find theirs, and it can carry links, images, a size chart, a video. The knowledge base is a structured answer store for machines, with no layout and no browsing.
They share facts, not wording. Write the policy once as the source of truth, then express it twice: a readable page for people, a tight pair of sentences for agents. When the two disagree, the customer gets one answer from an agent and a different one from your site, and support eats the difference.
Don't delete the FAQ page. It still earns search traffic, and that traffic converts at a rate agentic surfaces have yet to match. If you're rebuilding that side of things, our content and SEO work treats both as one project rather than two.
Writing shipping answers for agentic commerce
A shipping answer an agent can repeat without lying has five parts: the cutoff with a time zone, the carrier, transit time by destination, the cost or threshold, and the exceptions. Vague beats wrong, but specific beats both.
Weak: Orders usually ship in 1-2 business days and most customers receive them within a week.
Usable: Orders placed before 1 PM ET on a business day ship the same day from our Columbus, OH warehouse. UPS Ground is 1-4 business days to the lower 48. Free over $75 after discounts. Alaska, Hawaii and Puerto Rico ship USPS Priority, 3-7 business days, and are not eligible for free shipping. We ship to APO/FPO addresses via USPS only. We do not ship to PO boxes on UPS orders.
That second version prevents three support tickets and one chargeback. Write every shipping entry that way. And if you have more than one fulfillment location, say so, because a single transit-time promise breaks the moment an order routes from the other coast. We get the location nuance wrong on the first pass more often than we'd like, usually because the warehouse team never told anyone that West Coast orders stopped going out of Nevada.
Two US-specific entries worth adding: what happens between Thanksgiving and the carrier cutoff dates in December, and whether you charge sales tax in the customer's state. On tax, don't state a rate. Say that tax is calculated at checkout based on the shipping address and the states you're registered in. If you're running Shopify Tax or Avalara, the checkout is already correct; the knowledge base only needs to stop an agent from inventing a number.
Returns, and the arithmetic that justifies writing them properly
Take an apparel store doing 900 orders a month at a $68 average order value. Returns run 22%, so 198 returns a month. Reverse shipping on a prepaid label runs about $9, and inspection, repack and restock adds roughly $7. That's $16 a return, $3,168 a month, before you count the orders you lose entirely because a customer who can't get a straight answer about exchanges just doesn't buy.
Now say clearer sizing and exchange answers pull the return rate from 22% to 19%. That's 27 fewer returns. $432 saved on handling, plus 27 orders that stay sold: 27 × $68 = $1,836 in revenue, at 55% gross margin, $1,010 in retained margin. Call it $1,442 a month, around $17,300 a year, from writing maybe twenty answers properly.
Run your own numbers. The point is that this is a revenue line, not a settings chore.
The returns entry needs: the window in days and whether it counts from delivery or order date, who pays return shipping, condition requirements including tags and original packaging, what's final sale, how exchanges work and how long they take, and what happens to a refund on a discounted item. Say the uncomfortable parts out loud. An agent that tells a customer returns are free when they're $7.95 has created a dispute you will lose.
Sizing: the answer most stores never write down
Agents get asked "will this fit me" constantly, and a size chart image is worthless to them. Alt text on a JPEG of a table isn't data.
Write the fit answer as text. How the brand runs against standard US sizing, what to do if you're between sizes, the model's height and the size worn in the photos, and garment measurements for the sizes you actually sell. If a style runs small, put that in the product description and in a knowledge base entry, because the agent may only see one of them.
How agentic sales channels use product data
Attributes do the comparison work. Material, fit type, color family, sleeve length, heel height, capacity, compatibility, whatever your category hinges on. If those live only inside a paragraph of marketing copy, an agent filtering for "waterproof under $120" won't find you, and no knowledge base entry rescues that.
The fix is unglamorous: populate Shopify's standard product taxonomy, fill in variant options properly instead of stuffing size and color into the title, keep GTINs where you have them, and use metafields for the attributes your category needs. Clean attribute data serves two masters at once, because the same fields power your own collection filters and on-site search. If that side of the store is weak too, FilterPro is one way to get filters and AI search running off those attributes without a theme rebuild.
Keeping knowledge base answers in sync with your policy pages
Pick one source of truth. Usually it's the policy page, because that's what a customer can screenshot and what a card network looks at in a dispute. Everything else copies from it.
Then give it an owner and a cadence. Ours is simple: one person on the support side reviews every entry quarterly, plus a mandatory pass in the first week of November before the Black Friday traffic lands, and an immediate update any time shipping rates, carriers, the free-shipping threshold or the returns window changes. Put the review date in the spreadsheet column next to each entry. Entries older than six months get re-read, not assumed.
The failure mode is always the same. Marketing changes the free shipping threshold from $50 to $75 for margin reasons, updates the banner and the cart, and nobody touches the knowledge base. Six weeks later an agent is still quoting $50 and you're honoring it.
Getting a US store ready for AI-assisted shopping
Order of work, if you're starting today:
- Pull your last 500 support tickets and count the question types. The top fifteen are your first fifteen entries. Don't guess at what people ask.
- Rewrite shipping, returns, exchanges and sizing with real numbers and named exceptions.
- Audit product attributes for your top 50 SKUs by revenue. Fix the fields, not the copy.
- Reconcile the knowledge base against the policy pages line by line, and fix whichever one is wrong.
- Move anything still sitting in the standalone app, then uninstall it.
Accessibility deserves a mention here, since the same structural discipline helps: real text instead of images of text, proper headings, tables marked up as tables. Demand letters over inaccessible storefronts are a live risk in the US, and a size chart trapped in a JPEG fails screen readers and agents for the same reason.
If you want someone to go through the ticket log and the policy pages with you and come back with the entries drafted, that's a half-day of work and we offer a free store audit to scope it. Merchants running larger US catalogues can also see how we work on the US team page.