Most businesses already have the documents that would make their website assistant good. The price list the office manager keeps updated. The cancellation policy sent with every booking confirmation. The sheet of aftercare instructions handed out at the counter.
The trouble is that those documents were written for people who already have context. A customer holding a printed aftercare sheet knows which treatment it relates to. An assistant reading a file called "aftercare_v3_FINAL(2).pdf" does not. When you upload documents to a chatbot, the job is less about uploading and more about making each file readable by someone who has never met your business.
That preparation takes an afternoon, and it is the difference between an assistant that quotes the right price and one that quotes last year's.
How the assistant uses what you upload
Uploaded files become part of the assistant's knowledge base, alongside anything crawled from your website, your plain-language description of the business and any Q&A pairs. When a visitor asks something, the assistant looks for the most relevant passages and answers from them.
Two consequences follow. First, the assistant works with passages, not whole files, so a heading like "Prices" on page 7 is only helpful if the passage beneath it makes sense without pages 1 to 6. Second, it has no way to know which of two conflicting passages is current. It sees text, not your filing system.
If nothing relevant exists, it should say so and offer to take the visitor's details. That is good behaviour, but it only covers missing information. Wrong information sails straight through.
Four problems that ruin good documents
These are the issues that come up again and again when businesses gather their files.
- Scanned pages. A PDF made by scanning paper is a photograph of text. There are no words in it for the assistant to read. Export the original from Word or your design tool, or retype it.
- No dates. A price list without an effective date looks the same in 2026 as it did in 2023. If you have ever kept an old version "just in case", it is probably still somewhere in the pile.
- Internal shorthand. "Std clean (excl. ovens) £85 / w-end +20%" is clear to your team and ambiguous to everyone else. Is that plus 20% on Saturdays only? Bank holidays?
- Private material mixed in. Staff rotas at the back of a policy pack, a note saying "Mike can do 10% off for regulars", a supplier's trade price. Anything in the file can shape a public answer.
A price list, before and after
Here is a typical cleaning company price sheet, and the same content made ready for upload. The information is almost identical. The difference is that every line now stands up on its own.
| Before | After |
|---|---|
| File name: prices_new.pdf | File name: price-list-effective-1-september-2026.pdf |
| No date | "Prices effective from 1 September 2026. Replaces all earlier price lists." |
| Std clean (excl. ovens) £85 | Standard clean of a 3-bedroom house: £85. Oven cleaning is not included. |
| Oven +£40 | Oven cleaning, added to any clean: £40. |
| w-end +20% | Saturday and Sunday appointments cost 20% more. Bank holidays are not available. |
| Regulars ask Mike | (removed) |
| Min 2hrs | Minimum booking is 2 hours. |
| Scanned table image | Typed table with a heading row |
Notice the removed line. A discretionary discount that depends on a person's judgement is not something a website assistant should announce. If you want regulars to ask about loyalty pricing, write a Q&A pair saying the team can discuss it and offering to take their details.
Policies need a different kind of care
Price lists go wrong through age. Policies go wrong through vagueness and exceptions.
A cancellation policy that says "we may charge for late cancellations" leaves the assistant with nothing definite to say, so the visitor gets a vague answer and books with someone clearer. Write the rule as you actually apply it: how much notice, what the charge is, and what happens in genuine emergencies.
Exceptions are fine, but name them. "If you cancel with less than 24 hours' notice we charge 50% of the booking. We waive this for illness if you let us know before the appointment time." That is a sentence an assistant can repeat accurately.
If a policy touches health, money or legal rights, such as aftercare after a treatment or the terms of a deposit, add a line saying who to contact with a specific concern. The assistant should give general information from the document and point anything personal or urgent to a person. It is not there to give clinical, legal or financial advice.
Getting your files ready in one sitting
Work through this list with your documents open.
- Gather everything a customer might ask about. Prices, terms, policies, aftercare, preparation instructions, service descriptions.
- Delete duplicates and old versions. Keep one file per topic. If you are unsure which is current, that is a question to settle before uploading, not after.
- Add a date line at the top of each file. "Effective from" plus "replaces earlier versions".
- Rewrite shorthand as full sentences. Read each line as a stranger would.
- Remove anything private or discretionary.
- Check every file against your website. If the site says £80 and the document says £85, fix one of them today.
- Save in a text-based format. PDF exported from a document, DOCX, TXT, CSV or Markdown. For spreadsheets, export a clean CSV with clear column headings.
- Name files so a human would understand them.
Then upload, and ask the assistant five questions each document should answer. Check the replies against the files.
Keeping them honest after launch
Documents are the part of an assistant's knowledge most likely to go stale quietly, because nobody visits them. Websites at least get looked at. A good habit is to tie document updates to the moment the real thing changes: when you print a new price list for the counter, replace the uploaded one the same day, and delete the old file rather than adding the new one beside it.
Where information changes often, such as seasonal opening hours or a waiting list, a short Q&A pair is easier to edit than a document. Our guide to writing Q&A pairs that answer the way your team does covers that, and what the crawler reads from your website covers the other half of the knowledge base.
Where to start
Open the folder where your price list and terms live and pick the two documents customers ask about most. Apply the eight steps above to just those two, then upload them and test. That small start will show you how much rewriting your other files need.
The docs explain the upload screen itself, and there is a longer piece on keeping a chatbot's knowledge up to date once you have more than a handful of files to manage.