When an agent answers badly, the cause is nearly always one of four things, and they are worth checking in this order. Three of the four are about content rather than the AI itself.
1. The knowledge is missing or out of date
An agent can only answer from what is in your Content Library. Open it and confirm the topic is actually covered, and that what is there is current. A website synced six months ago answers with six month old prices.
This is the most common cause by a wide margin. Fix it by adding or re-syncing the source.
2. The content exists but is not shared with this agent
Content in your library is not automatically available to every app. Each item has sharing settings that decide which apps may use it. If an answer exists in your library but the agent never uses it, check that the item is shared with this agent rather than kept internal or limited to another app.
3. The role or instructions are unclear
If the agent has the right facts but presents them badly, refuses reasonable questions, or wanders off topic, the problem is behaviour rather than knowledge. Check the agent role, and any Advanced Mode instructions, including the Assistant Style field which shapes replies separately.
Also check Creativity. A high setting makes an agent more likely to go beyond your content.
4. The language is wrong
If replies arrive in an unexpected language, or read awkwardly, check the agent's own language setting. It is separate from your dashboard language, so changing what you see does not change what customers get.
How to test a fix
Use the Playground preview and ask the exact question that failed. Change one thing at a time. If you change the model, the instructions and the content together and the answer improves, you have learned nothing about which one mattered.