How to write help content your AI agent can actually use

A customer asks your AI agent how long they have to return something from your summer sale. It confidently says 14 days. You extended the window to 30 days in June, but an old FAQ page still says 14, and the AI found it first.

That isn't really an AI problem. It's a content problem, and it's the most common reason AI agents disappoint. The model is usually good enough. What it reads is not. This guide shows how to write and maintain help content that an AI agent can answer from correctly, with enough time left to get it right before the holiday peak.

Why AI agents give bad answers

When you look at the transcripts behind a weak AI answer, you almost always find one of these:

• Two pages disagree. The returns page says 30 days, the FAQ says 14, the terms say "a reasonable period".
• The answer is buried. The rule sits in paragraph six of the terms and conditions, surrounded by legal text.
• The answer is in an image. Shipping tables as screenshots, size charts as pictures. Many AI agents can't read them reliably.
• "It depends" without saying on what. "Delivery times may vary" doesn't help anyone.
• The answer only exists in people's heads. Agents know you can change the address until the parcel ships; nobody wrote it down.
• Old content never died. Last year's holiday hours, expired promotions and discontinued products are still on the site.

Fix these and most of your AI agent's mistakes disappear.

How AI agents read your content

You don't need to know the technical details, but one idea changes how you write. Most AI agents don't read your whole site for every question. They search your content for the passages that look most relevant, then write an answer from those passages.

That has three practical consequences:

1. Every article should make sense on its own. If a passage only works with the paragraph above it, the AI may get one without the other.
2. Headings and wording should match how customers ask. "Can I return a gift?" is easier to find than "Reverse logistics policy".
3. Contradictions are dangerous. If two passages disagree, the AI may pick either one.

Step 1: Take an inventory

List every place your AI agent can read: help center articles, website pages, policy pages, the FAQ, product pages, and any documents you uploaded. For each, note what it covers, when it was last updated and who owns it.

Then pull your top 30 customer questions from recent conversations and check, for each one, where the answer lives today. You'll find questions with no answer anywhere, and questions with three different answers.

Step 2: Write one answer per question

Use the same simple structure for every article:

Part	What it does	Example
Title as a question	Matches how customers ask	"How do I return a gift?"
The answer, first	One or two sentences that answer it	"You can return unused gifts within 30 days of delivery for an exchange or store credit."
Conditions	Who, what, when, where	"Holiday orders placed from November 1 to December 24 can be returned until January 31."
Steps	How to do it	"1. Open the returns page. 2. Enter the order number…"
Exceptions	What doesn't apply	"Personalised items and gift cards can't be returned."
Next step	What to do if this doesn't solve it	"If the item arrived damaged, contact us with a photo."

Write it the way your best agent would explain it in chat: plain, specific and polite. Avoid "please note", "kindly" and legal phrasing.

Step 3: Make every policy explicit

Vague words create vague answers. Replace them with facts.

Vague	Explicit
"Orders ship quickly"	"Orders placed before 1pm CET on a working day ship the same day"
"Delivery takes a few days"	"Standard delivery: 2–3 working days in Germany, 3–5 in the rest of the EU"
"Returns accepted within a reasonable time"	"Returns accepted within 30 days of delivery"
"Some items are excluded"	"Underwear, earrings and personalised items can't be returned"
"Contact us for international shipping"	"We ship to the EU, UK, US and Canada. We don't ship to other countries yet"

Put tables as real text, not images, and spell out regions, currencies and time zones.

Step 4: Write down what only your team knows

Ask your agents: "What do you answer from memory that isn't on the website?" The list is usually long. Typical examples:

• whether an order can be changed after it's placed, and until when;
• gift wrapping, gift notes and sending to a different address;
• how to combine discount codes;
• what happens if nobody is home for delivery;
• bulk or business orders.

Add each as a short Q&A pair. Most AI agents, Tidio's Lyro included, let you add these directly without publishing a new help article.

Step 5: Give product questions something to work with

In online stores, a large share of pre-sale questions are about products. The AI agent can only answer them if the product data is complete:

• materials, dimensions, weight and compatibility as text, not only in images;
• fit notes ("runs small, size up if between sizes");
• care instructions and what's in the box;
• a size guide written as text or a real table.

Step 6: Tell the AI what not to answer

Good content also sets boundaries. Write down, in your AI agent's instructions or guidance:

• topics it must hand to a person: refunds outside policy, damaged goods, legal or medical questions, complaints;
• things it must never promise: delivery dates beyond your stated times, discounts, exceptions;
• the tone you want: formal or casual, emojis or not.

If a wrong answer would cost you money or a customer, the AI shouldn't answer it alone.

Step 7: Test with real questions

Before switching the AI on for customers, take 50 real questions from recent conversations, including awkward ones, and ask them all. Score each answer:

Result	Meaning	What to do
Correct	Accurate and complete	Nothing
Partly correct	Missing a condition or step	Add the detail to the article
Wrong	Contradicts your policy	Find and fix the conflicting content
No answer, handed over	AI didn't know and passed it on	Add content if customers ask this often

Repeat the test after every round of fixes. When most answers are correct and the rest hand over cleanly, go live.

Step 8: Keep it current

Content goes stale the moment something changes. Build updating into how you work:

• Give every policy page an owner and a review date.
• Update content as part of any change: a new shipping provider, a new returns window or a new product line isn't done until the help content is.
• Prepare seasonal content in advance: holiday shipping cut-offs, holiday returns rules and opening hours, then remove them afterwards.
• Read AI transcripts weekly and turn every gap into a new Q&A or a better article.

Common mistakes

• Writing for search engines instead of customers. Keyword-stuffed articles confuse AI as much as people.
• One giant FAQ page. Split it into focused articles.
• Keeping old versions "just in case". Delete or redirect them.
• Leaving it to the AI vendor. No tool can write your policies for you.

A two-week content sprint before the peak

1. Days 1–2: export your top 30 questions and map where each answer lives.
2. Days 3–6: rewrite the articles for shipping, returns, exchanges and order changes using the structure above.
3. Days 7–8: collect what your team answers from memory and add it as Q&A pairs.
4. Days 9–10: remove outdated pages, old promotions and contradicting copies.
5. Days 11–14: test with 50 real questions, fix, and test again.

With the content in place, follow our guide on how to automate customer support to roll out the AI agent itself, and get ready for the post-holiday returns wave with our guide to handling returns and refunds.