What counts as a valid video session
A session counts when the video actually started, the session actually ended, and its length is physically plausible.
Concretely: a video_ready event fired, a closing signal exists, and measured duration is greater than zero and no more than four hours — anything longer is treated as an abandoned tab, not a conversation. Sessions from a backgrounded browser tab are excluded, as are sessions where the video connection failed. The goal is simple: a session in the data should mean two people were really on screen.
How conversation duration is measured
Duration is the time from video_ready to conversation_ended, observed on the client.
A lightweight client-side probe watches route changes and the video element itself to timestamp those two moments. It records timing and interaction events only. It does not analyze the video stream, and it performs no face recognition of any kind — the research measures how long conversations last, not who is in them.
How match speed is measured
Match speed is the time from match_requested to video_ready — tap to face on screen.
That window includes finding a partner and establishing the video connection, so it reflects the full wait a user actually experiences, not just the server's matching time. Impossibly fast values are treated as instrument noise and dropped during cleaning.
How the survey works
The survey is an optional, anonymous, multi-select question shown to a fixed fraction of longer conversations.
A deterministic sample of 5% of sessions that lasted at least 30 seconds is offered one question: what made the interaction feel natural? Respondents pick any options that apply — or close the prompt and answer nothing.
Responses are tied to an anonymous fingerprint built with a daily rotating salt, and each fingerprint is counted once per study window, so one person cannot move a rate by answering repeatedly. The survey stores option counts only: no free text, no identity attributes, nothing about who answered.
What gets excluded
Anything that is not a real person having a real conversation is removed before any metric is computed.
- Automated traffic, filtered by bot and user-agent signals
- Duplicate events from retries or double-firing clients
- Internal, staff and development traffic
- Impossible durations — zero, negative or over four hours
- Sessions run in a backgrounded browser tab
- Video sessions that never successfully connected
- Repeat survey answers from the same anonymous fingerprint
Excluded records are counted and reported alongside each study's valid sample, so the cleaning step is visible rather than silent.
Limitations
These studies describe what happens on EasyGirl. They do not explain why, and they do not prove causes.
The research is observational: it watches behavior rather than running controlled experiments, so correlation does not imply causation anywhere in these findings. Timing is measured on the client, which carries small clock and network errors. Samples are finite, and small subgroups are direction-only — where a group is too small to lean on, the study page says so. Results describe one platform and its users; they may not transfer to other products, other audiences or other time periods.
Privacy
The research runs on anonymous session data and collects nothing it does not need.
Sessions are identified by anonymous, short-lived tokens rather than accounts or permanent identities. The system records interaction timing and survey option choices — not message content, not video frames, not contact details. Survey fingerprints are anonymized with a daily rotating salt, so even the research pipeline cannot tie an answer back to a person the next day. Personal information the research does not require is simply not collected.