Your Assistant relies on generative AI to answer based on your product knowledge and helpful content. To ensure reliability, it may sometimes choose not to respond. This guide explains how to use iAdvize analytics tools to identify these gaps and improve your Assistant’s performance.
1. Analyze responses
There are two complementary entry points to analyze your Assistant’s responses:
1.1 AI Shopping Assistant report
This report gives you an overview of recurring topics handled by your Assistant.
- Go to the Reporting tab > Advanced reports > AI Shopping Assistant
- Open the Recurring Topics table
In the “Response %” column, you’ll find your Assistant’s response rate.
The lower it is, the less your Assistant was able to provide an answer on that topic.
- Click the “Response %” column to sort responses in ascending order and identify priority topics that need to be enriched in your knowledge bases.
- Select a topic and click “Improve”
You’ll then see all unanswered questions from your Assistant for the selected topic.
Up to 5 main topics may be suggested. They are updated every two weeks.
To analyze these unanswered questions and understand their context:
- Click “Read” to open the related conversation
- Click “Improve answer”
- Then click “Analyze”
This will help you understand how knowledge was used by the Assistant and what led to this response.
- The Messages section provides information about the visitor’s message and your Assistant’s response.
- The No answer reason section explains why no answer was given (2. Understanding no-answer cases)
- The Knowledge record used section provides details about the knowledge your Assistant used to generate its response.
- The Ignored knowledge records section also provides information about other knowledge sources reviewed by your Assistant that did not contain relevant information to generate a response.
By reviewing this information, you’ll be able to enrich your knowledge base more effectively and provide your Assistant with richer content, enabling it to respond optimally to visitors.
1.2 Conversations report
You can also review each conversation individually to assess your Assistant’s response quality on a case-by-case basis.
- Go to the Reports tab > Conversations
- Add a filter “Answered by Shopping Assistant” > “True”
- Open any conversation you want.
- If your Assistant’s response isn’t satisfactory, click “Improve” as described above.
2. Understanding no-answer cases
The Shopping Assistant generates its responses in real time using:
- the latest artificial intelligence models and a general knowledge base
- the conversation context
- the data available in your knowledge bases (product catalog and helpful content)
This mode of operation gives it real flexibility: it can rephrase answers, adapt its tone, or suggest alternative wordings, as long as it stays faithful to the information available to it.
No-answer cases can be caused by:
- Missing information
Your Assistant couldn’t find any information in its knowledge base to answer the visitor.
- Detected hallucination
Your Assistant detected a hallucination risk and decided not to send its response.
This situation typically occurs in the following cases:
- The system detects that the generated response contains elements not found in the knowledge base. The response is therefore blocked to avoid hallucination.
- The knowledge base provided in iAdvize is incomplete, poorly structured, or lacks detail: the more information is missing, the higher the risk of errors.
- The content is unclear or ambiguous (e.g. vague or insufficiently factual information).
- Multiple sources are contradictory, making it difficult to select a coherent response.
- The Assistant didn’t identify any relevant information in the iAdvize knowledge base. It therefore attempted to rely on its general knowledge, but the response was blocked to prevent potentially incorrect information.
The best way to reduce these behaviors is to maintain knowledge bases that are as complete, clear, and well-structured as possible, and to continuously enrich them. The stronger your knowledge bases, the more reliable and relevant your Assistant’s answers will be.
3. Add knowledge
Manually added knowledge will therefore apply to all of your products.
If you want to edit your product catalog, go to the Product catalog tab.
- Click the “Add knowledge” button
A panel will open where you can define:
- The knowledge base where this Q&A should be added
- Topic: the question originally asked by the visitor. You can edit it as needed.
-
Content: the response the Assistant should provide. This will be added to your Helpful Content knowledge base and applied to all products. It does not modify your product catalog in any way.
- Click “Save”.
Product catalog
If the visitor’s question relates to a specific product and the Assistant’s response isn’t satisfactory, the “Add content” button will not be displayed. You must update your product catalog:
- If your catalog is automatically synchronized (URL, Shopify, etc.): update the content on your website. An automatic refresh occurs daily. You can still force an update by clicking “Refresh source” in Shopping Assistant > Knowledge:
- If your catalog is uploaded manually (e.g. CSV file): download the file, edit it, and upload it again with your changes. Make sure to disable or delete the previous version.
4. Why don’t I see the Improve button?
If it doesn’t appear, it’s probably because:
the response comes from a predefined conversation starter,
This is a pre-generated, cached response to a predefined question. It is not dynamically generated by AI at the time of the interaction.
→ In this case, no analysis or improvement is possible.
→ To modify these responses, you need to regenerate the starters (for example, after updating your product catalog).
or it comes from a custom behavior
This is a static response defined in a custom behavior, triggered by intent recognition.
→ Analysis is possible.
→ Improvements require editing the intent definition or associated content in AI Copilot > edit your Copilot in the Copilot Builder > Custom behaviors.