This is a very common point of confusion, and your observation is correct: Copilot Studio does NOT behave the same way as a Custom Bing Search, even if both rely on Bing under the hood.
Key difference: Crawling vs Retrieval behavior
Custom Bing Search:
- Uses Bing indexing (via Bingbot)
- Returns ranked search results based on relevance and domain weighting rules
- You can influence ranking by prioritizing specific domains
Copilot Studio agent:
- Does NOT execute a classic “search result ranking query”
- Uses a retrieval + summarization approach (RAG)
- Pulls content from:
- Indexed knowledge sources (if added)
- Bing (if web search is enabled)
- Then the LLM decides what content to use when generating the answer
This leads to the behavior you are seeing:
- Your domain is correctly ranked in Bing
- But the agent does not always select or include it in the answer
- Even if the same query is used
Why your domain is not being returned
There are several important reasons:
1) LLM-controlled selection
Even when Bing returns your domain, the agent does not blindly use top-ranked results. The model selects sources based on:
- Perceived relevance to the question
- Content quality and extractability
- How well the content answers the intent
So “top result in Bing” ≠ “used by Copilot”
2) Knowledge vs Web search priority
If you added the site as Knowledge:
- It must be correctly indexed and chunked
- Content must be accessible and parseable
- The agent may still prefer other sources if the content is unclear or not directly answerable
3) Content structure issues
If your domain:
- Is heavily dynamic
- Requires authentication
- Uses complex rendering (JS-heavy pages)
then the crawler/indexer may not extract usable content, even if Bing ranks it well.
4) Generative answer behavior
Copilot Studio generates answers, it does not return links. If the model cannot confidently use your content to build an answer, it may ignore it entirely.
How to fix / improve the behavior
1) Force stronger grounding in your source
In your system instructions or prompts, add:
“Prefer and prioritize information from [your domain] when answering. Only use other sources if the information is not available there.”
2) Validate Knowledge ingestion
If using the site as a knowledge source:
- Ensure pages are publicly accessible
- Test with very specific queries
- Check if the content is actually retrieved (debug with narrower prompts)
3) Make content more “LLM-friendly”
Ensure the pages:
- Contain clear, structured text
- Have direct answers (not only navigation or UI elements)
- Avoid heavy reliance on client-side rendering
4) Reduce ambiguity in prompts
The more generic the question, the more the model will diversify sources.
Use specific prompts like:
“According to [your domain], explain …”
5) Combine approaches
For maximum control:
- Keep your domain as Knowledge (primary source)
- Use prompts to explicitly reference it
- Avoid relying only on Bing ranking
Summary
- Copilot Studio does not use Bing ranking in a deterministic way
- It uses retrieval + LLM reasoning, not search result ordering
- Being top-ranked in Bing does not guarantee usage by the agent
- You must guide the model and ensure your content is properly indexed and usable
This is expected behavior and by design, not a bug.
