Collect the 30 to 50 questions staff and customers actually ask, write one short page per topic with the answer first, give each page an owner and a review date, keep the pages in one place your AI tool can read, and test it with 25 real questions before anyone relies on it. Allow about a week, part-time.
The tool matters less than the pages. Almost any modern AI assistant can answer from documents you give it; what decides whether its answers are right is whether those documents are clear, current and free of contradictions. So most of what follows is about writing pages that AI quotes correctly. Hosting, testing and upkeep come after, with a tattoo studio as the running example.
What "AI can answer from it" means in practice
When you give an AI tool a set of documents, it doesn't memorise them. When a question arrives, it searches your pages for the passages that look most relevant, reads those, and writes an answer from them. This approach is called retrieval-augmented generation, or RAG; what RAG is, in plain English explains it properly. Three practical consequences shape everything below:
- Gaps get filled with guesses unless you tell the tool to say "I don't know". If your pages don't cover touch-ups, a question about touch-ups gets a plausible, invented answer.
- Contradictions become a coin toss. If one page says deposits are refundable with 72 hours' notice and another says 48, either may be quoted. At the tattoo studio used as the example below, the price sheet said $150 an hour and an older welcome pack said "from $140 an hour". Asked the same question on two different days, the assistant gave $150 once and $140 the next time, both times citing a real document. Neither answer was a malfunction; the pages disagreed.
- Buried answers get missed. A policy stated in paragraph nine of a rambling document may not be the passage the search picks up.
Harvest the questions people actually ask (two weeks, minutes a day)
Don't start by writing what you think people want to know. Start by collecting what they ask. For two weeks, keep a tally sheet by the phone and the front desk, and add to it from these sources:
- Your inbox and messages: search for question marks in customer emails and direct messages from the last three months.
- Staff questions to the owner or manager: "what do we do if…?" questions are knowledge-base gold.
- Complaints and misunderstandings: every "but I thought…" points to a missing or unclear page.
You'll end up with 30 to 50 distinct questions. Group them into topics; most small businesses land on 12 to 20 topics. Each topic becomes one page.
Write each question down the way it was asked, with a tally mark each time it recurs. A few lines of the studio's sheet after the first week (illustrative):
QUESTION (as asked) SOURCE TALLY
"how much is a deposit" DM |||| ||
"do I get my deposit back if I can't make it" email ||||
"can I bring my friend in with me" desk |||
"can my mum sit with me, I'm nervous" phone |
"is it free to get it touched up" DM |||
"what cream should I use after" DM ||||
"do you tattoo under 18s with a parent there" phone ||
"can you cover up an old tattoo" email ||The wording is worth keeping. "Can I bring my friend" and "can my mum sit with me" are one topic (who can come to the appointment), but the second is really a question about nerves, and a good page answers both. The under-18s question came in only twice, and it's the kind where a wrong answer is costly, so it gets a page of its own regardless of the count.
An AI assistant is handy for the grouping, as long as you check its choices:
Here are 46 questions customers and staff asked us, one per line.
Group them into topics for a knowledge base, 12 to 20 topics.
One topic per policy: if two questions are answered by different
rules, they belong in different topics. List the questions under
each topic heading.
ILLUSTRATIVE REPLY (EXCERPT)
Aftercare and healing
- what cream should I use after
- is it free to get it touched up
- how long before I can swim
Appointments
- can I bring my friend in with me
- can my mum sit with me, I'm nervous
- do you tattoo under 18s with a parent thereTwo corrections. Touch-ups aren't aftercare: they have their own rule (free within a set period, with conditions), so they need their own page, and grouping them under aftercare would have buried that rule in paragraph six. And the under-18s question is an ID and age policy, not an appointment detail. The prompt asked for one topic per policy; the model grouped by what the questions sound like instead, which is exactly the mistake to look for.
Topics the tally usually misses
A fortnight's tally catches the everyday questions and misses the occasional ones that cause the most trouble when answered wrongly. Before you start writing, check your list covers these:
- What happens when something goes wrong: complaints, refunds, redoing work, and who decides.
- Who to contact for what: bookings, billing, urgent problems, and the hours each is answered.
- Exceptions to normal hours: holidays, closures and what happens to bookings on those days.
- What the price includes and excludes: the single biggest source of "but I thought…" complaints.
