To create an FAQ is to answer questions people really ask, in the words they ask them in. The best source for that is your own sales, support and advisory conversations: that is where the phrasings, misunderstandings and follow-up questions live that your website currently misses. The path from there to a published question has four steps: collect questions verbatim with their trigger, cluster only the variants that hit the same concern, pick your clearest reviewed answer so far for each cluster, and maintain the entry with source, owner and review date. A copyable template separates short answer, detailed answer and evidence. Whether the FAQ belongs on the website, in WordPress, Shopify, SharePoint or an internal knowledge base is decided by the readership, and an internal FAQ may hold more than a public one. AI helps with searching, sorting and phrasing; which answer gets published remains a decision with an approval.
In my business the FAQ questions come out of the pattern register that Gustav distils from my approved calls: the recurring questions of my clients, each paired with my clearest human-reviewed answer so far. The FAQ is one of the four outputs of that register, next to the playbook, the objection library and raw material for content. How the register comes into being is in Call analysis with AI: the pattern register; which role takes on which task in my business is collected under AI employees. This guide starts at the question which entries from the register become an FAQ line.
Which questions belong in an FAQ?
Include a question when at least one of these signals applies:
- it appears in several conversations;
- the same concern is expressed in different words;
- your first answer regularly triggers a follow-up;
- you explain the topic unprompted in almost every call;
- the answer affects a purchase, product use or support workload.
Frequency alone does not decide. A rare question may matter if a wrong expectation causes serious harm. Conversely, a highly individual question does not automatically belong on a public page. The fourth signal is the underrated one: if you explain something in almost every call without being asked, your website has a gap there.
Create an FAQ from five calls: from raw sentence to published question
The following example is constructed so that the method can be shown without real conversation data. A fictional provider who sets up a first automation workflow for small businesses analyses five initial conversations. Calls, timestamps and quotes are demo data.
Step 1: collect raw sentences with their trigger
The quote stays verbatim, because you will need the phrasing as a heading later. The trigger stays with it, because it shows what the question was reacting to.
Step 2: cluster what needs the same answer
Four of the five sentences are about time, yet they become two entries and one open item. The quotes from calls 01, 03 and 05 belong together: behind all three sits the worry about effort on their own side, and the same answer helps all three. The quote from call 02 does not ask about effort but about a point in time; only a date helps there, so it is an entry of its own. The quote from call 04 concerns the time after handover and gets the status “open”, because the business answered it differently in every call. The test question is not “same word?” but “does the same answer help both?”.
Step 3: pick the clearest reviewed answer
For the first entry, three answers from three calls are available. In demo call 03, the answer “We start with one workflow that we get stable together before a second one is added” was followed by no further question, and the customer asked which workflow would be a good first one. That is the signal on which the entry gets this answer. Observation, not proof: a missing follow-up can also be politeness. That is why the field is called “clearest reviewed answer so far”.
Step 4: the finished entry
What changed from raw sentence to question: “for us” and “my team” are gone, because the question on the website has to apply to everyone. The variants are kept, because they are search language. The evidence stays in the internal register and does not go on the page.
A copyable FAQ template
Question in customer language:
Alternative phrasings:
Short answer in 1 or 2 sentences:
Detailed answer:
- requirement
- process
- boundary or exception
- next useful step
Source / evidence:
Expert owner:
Last reviewed:
Applies to:
Public, internal or confidential:
The short answer must stand on its own. The detailed answer explains conditions and limits. Marketing filler, invented numbers and guarantees belong in neither.
Internal FAQ or website FAQ?
Both come from the same source, but they have different readers, and the last field of the template decides which version goes where. This is how I divide it in my business:
The website FAQ is a subset of the internal one: the same question, the same reviewed answer, without evidence, without exceptions for individual cases and without statements that only apply to existing customers. What sits under “boundary or exception” in the internal entry is checked individually before publication: some of it belongs on the website because it prevents false expectations, some stays internal because it reveals negotiating room.
Where should the FAQ live?
