What Is QoE in Video Streaming, and Why It Matters

QoE (quality of experience) in video streaming is how a playback session feels to the viewer: how fast it starts, whether it stalls, and whether the picture stays steady. QoS measures how well the network delivers, and a fast network can still give your viewers a bad session.

TL;DR:

• QoE measures the viewer's experience at the player, and QoS measures network performance such as throughput, latency and packet loss. Good QoS makes good QoE possible but doesn't guarantee it.
• The standard model, ITU-T P.1203, predicts a 1 to 5 mean opinion score (MOS) for an adaptive streaming session from its video, audio and playback events.
• In a Columbia University study of 409,511 YouTube views, rebuffering raised abandonment six times as much as startup delay, and even a bitrate increase made viewers leave more often.
• AIOZ Stream's docs model QoE as perceived quality (VMAF) minus rebuffer and startup penalties plus stability, with starting targets of under 2 seconds to first frame and a rebuffer ratio under 0.5 to 1%.

Updated Oct 19, 2026: definition-first opening, question headings, a labelled QoE diagram, the exact study figures, what the public Analytics API exposes, and a References section.

What is QoE in video streaming?

QoE is the outcome a viewer notices, and it isn't a single number. AIOZ's QoE metrics docs put it plainly: "People remember whether the video started quickly, whether it froze, and whether the picture held steady without distracting jumps." If you run a video service, QoE is the score your viewers keep in their heads, whether or not you measure it yourself.

Picture a 4K stream that takes six seconds to start and stalls twice a minute. On a spec sheet its picture quality is excellent, but nobody would call that a good session. QoE exists as a separate discipline to catch exactly that gap between what the pipeline delivers and what the viewer sits through.

What is the difference between QoE and QoS?

QoS (quality of service) measures the network and the delivery pipeline, while QoE measures the result at the player. Mux's QoE guide sums up the split in two questions: QoS asks "is the network performing well?" and QoE asks "are viewers satisfied with their experience?"

The two are linked, since you can't have good QoE on a network that drops half its packets. The reverse doesn't hold, though, and your session can meet every QoS target and still feel bad. That happens when the player starts on too high a rendition, switches quality every few seconds, or waits too long for the first segment. That is why your monitoring needs player-side events and not only server logs.

How is QoE measured?

There are two common approaches: a standard model that predicts a viewer's score, and a set of player metrics you track directly.

The standard model is ITU-T Recommendation P.1203, built for HTTP adaptive streaming over longer sessions. It has three modules, one for video quality, one for audio quality and one that combines them with playback events into a session score. The output is a mean opinion score (MOS) from 1 (bad) to 5 (excellent). It predicts how a panel of real viewers would rate the session, so you don't need to run a lab test each time.

Most teams start with player metrics instead, because they're cheaper to collect and easier for you to act on. The AIOZ docs list six core metrics you can track:

• Startup time: time from pressing play to the first frame.
• Rebuffer ratio: total stall time divided by watch time.
• Average VMAF: a proxy for perceived picture quality across the segments the viewer watched. Our VMAF guide explains the score.
• Stability: time spent on the highest sustainable rung, and how often the player switches.
• Effective bitrate: what the viewer received over time, compared with the ladder you planned.
• Error rate: fetch or decode errors that interrupt playback.

Our stream health metrics guide shows how to calculate the first two from player events.

The QoE problems that make viewers leave

Not all of these metrics hurt your audience equally. Nam, Kim and Schulzrinne at Columbia University analyzed 409,511 YouTube views from more than 900 viewers in 102 countries, collected between April 2014 and July 2015. Three findings from their paper, "QoE Matters More Than QoS: Why People Stop Watching Cat Videos", matter for your priorities:

• Rebuffering raised abandonment six times as much as startup delay, and most of the startup delay they saw came from pre-roll ads.
• A single rebuffering event had three times the impact of a single bitrate change.
• Bitrate changes still hurt, even upward ones. When the bitrate changed during playback, viewers abandoned videos more than four times as often.

The last finding surprises most people, because a sharper picture sounds like good news. Viewers notice the switch itself, so a player that holds a steady rung often beats one that chases every gain in bandwidth. Our adaptive bitrate guide covers the rules players use to switch less.

The QoE model in AIOZ Stream's docs

AIOZ's docs describe QoE as a balance you can reason about: QoE ≈ perceived quality (VMAF) − rebuffer penalty − startup penalty + stability. The docs say it "isn't a rigid formula," and they use it to explain trade-offs. If stalls rise, the rebuffer penalty can outweigh a sharper picture, and a steady session at a modest rung can score well because stability counts.

