In this guide · 8 sections
- 01Why content preparation matters more than the chatbot tool
- 02What to collect: the content checklist
- 03Clean it: remove contradictions and old versions
- 04Take out personal data before you upload anything
- 05Decide what the bot must not answer
- 06Set an owner and an update routine
- 07Do it yourself or get help?
- 08Frequently asked questions
Short answer
Collect your FAQs, prices, opening hours, service areas and policies into a few short, dated documents. Remove contradictions and outdated versions, take out personal data, write down what the bot must not answer, and name one person who updates the source whenever a price or policy changes.
Why content preparation matters more than the chatbot tool
Most website chatbots built on large language models use retrieval: when a visitor asks a question, the system looks up the most relevant passages in your documents and the model writes an answer from them. Our guide to RAG, or chatting with your documents explains the mechanics. The practical point is simple: the quality of the answers is capped by the quality of the passages it finds.
OWASP's list of top risks for LLM applications names misinformation as a core risk and recommends retrieving from "relevant and verified information from trusted" sources to reduce it. That puts the job on the business, not the software vendor: decide which source is the trusted one, and make sure it is correct.
What to collect: the content checklist
Start with what customers actually ask. Look through your last few months of emails, WhatsApp chats and enquiry forms and note the ten questions that come up most. Then gather the documents that answer them.
Gather before building
Format matters less than people think, but short, plain documents work best. One topic per document or per heading, with the answer near the top, makes it easier for retrieval to pick the right passage. Scanned PDFs of printed brochures often extract badly; a typed copy is safer.
Clean it: remove contradictions and old versions
Contradictions are the most common cause of a bot giving different answers to the same question. If the website says "free delivery above S$80" and a 2023 promotion PDF says S$60, the bot may quote either.
Put every fact in one master place
A shared document or a simple sheet: fact, current value, date it applies from, source document.
Search for duplicates
Prices, delivery thresholds, hours and refund days usually appear on several pages. Make them match, or remove the copies.
Exclude old versions from the upload
Archive past promotions, old price lists and superseded T&Cs. Do not give the bot a folder with "v1", "v2" and "final-final".
Date anything that changes
Write "Prices from 1 November 2026" in the document itself, so a reviewer can see when it is stale.
Write out the gaps
If customers often ask something your documents don't answer (parking, wheelchair access, lead times), add a short answer or tell the bot to hand over.
Take out personal data before you upload anything
It is easy to upload a folder that contains more than you meant: an order export, a quotation with a customer's address, or an email thread pasted into an FAQ. If the bot can read it, a clever question might make it repeat it. OWASP lists sensitive information disclosure as a top risk and recommends scrubbing or masking sensitive content and limiting what the model can access.
The PDPC's Advisory Guidelines on Use of Personal Data in Generative AI, issued 20 July 2026, encourage organisations to anonymise data as much as possible and practise data minimisation, and note that access and correction requests can extend to personal data held in a retrieval (RAG) database. The simplest way to stay out of that is not to put personal data in the knowledge base at all.
- Remove names, phone numbers, emails, addresses and NRIC or FIN numbers from examples and templates.
- Don't upload customer lists, order exports, CRM exports or HR documents to a public-facing bot.
- Replace real case examples with generic ones ("a 3-room flat", "a 10-person office").
- Check the vendor's settings for how long chat logs are kept and whether they are used for model training.
This is general information, not legal advice. Our guide to chatbot conversation logs and the PDPA covers what happens to the questions visitors type in.
Decide what the bot must not answer
Write a short list of topics the bot should refuse or hand to a person. This becomes part of the bot's instructions and saves embarrassing answers later.
| Question type | Why not let the bot answer | What it should do |
|---|---|---|
| Custom quotes | Price depends on details it can't see | Collect details and pass to sales |
| Complaints and refunds in dispute | Needs judgement and a record | Apologise, take contact details, hand over |
| Medical, legal or financial advice | Regulated, high-risk if wrong | Point to a qualified person |
| Anything not in your documents | The model may guess | Say it doesn't know and offer contact options |
| Order or account status | Needs a secure lookup, not documents | Link to the order page or a human |
Our guide to AI chatbots for Singapore websites covers choosing a tool and setting up hand-over to WhatsApp or email.
Set an owner and an update routine
Content goes stale quietly. A new price list goes on the website, but nobody updates the chatbot's source documents, and for weeks the bot quotes old prices.
- One owner: the person who already approves price and policy changes.
- A trigger: any change to prices, hours, policies or services includes "update chatbot source" as a step.
- A monthly check: ask the bot your top ten questions and compare answers with the master sheet.
- A log review: read the questions the bot couldn't answer and add content for the common ones.
Do it yourself or get help?
The content work is something only you can do well, because you know which price is current and which policy is real. Expect a few hours for a small business with one service line, longer if you have several outlets or product ranges.
Our AI Website Assistant (S$790) builds the assistant on your website, FAQ and documents, captures leads to your inbox, hands over to WhatsApp or email, and adds guard-rails so it only answers about your business, with a usage report after the first 30 days. Preparing content with the checklist above first makes the build faster and the answers better.
Frequently asked questions
At minimum: your common questions and answers, prices or price ranges, opening hours and locations, booking and refund policies, a description of what you do and don't offer, and how to reach a human. Short, typed documents with one topic per heading work better than long scanned brochures or a dump of every file you have.
Often yes, as a starting point, if your website is current and consistent. Many businesses find their site has outdated pages, conflicting prices or missing answers, so it is worth checking the pages the bot will read. Add a short FAQ document for questions customers ask that the website does not answer.
For a public website bot, avoid it. Those records contain personal data that the bot could repeat to other visitors. Write anonymised FAQs based on the questions instead. PDPC's July 2026 guidelines on generative AI encourage anonymisation and data minimisation, and access requests can reach personal data stored in a retrieval database.
Whenever a price, policy, service or opening hour changes, as part of the same change. On top of that, do a monthly check: ask the bot your top ten questions, compare the answers with your master sheet, and review the questions it could not answer to find gaps worth filling.
Usually because the source content contradicts itself, such as two price lists or an old promotion still in the upload, and the bot finds a different passage each time. Keep one current, dated version of each fact and remove older copies. Vague documents with no clear answer near the top also make answers less consistent.