FAQ page examples
How to write FAQs your AI agent can answer from
Your FAQ page used to have one kind of reader: the customer who scrolled far enough to find it. If you use an AI support agent, it now has a second. Every time a customer asks a question, the agent searches your FAQs and writes its reply from what it finds.
That changes what a weak FAQ costs. A vague answer or a missing policy used to mean the customer gave up and contacted your team. Now the AI agent repeats the vague answer, or fills the gap itself. Gartner found that 61% of customer service leaders have a backlog of knowledge articles to edit (Gartner, 2024). In a separate survey, customers whose self-service attempt failed couldn’t find content relevant to their issue 43% of the time (Gartner, 2024).
Below are five rules for writing FAQs an AI agent can answer from, each with a weak and a good example, and a full returns section rewritten from scratch.
Where should your FAQ questions come from?
Start with your support tickets, chat transcripts and site search, because they show the questions customers actually ask and the words they use. Write an FAQ for each high-volume question that has a general answer, and add new ones when your team keeps typing the same reply by hand.
Sort your contacts by intent, start at the top, and copy the customer’s phrasing as you go. Then keep the list moving. If your team writes the same answer three times in a week, that’s your next FAQ. Here’s more on how to use your ticket data to decide which questions to answer.
Five rules for writing FAQs, with examples
These rules come from our work setting up DigitalGenius for ecommerce brands, and from seeing which FAQs the AI agent answered well and which it struggled with.
1. Organise questions by topic
Group your FAQs into the categories customers think in. It helps people scan the page and shows you where coverage is thin. A good starting set for ecommerce:
Orders and delivery
Returns and refunds
Products and sizing
Payments
Accounts
Contact and support
Weak: one “General questions” list of 40 items, in the order they were written.
Good: six labelled categories, each with five to ten focused questions.
Sizing questions come before a purchase, so it’s worth knowing how to train an AI agent on your FAQs and product content.
2. Write questions in the customer’s language
Use the words customers type, which are rarely the words in your policy documents. This is why it’s important to check the exact questions customers have been asking in your ticket backlog.
Weak: “Returns procedure”
Good: “How do I return an item?”
The question is part of what the AI agent matches against, so the customer’s wording makes a match more likely. Use everyday terms such as “refund”, “tracking” and “damaged item” in both the question and the answer.
3. Give one direct answer per question
Start with the answer, then add any conditions or steps.
Good:
When will I receive my refund?
After we receive and inspect your return, we’ll process your refund within 14 working days. Your bank may take additional time to show it in your account.
Weak:
Refunds
Refunds are handled in line with our returns policy. Please see our Terms and Conditions for details.
An AI agent can only repeat what the article says, and the weak version gives it nothing to work with. The industry advice is: write each answer so it makes sense if it’s quoted on its own, and avoid bare “yes” or “no” answers that depend on the question for their meaning.
4. Keep each FAQ focused
Give each question its own article. A single “Returns policy” page covering the returns window, exclusions, refund timing, exchanges and international returns answers five questions, each less well. Split it:
What is your returns policy?
How do I return an item?
When will I receive my refund?
Can I return a worn item?
Because each article is stored and matched as one unit, a focused article is a clean match for one question. A mixed article is a partial match for several, and the agent may answer from the wrong part of it.
5. Be precise
Include the details that change the answer:
Time limits
Eligibility and exclusions
Regions or countries
Costs
Required steps
Expected processing times
Weak: “Returns are usually accepted within a reasonable time.”
Good: “You can return unworn items within 30 days of delivery. UK returns are free. International returns cost £5.95, which we deduct from your refund.”
When an FAQ is vague, or two pages say different things, the AI has to fill the gap, and the brand owns whatever it says.
In the famous Air Canada lawsuit (Moffatt v Air Canada, 2024), the airline’s chatbot told a customer he could apply for a bereavement fare within 90 days, even after travelling. The airline’s own bereavement page said requests couldn’t be made once travel was complete. The tribunal held Air Canada responsible: “It should be obvious to Air Canada that it is responsible for all the information on its website.” In another case, Cursor’s support bot told a user that a subscription only worked on one device, a policy that didn’t exist.
Keep one version of each policy, with exact numbers, and remove older pages that contradict it. Where a policy decides money, such as refund eligibility, it’s also worth enforcing it in a workflow. Here’s why prompt-only AI agents get policies wrong.
FAQ page example: a returns section, before and after
Here, a catch-all returns article for an example clothing brand becomes four short FAQs, each answering one question in the customer’s words, with exact limits and costs.
Before:
Returns
We want you to love your order! If you’re not happy, you can return most items. Returns are accepted within a reasonable time frame and must be in their original condition. Some exclusions apply (see our T&Cs). Refunds are processed once we receive the item and will be with you soon. Different rules apply to international orders. To start a return, follow the steps in the image below.
