A chatbot doesn't know anything about your business until you tell it. Upload the right documents and it answers like your most experienced employee. Upload a messy pile of outdated PDFs and it answers like someone who skimmed the wrong file five minutes before their shift. The documents you feed it decide everything.
Most people rush this part, dumping every file they can find and hoping the AI sorts it out. It won't. A little care up front, choosing the right sources and prepping them well, is the difference between a bot customers trust and one they learn to ignore. Here's how to do it properly.
Start with the documents that answer real questions
Don't upload everything. Upload the material that answers what customers actually ask. Your internal strategy deck and your 2019 press release aren't helping anyone at 11pm.
Pull together the sources your team already reaches for when replying to customers:
- Help articles and FAQs
- Product descriptions, specs, and compatibility notes
- Policies: returns, shipping, warranty, privacy
- Pricing and plan details
- Setup and how-to guides
- Your saved replies and canned responses (these are gold, since they're already the answers)
If you kept an audit of your top customer questions, match documents to that list. Every common question should have a source that answers it clearly. Any question without a matching document is a gap the bot can't fill, no matter how clever it is.
Prep the files so the AI can actually read them
A chatbot learns from text. If your document is a scanned image, a locked PDF, or a design-heavy file where the words live inside graphics, the AI may read little or none of it. Garbage in, silence out.
Before you upload, run through this checklist:
- Is the file text-based, not a scanned image? Copy a sentence out of it. If you can't, the bot probably can't read it either.
- Is it current? Remove versions with old prices, discontinued products, or expired policies.
- Is it self-contained? A document that says "see the attached spreadsheet" leaves the bot guessing.
- Is it free of internal-only notes? Strip anything you wouldn't want quoted back to a customer.
- Does one answer live in one place? Conflicting duplicates are a top cause of wrong answers.
That last point trips up a lot of teams. If three documents state three different return windows because two were never updated, the bot might quote any of them. Pick the source of truth and delete the rest.
Choose your training method
Most chatbot tools give you a few ways to load knowledge, and they're not either-or. You'll usually mix them.
| Method | Best for | Watch out for |
|---|---|---|
| Website crawl | Getting broad coverage fast from existing pages | Picks up outdated or off-topic pages too |
| Document upload | Policies, specs, guides not on your site | Needs clean, text-based files |
| Q&A pairs | Your highest-volume, must-be-right answers | Manual to write, but the most precise |
A common and effective setup is to crawl your site for breadth, upload key documents for depth, and hand-write Q&A pairs for the handful of questions that absolutely must be answered a specific way. In SpideyChat you can do all three, then the bot draws on whichever source best matches the question.
A word on the website crawl, since it's the one that quietly causes trouble. Crawling is fast and covers a lot of ground, but it also sweeps up whatever's on your site, including that old landing page with last year's pricing and the blog post describing a product you discontinued. The bot can't tell which pages are current. Before you rely on a crawl, decide which pages should be in scope and exclude the rest, or you'll spend your first week of testing chasing down answers the bot pulled from pages you forgot existed.
Walk through a real setup
Take Fernwood Ceramics, a small studio that sells pottery and runs classes. Here's the order they trained their bot in:
- They crawled their website, which covered the shop pages, class schedule, and about page.
- They uploaded a PDF of their studio policies: kiln-firing turnaround, class cancellation rules, and care instructions.
- They wrote ten Q&A pairs for the questions that came up constantly: "do you ship internationally," "can I reschedule a class," "are the mugs dishwasher safe."
- They tested by asking each question a few different ways, including with typos.
- They fixed two answers where the bot pulled an outdated class price from an old page, then removed that page from the crawl.
Total time was an afternoon. The result was a bot that answered the shop's real questions accurately, because every common question traced back to a clean, current source.
Test before you trust
Never assume the upload worked just because it finished. Open the chat and interrogate it like a picky customer.
Ask your most common questions in the messy way people really type them. Ask the edge cases. Ask something the documents don't cover, and confirm the bot admits it doesn't know and offers a human, rather than inventing an answer. When it gets something wrong, trace it back: is the source missing, outdated, or contradicted by another file? Fix the document, not just the symptom.
This testing pass is where you catch the confident-but-wrong answers that would otherwise erode trust one customer at a time.
Keep it fresh, because documents rot
Uploading isn't a one-time task. The day your return policy changes, your bot is wrong until you update the source. Prices shift, products retire, processes change, and every one of those events can silently break an answer.
Build two small habits. First, whenever a policy, price, or process changes, update the source document and re-sync it, the same way you'd update a help page. Second, skim real transcripts regularly to spot answers that have drifted stale, then trace them back to the file that needs fixing. Twenty minutes a week keeps the whole thing honest.
A chatbot's intelligence is borrowed entirely from what you give it. Feed it clean, current, well-chosen documents and it becomes a genuinely useful member of your team. Feed it a junk drawer and it'll confidently pass that junk to your customers. Start with the sources that answer your real questions, prep them so the AI can read them, and keep them current. The quality of your answers will always trace straight back to the quality of what you uploaded.