FAQ from customer questions: building new topics from customer language

Short answer

Building an FAQ from customer questions takes five steps: collect the questions that repeat in conversations, group them into sets that need the same answer, assign an owner for the correct answer in each group, have that owner write and approve the answer, and only then publish. The question's heading should stay close to the customer's own words; the answer should reflect the company's approved position.

The biggest risk is AI writing the answer itself. The question comes from the customer, but the answer is a company decision: price, warranty, returns rule, delivery time. A model can guess these, but a guessed answer in an FAQ becomes a promise to the customer.

Why FAQs fill up with the wrong questions

Most FAQ pages are written by the marketing team when the site launches and are hardly ever updated. The questions are the ones the company thought of: "Why choose us?", "When was the company founded?". Real customers ask completely different things: "can I change delivery from the office to my home?", "I lost the receipt — can I still return it?". This gap between the FAQ and real questions comes back to agents every day as extra work.

Step 1: collect repeating questions

If you already have a topic field, start from it; if not, use the reading method from customer inquiry analysis. An FAQ needs a slightly finer level than a topic list: not the topic "delivery" but the question "can I change the delivery address?". Pick the 3–5 most repeated concrete questions within each topic.

• Frequency: does the question come up at least several times a month?
• Stability: it is not tied to one campaign but repeats for months.
• Answerability: the question has one standard answer, not an individual one per customer.

Step 2: one answer, one question

The grouping rule is simple: questions that need the same answer are one FAQ item. "Can I pay by card at the door?" and "does the courier bring a card terminal?" are one item. "Can I pay by card?" and "can I pay in instalments?" are two, because the answers differ. Squeezing very different questions into one item makes the answer long and confusing.

Step 3: the answer's owner

Every FAQ item's answer should be written or approved by the person responsible for that area:

1. 1Price, discounts, paymentThe sales or finance lead.
2. 2Delivery, returns, warrantyThe operations or logistics lead; a lawyer if there is legal effect.
3. 3Product featuresThe product team or a technical specialist.
4. 4Personal data, contractsA lawyer and the person responsible for data protection.

Step 4: a heading in the customer's language

An FAQ heading should use the words customers use in search and in conversations, not internal terms. "How many days does my order take?" rather than "Logistics SLA". Real phrases taken from conversations are the best source for headings. The two or three most common phrasings for an item can also be used inside the answer, so customers recognise their own words.

Step 5: publish and measure

1. 1WhereWebsite FAQ, the bot's knowledge base, agents' answer templates — the same answer must be the same everywhere.
2. 2VersionRecord the answer's date and owner; when the rule changes, the FAQ changes too.
3. 3MeasurementA month later, check the question's share in conversations and the number of repeat questions.

The FAQ's effect may not show immediately: customers may keep writing without reading the page. But when the bot and agents give the same approved answer, conversations get shorter and contradictory answers decline.

When to update the FAQ

FAQ updates should be tied to signals, not to a calendar. Four signals are enough: a new question has entered the top ten of the topic list; the share of conversations about an existing item did not fall after publishing, so the answer is unclear; a price, rule or service has changed; the same phrase keeps appearing in "other". When any signal appears, the item's owner is notified and re-approves the answer. This turns the FAQ from a static page into a living document.

The invented-answer risk

AI can read conversations and write the FAQ answer too, and the text will look convincing. The problem is that the model does not know the company's rule; it only generalises what it saw in conversations. If agents gave different answers, the model writes their "average", which matches no official rule. Never publish an FAQ answer about price, timing, warranty or legal terms based only on AI-written text.

When an FAQ is not enough

• Individual questions: "where is my order?" is not answered by an FAQ; it needs access to the system.
• Complex choices: "which plan suits me?" needs an interactive conversation.
• Emotional situations: a complaining customer wants to talk to a person, not get an FAQ link.

The difference between what an FAQ page and a bot do is explained in chatbot vs FAQ page.

Illustrative example

This is an illustrative example. An optician's FAQ has 8 questions, none of which are among the 15 most asked in the last three months. The most repeated questions in conversations: "can I bring a prescription from another doctor?", "how many days until the lenses are ready?", "can I change the frame later?". Answer owners are assigned, 12 new items are written and added to both the website and the bot. The next month, the average length of conversations on these questions is tracked.

Typical mistakes

• Writing the FAQ once and never comparing it with conversations.
• "Cleaning up" headings into internal language — customers no longer recognise their question.
• Keeping different answers on the website, in the bot and in agent templates.
• Ownerless answers: the rule changes, the FAQ stays old.

What Vexvon provides for an FAQ

In Vexvon, question-and-answer pairs from conversations can be exported to Excel — a working draft for an FAQ and knowledge base. With a "question topic" field you can count the most-asked questions by period and read the conversations behind each value. Approved answers are loaded into the Vexvon knowledge base, from which the bot builds its replies. More: the knowledge base and analytics.

Next step

Put your current FAQ next to last month's 15 most-asked questions: how many match? If the gap is large, start the five-step process. We can do it on your own conversations in a demo.

Further reading on this topic: unanswered customer questions.