How to Write a Shopify Help Center for an AI Support Agent

An AI agent is only as strict as the pages you give it. A 2,000-word 'Shipping' essay that hides the return window in paragraph nine will produce confident wrong answers.

Who this is for

Stores moving from a human-written FAQ to an agent that answers tickets 24/7.

The actual problem

Help centers are written for browsers. Agents need atomic rules: 'US returns are 30 days from delivery, not from purchase. Final-sale collection is excluded.'

If the page says 'usually' or 'in most cases', the agent will improvise. Improvisation is how you refund a custom cake.

The playbook

• Split Shipping, Returns, Exchanges, Damaged, Lost, Wholesale and Subscriptions into separate URLs.
• Put the number first: days, dollar caps, who pays return shipping.
• Name excluded collections and tags the same way they appear in Shopify.
• Add a one-line exception block at the top of each article, not a footnote.
• Remove 'email us' as the only next step. Give the agent an action instead.
• Rescrape after every policy change. Stale pages are worse than no pages.

Rules the agent should follow

• One decision per heading. Do not mix prepaid labels and customer-paid returns in the same list without labels.
• Date math: delivery date vs order date must be explicit.
• International and domestic never share a paragraph without saying which is which.

Mistakes that keep the queue full

• PDF policies the agent cannot fetch cleanly.
• Three overlapping articles that disagree on the window.
• Marketing language ('we always make it right') that overrides the written cap.

What to measure

Tickets the agent escalated with reason 'no matching article'. That list is your next writing queue.

FAQ

How many articles does an AI agent need? Start with the ten ticket types that are 70% of volume. Depth beats a 90-page encyclopedia.

Should I upload old tickets? Useful for tone and edge cases. Dangerous if old tickets contain exceptions you no longer honor. Curate them.

What about product pages? Yes for materials, dimensions and care — that is how you kill pre-purchase tickets. Not as a substitute for policy.

How does Relay ingest this? Point Relay at the live URLs. It uses those pages as the source of truth, then you add rules on top.

Feed Relay your policy URLs and watch which questions still have no article.