Helping Robots Read

Why website chatbots give wrong answers

If you have built a widget that answers questions from your own content, you may have seen this happen.

• A question comes in.
• The answer exists on the site.
• The response is vague, incomplete, or incorrect.

In many cases, the problem is not the answer itself. It is how that answer is located.

Most modern website chatbots use retrieval-augmented generation (RAG), where the answer is generated from a small set of selected sections of your content. Chatbots that rely on retrieval can only work with the information they are given.

This article explains why that selection step fails more often than expected, and how small changes to wording can make content easier to retrieve and reuse. Those same changes often improve search visibility and reduce follow-up questions.

When relevant content is not selected, it does not matter how capable the language model is.

The missing information never reaches it.

You might recognise questions like:

How answers reach a chatbot

Most systems that answer questions from a website follow a similar process.

Content is divided into smaller sections. When a question is asked, the system searches those sections and selects a limited subset that appears relevant. Only that subset is passed to the language model to generate a response.

Each step narrows the available information. By the time the model produces an answer, it is working from a small slice of the original content.

This explains why an answer can be visible on the page but absent from the response. The full page is available to the reader. The system is working from selected fragments.

Why clear content is still overlooked

Many FAQs and help pages assume shared context.

• They rely on pronouns.
• They compress details.
• They assume surrounding headings will always be present.

This works when the page is read as a whole, however, it can become unreliable when a single paragraph is evaluated on its own.

Selection systems look for textual overlap between the question and the available content. When key terms are missing or implied, relevance becomes harder to establish. The system may select a different section that appears closer to the query, even if it contains less useful information.

Reviewing answers in isolation

A practical way to find these problems is to remove the surrounding context mentally.

Read an answer as if it were detached from its heading and page title. If the subject is unclear or the scope is ambiguous, the answer is likely to be missed during retrieval.

Making the subject explicit usually resolves this without changing the intent or length of the answer.

Writing answers that are easier to select

Most improvements come from straightforward edits.

• Replace vague references with the thing being discussed.
• Name the concepts involved using familiar terms.
• Avoid relying on implied meaning.

These changes reduce uncertainty. They also make answers usable outside their original page, which is how retrieval systems encounter them.

This approach can work extremely well for tools like Parrot Answers, which intentionally restrict what the model can answer and focuses on improving how content is written and selected rather than letting a chatbot improvise. It's designed for locked, reviewable answers rather than open-ended conversation.

Example: pricing

Example: cancellations and refunds

Example: limits and quotas

Example: access and visibility

Effects beyond chatbots

Clear questions and explicit answers tend to perform better in several contexts.

• Search engines can interpret pages more accurately.
• Snippets are less ambiguous.
• Support requests decrease because boundaries are clearer.
• Content remains usable as products and policies evolve.

What to avoid

Rewriting for retrieval does not require expanding answers unnecessarily.

• Adding large lists of related terms often makes content harder to read.
• Repeating names or concepts without purpose adds noise.
• Avoiding clear statements to hedge against edge cases usually creates more confusion.

Specific wording with defined boundaries is easier to retrieve and easier to rely on.

Applying this approach

Start with the pages that generate the most questions. Pricing, billing, limits, onboarding, and cancellations usually benefit first.

Review each answer on its own. Replace implied references with explicit ones. Name the concepts involved. Clarify behaviour where uncertainty leads to follow-up questions.

Those changes tend to improve retrieval reliability without altering the overall structure of the content.