[Screenshot of the returns portal]
It answers four questions at once, gives no time limit, exclusions or refund timeline, and puts the steps in an image the AI agent can’t easily read.
After:
What is your returns policy?
You can return unworn, unwashed items with the tags attached within 30 days of delivery. Swimwear, earrings and personalised items can’t be returned unless they’re faulty.
How do I return an item?
Go to our returns portal, enter your order number and email address, and select the items you’re returning. For UK orders, we’ll email you a free returns label to print and attach to your parcel.
When will I receive my refund?
After we receive and inspect your return, we’ll process your refund within 14 working days. Your bank may take additional time to show it in your account.
Can I return an item if I live outside the UK?
Yes. International returns cost £5.95, which we deduct from your refund. Send your parcel with a tracked service to the address shown in our returns portal.
Each answer now matches one question (rule 4), uses the customer’s wording (rule 2), starts with the answer (rule 3) and gives exact limits and costs (rule 5). The screenshot has gone because the steps are written out.
What should you leave out of an FAQ an AI agent reads?
Leave out anything the AI can’t read or can’t use on its own. Put the full answer in plain text, in one place.
Answers inside images, screenshots or PDFs. Check whether your AI agent can consistently read screenshots, but consider making the answer in plain text to avoid any failures.
“See our T&Cs”. Copy the relevant sentence into the FAQ.
Duplicate articles that disagree. Merge them into one and delete the old version.
Words that point backwards. Repeat the noun: “Sale items can’t be returned.”
Complex tables. Language models find tables harder to read, so turn the rows that matter into sentences.
Internal jargon. Spell out terms such as “RMA” or “DC”, or use the customer’s word.
Accordions and drop-downs are fine, as long as the answer is real text on the page.
How do you keep FAQs accurate once an AI agent is using them?
One way is to give every category an owner, update FAQs whenever a policy, carrier or price changes, and check the conversations where the AI agent couldn’t answer or got it wrong. Those gaps become your next FAQs.
This is where most knowledge bases fall behind: in the same Gartner survey, more than a third of service leaders had no formal process for revising outdated articles (Gartner, 2024). A simple routine covers most of it:
Give each FAQ category a named owner.
Update the FAQs as part of every policy change: returns windows, carriers, delivery cut-offs and prices.
Sort articles by date last updated and review the oldest each quarter.
Each week, read a sample of conversations where the AI agent couldn’t answer, or answered wrongly, and fix the article behind them.
You should also test your AI to check whether it is giving the correct answers.
What can an FAQ answer, and what needs order data?
An FAQ can answer policy questions such as “How long do refunds take?”. It can’t answer “Where is my refund?”, because that depends on one customer’s order, return and payment. Most ecommerce contacts are the second kind, so the FAQ is the foundation and order data does the rest.
“Where is my refund?” depends on whether the parcel was dropped off, whether the warehouse has received it and whether the payment has gone through. We’ve estimated that a bot answering from a knowledge base alone covers perhaps 10–20% of questions at best.
That’s why DigitalGenius uses your FAQs for the policy wording, while specialist agents for order status and returns check order, carrier and warehouse data and take the action, such as a refund or a replacement, based on your business rules. When the data doesn’t settle it, for example a damage claim with unclear evidence, the case goes to a person with the context attached.
The FAQ still matters in those replies: when the agent confirms a return has arrived, it uses the FAQ to explain when the refund will follow. It’s also the difference between answering a question and resolving it, which is what automation rate really measures.
Nobody’s Child is a good example. They started with returns, expanded across WISMO and FAQs, and automated 30% of tickets in two weeks. They’re now at 63%.
FAQ
Is an FAQ page the same as a knowledge base?
No. An FAQ page is a set of short questions and answers, usually about policies such as delivery, returns and payments. A knowledge base is the wider help centre, including how-to guides and product information. An AI support agent can use both, and the same rules apply to each.
How long should an FAQ answer be?
Long enough to be complete, and no longer. Most FAQ answers need two to four sentences: the direct answer first, then any conditions, limits or steps. If an answer needs several paragraphs, it’s probably covering more than one question.
Can an AI agent read FAQs in PDFs or images?
Usually not reliably. Most AI support agents work from text, so answers inside screenshots, scanned PDFs or image-only manuals are missed or only partly read. Put the full answer in plain text on the page.
How often should you update your FAQs?
Update an FAQ every time the policy behind it changes, such as a new returns window or carrier. Review the rest at least quarterly, and check weekly for questions the AI agent couldn’t answer.
Start with your top 20 questions
Your AI agent will be about as precise as your FAQs. That makes rewriting them one of the cheapest improvements you can make, before you launch an AI agent or after. Start with your 20 highest-volume questions. Split any combined articles, put the customer’s words in the titles, lead with the answer and add the exact numbers. Then read the conversations the agent couldn’t handle, and write the FAQ that was missing.
If you’d like to see how your FAQs and your order data can work together, take a look at the DigitalGenius AI agent for customer care or speak to the team.
