Watch time is not one number — the system cards define it eight ways

You posted a reel. You check the insights at 24 hours. Average watch time is low. So you go back and trim the first three seconds. Cut the dead space. Make it hook faster. Everyone says watch time is what matters most.

Adam Mosseri says it too. The head of Instagram called it out by name. In January 2025 he said watch time is one of the top three signals for ranking. And the advice followed. Optimize for watch time. Make people watch longer. That is what the algorithm wants.

We went looking for what watch time actually means. In the documents Meta publishes about how Instagram works. What we found is not one definition. It is at least eight. Different thresholds on different surfaces. Some count seconds. Some count percentages. One compares you to other creators. The thing everyone calls watch time is not one metric. It is a bundle. And the bundle changes depending on where your post shows up. Instagram already scores a carousel and a reel on different predictions. Watch time is where that splitting goes furthest.

What follows is a reading of Meta's own system cards. The pages it publishes to explain how each part of Instagram decides what to show you. We also read what Mosseri said. Both things are true at once. He names watch time as a top signal. And the cards describe it in ways that do not agree with each other. That gap is the subject of this post. Every source below is linked, dated and quoted. One section shows where the cards contradict Mosseri's own statements.

Mosseri calls it the top signal. The cards call it eight different things.

Start with what Mosseri said. On 22 January 2025, he posted a video to Instagram. Social Media Today reported it the same day. His words:

The top three signals that matter most for ranking are watch time, likes and sends. So when looking at your insights, pay close attention to average watch time, likes per reach, and sends per reach. — Adam Mosseri, quoted in Social Media Today, 22 January 2025

Watch time. Singular. As if it is one thing you can measure and improve.

Then we read the system cards. Meta publishes one for each part of Instagram. Each card lists what the software tries to guess about you. Those guesses are called predictions. And each prediction runs on facts the system can read. Those facts are called signals. The cards list both.

Three surfaces matter here. Your feed is one. The Explore tab is another. And Reels Chaining is the third. Chaining is Meta's word for the run of reels that keeps arriving once you open one. Each surface has its own card. We read all three on 17 August 2026. That is the date Meta's site stamped them.

Every card defines watch time differently. Not one way. Many ways. Different thresholds, different calculations, different comparisons.

Reels Chaining counts eight different ways of watching

The Reels Chaining card lists nine predictions. Three of them are about how long you watch. Each one draws its line in a different place.

The clearest one is a threshold. The system guesses how likely you are to watch less than three seconds of a reel. That is from the card's own list. A signal feeding that guess is blunt: "How many times the post has been skipped within two seconds of being opened."

So two seconds is one threshold. Three seconds is another. Both show up on the same card as different measures of the same behavior.

Then comes the comparison. The card says the system guesses "How likely you are to watch more of a reel than 95% of users who watched reels of the same length."

That is not a fixed number of seconds. It is a rank. Your watch time set against other people's. On reels of the same length. So a 15-second reel and a 90-second reel are not compared to each other. They each have their own distribution. And 95% is the threshold inside that group.

The signals feeding that prediction are just two. One is the length of the reel. The other is "How much time people have spent viewing the reel." No percentage. Just time. In seconds, we assume. The card does not say.

Other signals on the same card name other thresholds. "How many times users watched at least 3 seconds of the reel anywhere on Instagram." Three seconds again. But now it counts across the whole app. Not just in Reels Chaining.

Another: "Whether or not you've been viewing reels for more than 15 seconds." Fifteen seconds. Not of one reel. Of reels in general. It measures whether you are in a long session.

Count them. Two seconds. Three seconds. Fifteen seconds. Ninety-five percent of same-length reels. Absolute time in seconds. And one more, under a different prediction: "How much time has viewers from your country spent on this reel." That is aggregated by country. Not by you.

Six different ways of measuring watch time. On one card. For one surface.

Explore and Feed add their own thresholds

The Explore card adds two more. It guesses "How likely you are to spend more than five seconds viewing a post." Five seconds. A new threshold. And it guesses "How likely you are to watch more than 95% of a video." Ninety-five percent. A fixed completion rate. Not a comparison to other users. Just a percentage of the video itself.

The signal feeding that second prediction makes it stranger. "How many people have watched more than 95% of the video." That counts other people. Their completion rate becomes a signal the system reads when it guesses yours. You are not scored on your own 95%. You are scored on how many other people hit 95%. At least in part.