An important SEO note: Google removed its FAQ rich result documentation on 15 June 2026, stating that the feature is no longer shown in Search results. An FAQ still helps when it answers readers' questions clearly. Do not build one for a search feature that has disappeared. Which structured data Google still shows instead is covered in Schema Markup Checker: Which Tool Tests What.
Use AI to create an FAQ without inventing answers
Do not give the AI only a topic. Give it approved entries with sources and clear rules:
- Use existing statements only.
- Flag contradictions instead of smoothing them over.
- State openly when the evidence is missing.
- Output short and detailed answers separately.
- Carry the owner and review date along.
The difference from an empty generator box lies in the input: a model that knows only the topic writes a plausible answer. A model that knows three pieces of evidence and your reviewed answer rephrases them and flags what is missing. A knowledge graph from calls can additionally show which FAQs depend on the same answer or source. When one answer changes, the affected pages become easier to find.
Privacy and the publication boundary
A public FAQ normally needs no personal data. Remove names, companies, places, rare case combinations and distinctive quotes. Pseudonymisation alone does not make a call excerpt anonymous. The legal basis, consent and purpose must already be clear before recording. More on that in the legal rules for transcribing conversations. This is not legal advice.
How Gustav turns this into a workflow
Gustav transcribes approved calls locally with Whisper; the raw transcript never leaves the machine. The pseudonymisation pass is the fixed first analysis stage, then he distils recurring questions with evidence in batches of ten to twenty calls, with a sample review after every batch. From the reviewed register come FAQ drafts in the template above; which of them go on the website is your decision at approval. His most important rule: he distils what you have said yourself and invents no answers on top of it. The same collection feeds sales objection handling and knowledge management with AI.
Frequently asked questions
How do I create an FAQ with AI without inventing answers?
By giving the AI evidenced statements instead of a topic. The input is the quotes from the conversations, your reviewed answer and the rule to phrase only from those and to flag gaps. You check the output against the evidence before it is published. Without that input, a model writes a plausible answer, and plausible is not the same as correct.
Is there an FAQ template I can copy?
Yes, the template above in this guide, with the question in customer language, variants, short answer, detailed answer in four parts, source, owner, review date, scope and publication level. It works in a text file, a spreadsheet or a CMS field. What matters is not the tool but that no field stays empty.
How do I collect customer questions systematically?
From the conversations you have anyway, with consent and verbatim. Three fields per hit: the quote, the trigger and the evidence. After each conversation only collect, do not cluster; clustering comes in the quarterly run. Emails and support tickets are a second source; they are already text and need no transcription.
How many questions belong in an FAQ?
As many as the selection rules above yield, no fixed number. An FAQ with five questions that are really asked helps more than one with thirty made-up ones. If you add questions only because the page would otherwise look short, you dilute the ones that count. Open entries from the register are candidates for later, not for now.
Is an FAQ still worth it now that Google no longer shows FAQ rich results?
Yes, because it is built for readers, not for a search feature. Google removed the documentation in June 2026 because the feature is no longer shown. A page that answers questions in the customers' language stays readable, linkable and useful for sales and support. I do not promise a visibility effect for it.
Do I need to ask customers before their question ends up in the FAQ?
Recording the conversation needs the prior consent of everyone involved; that comes before anything else. The published question itself no longer carries personal data: no name, no company, no distinctive quote, no rare case combination. Check every entry individually before you treat it as anonymous.
How to carry on
Take the last five conversations you have clean consent for and write out only the questions, verbatim, with the sentence before each. Strike out names, companies, places and amounts. If you then have three quotes that the same answer helps with, fill in the template once and decide at the last field whether the entry stays internal or goes on the website. You build the first entry yourself, and you are not alone while doing it: in the weekly live Q&A call in the community we go through the clustering together, including whether a short answer really stands on its own.
If you would rather not collect this by hand every quarter, Gustav is available as a ready-made package in my community, with the pseudonymisation pass, consent templates and a demo transcript to practise on. How you'd put Gustav to work in your business is something we'll look at together. Everything about it is under Community.
