Context
Citations are how people check an AI-generated answer: they connect each claim to the source it came from. In Microsoft 365 Copilot experiences, responses arrive as a stream and are rendered in the client, and their citations travel with them. As part of my AI quality and reliability work, I worked on how those citations are handled, rendered and monitored.
Problem
A citation is only useful if it is present, points to the right source and is shown consistently. When sources appear inconsistently, are grouped confusingly or drop out of a response, people lose the ability to verify what they are reading, and trust in the answer goes with it. Because citations pass through several stages before they reach the screen, a problem seen in the client can start anywhere along the way.
My contribution
- Implemented and improved citation rendering, grouping and lifecycle behaviour in AI response experiences.
- Contributed to citation monitoring and telemetry, so that citation problems could be seen rather than only reported.
- Investigated how references were represented in responses and why sources sometimes appeared inconsistently.
- Made streaming and client integration improvements around how citations reach and render in the client.
Technical approach
The work treated a citation as something with a lifecycle rather than a static link: it is produced with the response, carried through the stream, resolved to a source, grouped with related references and finally rendered. Improving reliability meant looking at each of those steps, not only at the final rendering, and adding telemetry so that problems could be detected and traced to the stage where they started.
Challenges and trade-offs
Inconsistent citations are often intermittent: they appear in some responses and not others. Investigating them meant tracing how references were represented through the response rather than only changing how they were drawn.
Outcome
- Completed: the changes to citation rendering, grouping and lifecycle behaviour described above.
- Aim: supporting sources shown consistently and usefully, so people can check where each claim came from.
Reflection
Reliability for an AI response includes its evidence. A response that is right but cannot be checked is less useful than it looks, so the path from a source to the citation a person sees deserves the same care as the answer itself. The measurement side of response quality is in the Researcher case study.