GazeVideo Research

How GazeVideo Research Works

Every number on our research pages comes from the same pipeline: anonymous behavioral events, conservative definitions, aggressive cleaning, and publication only after a sample threshold is met. This page is the complete reference — read it before quoting us.

Core Definitions

Video ready. A conversation clock starts at the moment the remote video track first appears on screen — not when the match is requested, and not when the connection is negotiated. Time spent connecting is never counted as conversation time.

Skip. A skip is a deliberate switch: one person ends the current match and asks for someone new. A session that ends for other reasons — a connection failure, a closed tab — is recorded separately and never counted as a skip.

Session milestones. Each session is checked against fixed milestones measured from video ready: 5, 10, 15, 30, 60, and 180 seconds. A session either reaches a milestone or ends before it; there is no in-between state.

RTC Sampling

Calls run over the Agora RTC SDK. During a call, network statistics from the SDK’s stats interface are sampled at 5-second intervals and reduced to a per-session aggregate summary — for example the session’s median round-trip time. We store only that summary. No audio, no video, no frames, and no per-sample series are ever stored.

Latency buckets. For analysis, a session’s aggregate round-trip time is mapped into one of four buckets: under 100 ms, 100–200 ms, 200–300 ms, and 300 ms and above.

Survey Sampling

The in-product survey is shown to roughly 10% of eligible sessions, selected at random, and at most once per person within any rolling 30-day period. It is a single-choice question with fixed options; answering is voluntary, and no identifying information is attached to a response.

Exclusion Rules

A session is excluded from all studies if any of the following apply:

• The connection failed before video ready, or the remote video track never appeared
• The session duplicates another counted session from the same visitor
• The session ran in a background tab, where timers and visibility are unreliable
• The session came from internal testing or known automated (bot) traffic
• The recorded duration is impossible — negative, or longer than the session window allows

Excluded sessions are counted separately and reported alongside every published study, so readers can see how much data the cleaning rules removed.

Privacy Boundaries

GazeVideo research never records or screenshots conversations, and we do not perform face recognition, face analysis, or eye tracking of any kind. We collect no names, accounts, or contact details for research purposes, and published results are aggregates only — no individual session is ever identifiable in what we release. For how the product itself handles data, see the Privacy Policy.

Limitations

• Observational data. We measure what happened, not why it happened. No experiment assigns anyone to a condition.
• Correlation is not causation. When two metrics move together — latency and conversation length, for example — this research shows the association only.
• Single-site sample. All data comes from gazevideo.live. Patterns here may not generalize to other platforms, audiences, or formats.

Publication Rule

A study publishes only after it reaches its minimum sample size and the aggregate numbers pass operator review. Published snapshots are never edited in place — corrections or refreshes are released as new, clearly labeled snapshots.

Related: first 15 seconds study · latency vs conversation length · what matters survey · how random video matching works