The Feed card names a ten-second threshold. It guesses "How likely you are to spend more than 10 seconds on the first post." Another signal says "Whether that viewing time reaches 10 seconds."

And one more: "Whether you saw the first post for at least 3 seconds." Three seconds is back. This time for Feed. Not Reels.

The Feed card also names something new. "Which Reels-watch-time group you fall into." That is a category. A group. The card does not say how many groups there are. Or where the lines are drawn. It just says you are in one. And that group is a signal.

So now the count is eight. Two seconds. Three seconds. Five seconds. Ten seconds. Fifteen seconds. Ninety-five percent completion. Ninety-five percent compared to same-length reels. And a group assignment based on your reel-watching history.

Mosseri says it is two things at once

One month after naming watch time as a top signal, Mosseri clarified what it measures. On 25 February 2025, he posted again. This time about longer videos. Social Media Today reported it:

We don't want to penalize longer videos, which is why we look at not only the percentage of a video that was watched, but also the number of seconds. — Adam Mosseri, quoted in Social Media Today, 25 February 2025

Read that twice. Watch time is not one thing. It is two. Percentage watched and seconds watched. Both at once. The system looks at both. Mosseri does not say how they are weighted. He does not say which matters more. He says both.

So a ten-second watch of a one-minute video is the same as a ten-second watch of a ten-second video. In absolute terms. But in percentage terms, the first is 17%. The second is 100%. The system reads both numbers. And combines them somehow. The cards do not say how.

That is the most important sentence in this section. Mosseri names two measurements. The cards name eight thresholds. None of them agree.

The thing you optimize for is not the thing Meta describes

Here is the problem. You open Instagram Insights. You see "average watch time." One number. Seconds, probably. You assume that number maps to the signal Mosseri named. Watch time. So you try to make it go up.

But the cards say the system does not read one number. It reads eight. Some are thresholds you cross or miss. Three seconds. Five seconds. Ten seconds. Some are percentages. Ninety-five percent. Some are comparisons. Did you watch more than 95% of people who saw reels of this length? Some are aggregates. How much time did viewers from your country spend on this?

And the cards do not say how those eight are weighted. Or combined. Or whether one overrides the others. They just list them. As separate signals. Feeding separate predictions. On separate surfaces.

So when you optimize for the Insights number, you are optimizing for something. But not for the bundle of things the system actually scores. You are aiming at one metric. The system is reading eight. And you do not know which of the eight matters most. Because Meta does not publish the weights.

One number would be easier to game

This is not an accident. Meta does not publish one definition because one definition would be easier to reverse-engineer. And easier to game. If watch time were just "total seconds watched," every creator would know what to optimize for. Make longer videos. Or make videos people rewatch. The tactic would be obvious.

But if watch time is a bundle, the tactic is not obvious. Do you optimize for three-second retention? Or 95% completion? Or the comparative rank against same-length reels? Or absolute seconds? All of them pull in different directions. A 90-second reel with 50% completion is 45 seconds. A 15-second reel with 95% completion is 14 seconds. Which one wins? The cards do not say.

That ambiguity is a feature. Not a bug. It makes the system harder to manipulate. You cannot point all your effort at one metric. Because the metric is not one thing.

But it also makes Mosseri's advice impossible to follow precisely. He says optimize for watch time. The cards say there is no single watch time to optimize for. There are eight versions of it. On three surfaces. With no published weights.

What this means for what you make

Start with what you can know. The thresholds exist. Two seconds. Three seconds. Five seconds. Ten seconds. Fifteen seconds. Those are in the cards. As separate signals. So the first few seconds matter. That much is clear. A viewer who drops off at two seconds triggers one signal. A viewer who stays past three seconds triggers a different one.

The 95% thresholds exist too. Both kinds. Watching 95% of a video is one signal. Watching more than 95% of viewers who saw reels of the same length is another. So completion rate matters. But so does relative completion rate. Against a cohort. Of same-length reels.

Mosseri's February statement adds one more thing you can know. The system looks at both percentage and seconds. So a long video is not automatically penalized for low completion percentage. As long as it racks up seconds. But you do not know the weight of percentage versus seconds. He did not say.