- What you do with customers' information and photos: increasingly asked, and not something you want an AI improvising.
- Accessibility and practicalities: access, payment methods, what to bring.
If a topic isn't on your list and nobody has asked about it in two weeks, write a short page anyway when a wrong answer would be costly. Otherwise, leave it out and let the tool say it doesn't know.
Rules for pages an AI will quote correctly
Pages written for AI to answer from follow the same rules as good reference pages for people, applied strictly:
- Answer first. The first two sentences should answer the most common version of the question outright.
- One topic per page. "Deposits" and "Cancellations" are separate pages, even though they're related.
- Numbers and conditions in full. Not "a small deposit" but the amount; not "reasonable notice" but the number of hours.
- Exceptions spelt out. If there's an exception, say exactly when it applies and who decides.
- No cross-references. "As above" and "see the other document" mean nothing to a tool that only reads one passage. Repeat the key fact instead.
- Say what you don't do. "We don't tattoo over fresh sunburn" stops the AI inventing a policy.
- Headings phrased as questions, the way customers ask them, so the search finds the right section.
TOPIC: [e.g. Deposits]
Owner: [name] Last reviewed: [date] Next review: [date]
SHORT ANSWER
[Two sentences that answer the most common question outright.]
DETAILS
[Amounts, time limits, conditions. One fact per sentence.]
EXCEPTIONS
[When the normal rule doesn't apply, and who decides.]
WHAT WE DON'T DO
[Anything customers often assume we do.]
IF THE QUESTION ISN'T COVERED HERE
[e.g. "Say the front desk will confirm, and pass it to [name]."]
STAFF-ONLY NOTES (remove for customer-facing tools)
[Internal guidance, discretion, escalation.]A tattoo studio's deposit page, before and after
Take an illustrative tattoo studio with four artists and a front-desk manager. Here's how its deposit policy read in the old studio handbook:
Deposits are needed for most bookings and are usually $60 or so,
though bigger pieces may need more - speak to your artist. We can
sometimes move deposits if you give us enough notice, see
cancellations above.Every sentence in that paragraph invites a wrong answer: "most bookings", "or so", "bigger pieces", "sometimes", "enough notice" and "see cancellations above". Rewritten as a knowledge-base page:
TOPIC: Deposits
Owner: Front-desk manager Last reviewed: 2026-09-01 Next: 2026-12-01
SHORT ANSWER
Every booking needs a deposit before it's confirmed: $60 for pieces
up to 3 hours, and $120 for pieces over 3 hours or multi-session work.
The deposit comes off the final price.
DETAILS
Deposits are paid through the booking link; we don't take them by phone.
A booking is not confirmed until the deposit is paid.
You can move your appointment once, free, with at least 72 hours' notice.
With less than 72 hours' notice, or on a second move, the deposit is kept.
EXCEPTIONS
Illness or emergency: the studio manager may transfer the deposit once.
WHAT WE DON'T DO
We don't refund deposits as cash. We don't hold dates without a deposit.
IF THE QUESTION ISN'T COVERED HERE
Say the front desk will confirm within one working day.The staff-only section is where the studio's real discretion lives, kept out of anything customers will see. On the cancellations page it reads: "Manager may transfer a deposit a second time for a bereavement or hospital stay. Log it in the bookings notes with the reason. Artists don't make this call; send the client to the manager." When the pages are later loaded into a customer-facing tool, that section is removed first, so the tool can't offer an exception that a person is meant to decide.
It's longer, and every sentence can be quoted on its own without being wrong. That's the test for each line of every page. For the studio's other big topic, handling enquiries and deposits end to end, see how tattoo studios handle enquiries and deposits with AI.
Where the pages should live
Keep the master copies in your own documents, in one folder, whatever tool reads them. That way you can switch tools without rewriting anything; organising shared files so AI tools can use them covers the folder side. Then choose how the pages are served, based on who's asking:
| Who asks | Options (September 2026) | Worth knowing |
|---|---|---|
| Staff | A project in ChatGPT or Claude with the pages uploaded | Claude's free plan allows up to five projects; on paid plans, Claude switches a project to retrieval when its knowledge grows large, which lets it hold far more |
| Staff | Gemini Notebook, the new name for NotebookLM since July 2026 | Free accounts get up to 50 sources per notebook, each up to 500,000 words or 200MB |
| Staff | Copilot or Gemini searching a shared-drive folder | Answers from whatever the user can open, so keep the folder clean and current |
| Customers | A help-desk or chat tool with an AI agent, such as HubSpot's Customer Agent | HubSpot's needs a Professional or Enterprise subscription and uses 50 credits, about $0.50, per resolved conversation |
Put a number on the customer option before choosing it. If an illustrative 300 customer conversations a month were resolved by HubSpot's Customer Agent, that's 300 x 50 = 15,000 credits. Professional includes 3,000 credits a month, so about 12,000 would be extra: roughly $120 a month at $10 per 1,000, before any other feature that draws on the same credits. Credits you don't use in a month are lost; a quiet month can't be saved up for a busy one.