Two sessions with the same average bitrate show how this works, and both start in 1.5 seconds. The first never stalls and holds one rendition, while the second stalls twice for three seconds in total and switches rendition six times. QoS dashboards would rate them about the same, but the rebuffer and stability terms score the second session far lower.

The docs also give starting targets, which they call "starting points, not hard limits":

• Startup to first frame under about 2 seconds for short videos, a little higher for long-form or TV.
• Rebuffer ratio under about 0.5 to 1.0%, watched at the 95th percentile and not only the average.
• Average VMAF high and steady, instead of chasing the maximum with frequent swings.
• Fewer switches and longer time on the highest sustainable rung.

When asked which metric to fix first, the docs say to start with startup and rebuffering, because viewers feel those immediately. That matches the Columbia findings above, which reached the same order from field data.

Why decentralized delivery changes QoE monitoring

AIOZ Stream delivers through AIOZ DePIN, a network of independently operated nodes, so which node serves a segment depends on where your viewer is. The QoE docs say AIOZ also measures "how locality and node availability influence startup and stall risk." A single global average can hide a region with weaker node coverage, so compare each of your regions or countries with its own baseline first and then with your global median.

How AIOZ Stream handles QoE

According to the QoE docs, the AIOZ Stream player emits events for play request, first frame, rendition change, stall start and end, and errors. AIOZ aggregates them into session KPIs that can be sliced by device type, app version, ISP and region, and the same signals feed ABR and ladder tuning.

The public Analytics API gives you engagement metrics rather than those QoE numbers. It exposes play, start, end, impression, watch_time, view and retention, broken down by country, device type, operating system or browser. It doesn't return startup time, rebuffer ratio or VMAF, and it has no ISP breakdown. You can still spot QoE trouble with it, for example when your play rate or watch time drops in one country. This request is quoted from the Analytics API docs and wasn't run for this article:

const result = await client.analytics.GetBreakdownMetrics('impression', 'country', {
  from: thirtyDaysAgo,
  to: now,
  limit: 25,
  offset: 0,
  sort_by: 'metric_value',
  order_by: 'desc'
})
// result.data?.data: [{ metric_value, dimension_value, emitted_at }]
// result.data?.total

The docs suggest a simple first test: upload a short video, enable analytics, and watch the first 100 sessions. Read QoE Metrics in the AIOZ Stream docs for the full model and targets.

FAQ

Is a higher average VMAF always better for QoE?
No. A high average with frequent quality swings can feel worse than a slightly lower, steadier picture, which is why stability is its own term in the model.

Does QoE apply to live streaming as well as VOD?
Yes. Startup time, rebuffering and switching affect both, although live viewers have less buffer to absorb a slow segment.

What is a good rebuffer ratio?
AIOZ's docs suggest under about 0.5 to 1.0% as a starting target, checked at the 95th percentile so a healthy average doesn't hide a bad tail of sessions.

Is MOS from ITU-T P.1203 the same as a VMAF score?
No. VMAF scores picture quality on a 0 to 100 scale, while a P.1203 MOS is a 1 to 5 session score that also includes audio, stalls and startup delay.

Can you improve QoE without re-encoding?
Often you can, because slow first-segment fetches, ladder gaps that are too wide and switch rules that react too fast can all be fixed at the delivery and player layers.

Which QoE metric should I fix first?
Start with startup time and rebuffer ratio, as both the AIOZ docs and the Columbia study suggest, then work on picture quality and stability.

Related reading

• Stream Health Metrics: Rebuffering Ratio and Startup Time: the startup and rebuffer terms in depth
• What Is Adaptive Bitrate (ABR) Streaming?: the switching logic behind the stability term
• What Is VMAF? How Netflix's Video Quality Score Works: the perceived-quality term
• Video Analytics Dashboard for Watch Time and Retention: building on the Analytics API
• What Is AIOZ DePIN? How Its Nodes Power Video Delivery: the delivery network behind regional QoE
• Glass-to-Glass Latency: What It Is and How to Measure It: delay in live streams

References

• AIOZ Stream Docs: QoE Metrics: the composite model, the six metrics, starting targets and player events
• AIOZ Stream Docs: Analytics API: available metrics, breakdowns and the request quoted above
• ITU-T Recommendation P.1203: the standard QoE model for adaptive streaming
• Nam, Kim and Schulzrinne: QoE Matters More Than QoS (Columbia University): the YouTube abandonment findings
• Mux: Quality of Experience (QoE) in Video Streaming: the QoS and QoE distinction