What you cannot know is which of these eight matters most. Or how they combine. Or whether different surfaces weight them differently. The cards list them. They do not rank them. Mosseri calls watch time the top signal. But he does not break that signal into its eight components. And he does not say which component dominates.

So you are left with a heuristic. Not a formula. The first few seconds matter. Completion rate matters. Absolute watch time matters. Relative watch time compared to same-length reels matters. All of them matter. You cannot know which one tips the scale.

Watch time is not only an early metric

The common story about ranking is that the first few hours decide everything. Get the early views. Get the early watch time. Clear the early thresholds and the algorithm carries the post from there.

The cards do not support that story cleanly. Yes, the three-second threshold matters. Yes, the skip-within-two-seconds signal is negative. Those are early behaviors. But the 95% completion thresholds are not early. They are cumulative. They grow over time. As more people watch. And the comparison to same-length reels is not early either. It needs a distribution to compare against.

So watch time is not just an early metric. It is also a late one. The version of it that matters most might not be the version that shows up in the first hour. Or the first day. It might be the version that accumulates over a week. Or a month. The cards do not say.

We do not know if early watch time causes late watch time. The way early likes might signal to the algorithm that a post is good. What we know is that the cards define watch time in eight ways. And only some of them are measurable early. The rest take time to fill in.

What we could not find out

Three things we looked for and did not find.

Any weight for any version of watch time. The cards list the signals. They do not weight them. So every claim about which threshold matters most is unproven. Including claims in this post. We say the thresholds exist. We do not say which one dominates. Because the cards do not say.

How the eight versions combine. Do they add? Do they multiply? Does one override the others? Is there a hierarchy? The cards list them as separate signals. They do not describe a formula that combines them. So we do not know if hitting three seconds is a gate you must pass before 95% completion matters. Or if they run in parallel. Or if different surfaces combine them differently.

Whether the Insights number maps to any of these eight. You see "average watch time" in Insights. We do not know which of the eight card signals that number represents. Or if it represents a ninth thing entirely. Meta does not say. So when you optimize for the Insights number, you are optimizing for something. But we cannot tell you which of the eight it is.

How we did this

We read five Meta documents on 17 August 2026. Three system cards: Reels Chaining, Explore, and Feed. All three carry an "updated" stamp from Meta. Reels Chaining was stamped 11 November 2025. Explore was stamped 22 June 2026. Feed was stamped 29 June 2026. We also read Adam Mosseri's Instagram Ranking Explained post, published 31 May 2023. And we read two Social Media Today articles quoting Mosseri's January and February 2025 statements.

Meta's Transparency Center pages are built by code in the browser. A plain download returns an empty shell. So all six pages were read in a browser. Every quotation was copied as a reader sees it. Not from a summary.

Limits: five of seven sources are first-party statements. All from the company we are writing about. A system card is a description Meta chose to publish. Not an audited account. Meta rewrites these cards without saying when. So every quotation carries the stamp it showed on 17 August 2026. The Social Media Today articles quote Mosseri. We could not access his original Instagram posts directly. So we cite the trade press articles that transcribed his statements. That makes them second-hand sources. We mark them as such.

Sources

1. Instagram Reels Chaining AI system card, stamped UPDATED NOV 11, 2025. Meta Transparency Center. Read in a browser. Retrieved 17 August 2026.

1. Instagram Explore AI system card, stamped UPDATED JUN 22, 2026. Meta Transparency Center. Read in a browser. Retrieved 17 August 2026.

1. Instagram Feed AI system card, stamped UPDATED JUN 29, 2026. Meta Transparency Center. Read in a browser. Retrieved 17 August 2026.

1. Adam Mosseri, Instagram Ranking Explained, 31 May 2023. About Instagram. Retrieved 17 August 2026.

1. Andrew Hutchinson, Instagram Shares New Insights into its Algorithm for 2025, Social Media Today, 22 January 2025. Quotes Mosseri's statement about watch time, likes and sends as the top three signals. The original source is Mosseri's Instagram post at instagram.com/p/DFFyRp-pINJ/, which we could not access directly to verify the transcript. Retrieved 17 August 2026.

1. Andrew Hutchinson, Instagram Explains Why Longer Videos Aren't Penalized in its Recommendations, Social Media Today, 25 February 2025. Quotes Mosseri's statement about percentage watched and seconds watched. Retrieved 17 August 2026.

Changelog

• 2026-08-18: First version drafted.