One option that's closing: OpenAI is retiring custom GPTs. Existing ones stop running on 11 December 2026, and OpenAI's migration turns each one into a plugin, with its instructions becoming a skill and its knowledge files becoming reference files. If your knowledge base currently sits in a GPT, move the pages into a ChatGPT or Claude project, or a Gemini Gem (becoming a skill from November 2026), before then. For a fuller comparison of hosting options and costs, see AI knowledge base options for small businesses compared, and if you choose Claude, setting up Claude Projects as a shared team assistant walks through the setup.
Whichever tool you use, add a standing instruction: "Answer only from the provided pages. If the answer isn't in them, say so and suggest asking [name]." It's the single most effective line against invented answers.
Upload the knowledge-base pages and nothing else. It's tempting to add the old handbook, a few email threads and last year's price list "for context", but every extra document is another chance of a contradiction, as the studio found when it ran its first test.
A 25-question test before anyone relies on it
Build a test sheet from your harvested questions: 20 that the pages should answer, worded the way people really asked them (typos and all), plus 5 that the pages deliberately don't cover. Ask each one and score it:
- Correct: right answer, nothing invented.
- Incomplete: right as far as it goes, but missed a condition.
- Wrong: any incorrect fact, especially a price, a time limit or a safety point.
- Correct refusal: said it didn't know, for one of the five uncovered questions.
For staff use, aim for at least 22 out of 25, with no wrong answers on money or safety. For anything customers see, the bar is zero wrong answers on prices, policies and health or safety, and all five refusals correct. A tool that confidently answers questions it has no page for isn't ready to face customers. Stopping an AI chatbot giving customers wrong answers covers what to change when it fails.
Scoring is where "incomplete" earns its place. One test question, copied as a client sent it: "hi can i get a touch up for free its been like 2 months". The assistant answered (illustrative): "Yes, touch-ups are free at our studio. Just get in touch with the front desk to book." The touch-up page actually says touch-ups are free within three months of the original session if the aftercare was followed, and chargeable after that. At two months the "yes" happened to be right, so it wasn't wrong, but it dropped both conditions, and the same answer to someone at five months would be. The fix was on the page, not the tool: the short answer now reads "Touch-ups are free within three months of your session, as long as you followed the aftercare guide", so the conditions travel with any quote of the first sentence.
Keeping every page owned and dated
Knowledge bases fail slowly. Prices change, a policy gets tweaked in a staff meeting, and six months later the AI is confidently quoting last year's rules. Four habits prevent it:
- Every page has a named owner, and only the owner edits it.
- Every change to a price or policy updates the page the same day. Make it part of the decision: "we're changing the deposit" isn't finished until the deposit page says so.
- Every page has a next-review date, quarterly for anything about money, twice a year for the rest.
- Re-run the same 25-question test every quarter. If the score drops, something has drifted.
The studio's second quarterly run shows what drift looks like. The score slipped from 24 to 21. Two of the three misses were the minimum charge: the owner had raised it at a staff meeting, the artists were quoting the new figure, and the pricing page still had the old one, so the tool contradicted the staff. The third was a new guest artist's booking rules, which nobody had written up. Neither problem would have shown until a customer pointed it out. Twenty minutes of updating two pages brought the score back to 24, and the owner added "update the page" to the agenda template for staff meetings.
How the tattoo studio's first version came together
Illustratively, the studio's two-week tally produced 46 questions, which grouped into 14 topics: deposits, cancellations, pricing guidance, the design process, cover-ups, touch-ups, aftercare, preparing for an appointment, ID checks, health questions, walk-ins, gift vouchers, guest artists and parking. Writing the 14 pages took the manager and owner about eight hours across a week, roughly 6,000 words in total.
