In short: An AI chatbot that cites its sources tells the customer which document each answer is based on. That helps customers trust the reply, lets your team check answers quickly and reduces the risk of the bot drifting outside your knowledge base and making things up. In Puhera, the bot speaks only from your own sources and shows the source behind every answer.
What is a cited answer?
A citation is a note attached to the bot's reply that shows where the information came from. If a visitor asks about the return window, the bot replies "You can return this product within 30 days" and adds something like "Source: Return policy" underneath.
Behind this sits an approach called grounding. Instead of answering from general internet knowledge or the model's own best guess, the bot first finds the relevant passage in your knowledge base and then writes its answer based on that passage. The citation is the visible part of that process: the bot shows what it relied on.
How do citations build customer trust?
A customer chatting with a bot usually cannot tell whether a reply reflects the company's official information or a machine's guess. On topics involving money, delivery or returns, that uncertainty often pushes people to send an e-mail or call just to confirm.
- Transparency: Seeing that a reply comes from the business's own policy tells the customer the answer is official.
- A way to verify: The customer knows they can look at the source and read the details themselves.
- Consistency: Answering the same question from the same source every time reinforces the sense that the brand speaks with one voice.
Trust is a precondition for a chatbot to be useful at all. A customer who does not believe the answer will contact your support team anyway, however fast the bot was.
How do citations make answers easier to check?
Citations help the team running the bot as much as they help customers. Figuring out why an answer is wrong is far quicker when you know which document it came from.
- Spot the wrong answer. You see an incorrect reply during testing or in a live conversation.
- Look at the source. Which document did the bot base this on?
- Diagnose the problem. Is the document outdated? Do two documents contradict each other? Was the passage unrelated to the question?
- Fix the source. Update the old file, remove the contradiction or add the missing information.
Without citations, this loop turns into guesswork. It is hard to fix content when you do not know which content caused the problem, and the same mistake can keep coming back. To set your documents up properly from the start, read our guide on how to build a chatbot knowledge base.
How does grounding answers in your own documents reduce made-up answers?
Large language models are excellent at producing fluent, convincing text. The catch is that they can also produce a plausible-sounding answer when they do not actually know it. This is often called hallucination: a confident reply with nothing behind it.
In a customer service bot, a made-up answer is a real problem. Confirming a promotion that does not exist, quoting the wrong delivery time or describing a warranty you do not offer puts both the customer and your team in an awkward spot.
Grounding answers in your own documents reduces that risk in three ways:
- It narrows the scope: The bot is steered toward content you have approved instead of its general knowledge.
- It makes claims verifiable: Because every answer has a source, an unsupported claim is easier to notice.
- It exposes gaps: Questions with no match in the knowledge base reveal missing content, which you can then add.
To be honest about it, grounding does not eliminate the risk entirely. It reduces it and makes it manageable. That is why citations work best alongside a well-maintained knowledge base, regular testing and a human to step in when needed.
How should your sources look to get the most out of citations?
Whatever the bot displays, the source the customer sees is your document. So the value of a citation depends on how clear your documents are. A few principles make a real difference:
- Use meaningful file and page names. "Shipping and delivery terms" means something to a customer when it appears as a source; "doc_final_v3.pdf" does not.
- Keep each topic in one place. If the same fact appears in several documents in different forms, the bot may cite a different source each time, which undermines consistency.
- Make freshness visible. Stating an effective date on policy pages helps both customers and your team see how current the information is.
- Write short and clear. When the answer is not buried in the middle of a long paragraph, the bot finds the right section more easily and cites the right source.
Good citations start with well-written sources. The bot and your documents work as one system, and improving one raises the value of the other.
What can citations not solve on their own?
Citations are a strong tool, not a magic fix. Knowing their limits helps you set the right expectations.
- A wrong source produces a wrong answer. If an old price list is still in the knowledge base, the bot will quote that price and cite it. Keeping sources accurate is your job.
- Questions outside the knowledge base go unanswered. That beats making something up, but you need a next step so the customer is not left stranded.
- Personal questions need different data. "Where is my order?" is answered through a connection to your order system, not from general documents.
When the bot cannot answer, handing the visitor to a person is the safest move. Our live chat handoff page explains how that works.
How do citations work in Puhera?
In Puhera, your content is the bot's source of knowledge. The sources you add to the AI knowledge base are:
- Your website pages, crawled automatically
- PDF, Word, PowerPoint and text files you upload
- Your FAQ / knowledge base content
- Your product catalog, synced from a Google Merchant feed or T-Soft
When the bot answers a question, it draws on these sources and shows the source alongside the reply. A returns answer points to your return policy page; a stock answer points to your product catalog. Customers see where the answer came from, and when you test in preview you can see exactly which document produced which reply.
For questions your knowledge base does not cover, or moments that need a human touch, the Pro plan lets you connect the visitor to a live agent. All of this works without code. If you want the bigger picture of what a chatbot does in online retail, start with what an e-commerce chatbot is.
Frequently asked questions
Can a chatbot that cites sources still give a wrong answer?
Yes, because an answer is only as accurate as the source it relies on. Citations show you which document a wrong answer came from, so you can fix the source and stop the mistake from repeating.
Does every Puhera answer show a source?
Yes. In Puhera the bot answers from your knowledge base and shows the source behind every answer.
What does grounding mean for an AI chatbot?
Grounding means the AI bases its answer on specific documents rather than its general knowledge. The bot first finds the relevant passage in your knowledge base, then writes its reply from that passage.