The first 25-question test scored 19. Three wrong answers traced to two older handbook files that had also been uploaded and contradicted the new touch-up page; removing them fixed all three. Two more came from a missing page about ID checks. The retest scored 24, and the one miss was a question about a guest artist's availability, which the pages rightly didn't cover but the tool answered anyway. Tightening the "answer only from the pages" instruction fixed that too.
The front desk now uses it to answer phone questions in seconds, and new starters use it to learn the studio's policies in their first week. The studio hasn't put it in front of customers yet, and won't until it passes the stricter customer test two quarters running.
Questions about building the knowledge base
Can I just upload our existing documents instead of writing new pages?
You can start that way, and for internal use it's often good enough. The trouble is that existing documents contain old versions, contradictions and answers buried in long paragraphs, and the AI can't tell which parts are current. Upload them first, run the 25-question test, and rewrite only the topics that fail. That usually turns a week of writing into two or three days.
How long should each page be?
Long enough to answer every version of the question on that topic, and no longer: typically 150 to 500 words. If a page passes about 800 words, it probably covers two topics and should be split. Short, focused pages make it easier for the tool to find the right passage and easier for you to keep up to date.
Should staff and customers use the same knowledge base?
They can share the same source pages, but give customers a filtered set. Staff pages can include internal notes, such as when to make an exception or who approves a refund. Customer-facing tools should only see what you'd happily publish. Keeping the internal notes in a separate section or page makes the split simple.
Further reads
- How to Build the FAQ Your AI Chatbot Needs Before Launch — The customer-facing version of the question harvest, before a chatbot launch.
- How to Train an AI Chatbot on Your FAQs, Policies, and Prices — Loading these pages into a customer chatbot, step by step.
- Gemini Notebook (Formerly NotebookLM) for Small Business Teams — A closer look at one of the hosting options for a small team.
- Reduce Owner Dependency: Use AI to Capture What Only You Know — Use the same method to capture what only the owner knows.
- How to Turn Support Tickets Into Help Centre Articles With AI — Turn past enquiries straight into new knowledge-base pages.
- How to Test a Customer Chatbot Before It Goes Live — A fuller test plan once the pages face customers.
- Example AI Roadmap for a 12-Person Business, Month by Month — A 12-person driving school's first AI year, month by month: five workflows, about $30 a month in new tools, one idea postponed and why.
- How to Build a Shared Prompt Library for Your Team — An entry template, three sample prompts from a physiotherapy clinic, five storage options compared and a monthly routine that keeps the library useful.
- What Business Data Should You Start Collecting Now for AI? — Seven datasets worth capturing from today (enquiries, quotes, job actuals, questions, complaints, prices, feedback), with the fields that make them usable.
- How to Give AI Your Business Context So Answers Fit — Build a one-page business profile and task instructions, store them where each AI tool reads them, and test that answers fit your business.
- How to Document Your Processes Before Adding AI — How to write down a process so AI can follow it: capture methods, the seven elements to record, turning judgement into rules, and a template.
- Do You Need a Lot of Data to Use AI in Your Business? — How much data each common AI task needs, from three example emails to two years of sales history, and why scattered data is the bigger problem.
- AI Hallucinations Explained for Business Owners: Causes and Fixes — Why AI invents statistics, quotes and policies, where that hurts a business most, and two copy-ready prompts plus a checking routine that catch it.
- Does ChatGPT Remember Conversations? Memory Settings for Business — The three ways ChatGPT remembers, where the memory switches are in September 2026, and how to stop one client's details turning up in another's work.
- AI Glossary for Business Owners: 50 Terms in Plain English — Fifty AI terms in plain English, grouped by where you'll meet them, each with what it means for your decisions, plus five words vendors like to stretch.
- How Independent Travel Agents Use AI to Draft Itineraries Faster — Why 'plan me ten days' prompts are slow, and the brief-plus-block-library workflow that turns a two-hour first draft into roughly forty minutes.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Claude help centre on projects and retrieval for projects; Gemini Notebook (formerly NotebookLM) help pages on source limits and the July 2026 rename; OpenAI help centre on custom GPT retirement and migration to plugins; HubSpot Knowledge Base on Customer Agent availability; vendor pages for HubSpot credit prices. Checked September 2026.