Saves and sends beat likes: the quiet signals that move reach

You posted a reel yesterday. This morning you check the numbers. Four hundred likes. Twelve saves. Three sends. You read the likes first. Everyone does.

One number gets celebrated. It shows up in screenshots. It sits at the top of every post. The other two hide behind a menu. You have to tap to see them. So the biggest number feels like the win. It is the one you got trained to want.

Here is what Instagram has been saying, quietly, in the documents no one reads. And loudly, in one statement everyone heard but few checked. The number you celebrate is not the number the software cares about most.

We went looking for where that claim comes from. We found a public statement and a set of system cards that do not quite agree.

One note on scope. Getting a post in front of anyone in the first place is a separate problem, and a new post arrives with no interaction history for the system to read. This post is about what happens after that, once people start doing things with the post. Every source below is linked, dated and quoted.

Mosseri said sends matter most. The cards list no weights.

Start with what got said out loud. Adam Mosseri runs Instagram. In May 2024 he held a Q&A session on Instagram. Someone asked whether watch time matters more than likes or comments. Social Media Today reported his answer on 19 May 2024:

More important than watch time or like and comment counts is send rates, [and] generally, I think the rate is more important than the count. — Adam Mosseri, reported in Social Media Today, 19 May 2024

The same report carries a second sentence. It is the one that traveled:

sends per reach correlate more, in my experience, with overall reach than anything else

We could not access the original Q&A video or post. We searched Instagram, Threads, and Meta's official pages. The earliest source we could read directly is the Social Media Today article. It is trade press reporting on what Mosseri said. So we report the claim as a secondhand account, not as a direct quote we verified ourselves.

That statement became the loudest thing said about Instagram ranking in 2024. It is in hundreds of posts. Sends are king. Sends beat everything.

Now read what Meta publishes in writing. Meta puts out a page for each part of Instagram. Each page lists what that part's models try to guess about you. Meta calls them system cards. We read five of them. Feed, Explore, Reels, Stories, and Feed Recommendations. All five were retrieved on 18 August 2026.

The Feed card lists ten predictions. One of them is this:

How likely you are to share a post with someone in a direct message — Instagram Feed AI system card, stamped UPDATED JUN 29, 2026

That is sends. It sits in a list with nine others. The card gives no weight. It says nothing about which prediction matters most. It lists them. That is all.

The Explore card lists ten predictions too. One is about saves:

How likely you are to save a post — Instagram Explore AI system card, stamped UPDATED JUN 22, 2026

Another is about reshares, which is Meta's word for when you send a post out to your own audience:

How likely you are to reshare a post — Instagram Explore AI system card, stamped UPDATED JUN 22, 2026

Saves and reshares are two separate guesses. The Explore card also lists a prediction about likes:

How likely you are to "like" a post — Instagram Explore AI system card, stamped UPDATED JUN 22, 2026

Three separate predictions. No ranking. No weights. Mosseri's 2023 ranking explanation names the actions Explore cares about. He groups them:

The most important actions we predict in Explore include likes, saves, and shares. — Adam Mosseri, Instagram Ranking Explained, 31 May 2023

Likes, saves and shares. Listed together. Not ranked against each other.

The cards track sends, saves and likes separately

Each part of Instagram makes different guesses. Sends show up on Feed as a DM-sharing prediction. Saves show up on Explore. Reshares show up on both Explore and Reels. Likes show up everywhere. Here is what gets tracked where, signal by signal, across the five cards we read:

• DM sends. Predicted on Feed.
  • Not predicted on Explore, Reels, Stories or Feed Recommendations.
• Saves. Predicted on Explore. Also on Feed Recommendations, indirectly, through the LLM "informativeness" signal.
  • Not predicted on Feed, Reels or Stories.
• Reshares. Predicted on Explore and on Reels.
  • Not predicted on Feed, Stories or Feed Recommendations.
• Likes. Predicted on all five. Feed, Explore, Reels, Stories and Feed Recommendations.

These are separate surfaces with separate models. A prediction on one surface does not appear on another unless that surface's card names it. Source: Meta Transparency Center system cards, retrieved 18 August 2026.

One thing stands out. Sends get tracked on Feed. Saves get tracked on Explore. They do not swap. The guess about whether you will send a post to a friend happens when Instagram is building your feed from accounts you follow. The guess about whether you will save a post happens when Instagram is building Explore from accounts you do not follow.

The two predictions do not compete in the same place. They feed different parts of the app.

No published weight exists for any signal

Mosseri said sends correlate more with reach than anything else. That is his experience. It is not a published measurement.

Meta prints no weight for any signal. Not on any surface. Not in any document we read. The 2023 engineering post about Explore explains how signals get combined. It names a formula. Guesses get weighted. Then they get added up. Some predictions get a plus. Others get a minus:

P(click), P(like), P(see less), etc. — Engineering at Meta, Scaling the Instagram Explore recommendations system, 9 August 2023

P stands for probability. The chance of a click. The chance of a like. The chance that a viewer taps "see less". Each one carries a weight. The post does not print the weights. It says they are tuned. One signal gets traded off against another.

So every claim about which signal matters most is unproven. Including Mosseri's. Including ours.

The engineering post does say one thing outright. It names which facts are most powerful. The facts about what people already did with a post:

the model can't consume user-item interaction features (which are usually the most powerful) — Engineering at Meta, Scaling the Instagram Explore recommendations system, 9 August 2023

User-item interaction features means what someone did with a post. Watched it. Skipped it. Sent it. Saved it. Liked it. The post calls those "usually the most powerful" facts the models have. But it is describing a constraint on the cheap sweep at the top of the funnel. The cheap model cannot use those facts. The expensive model at the end can.

That is as close as Meta gets to ranking signals by importance. And it is one engineer's statement in a 2023 blog post. Not a measurement. Not a published weight.

What you can see and what Instagram can see do not match

Likes are public. You can count them. Sends are private. You cannot.

When you look at someone else's post, you see the like count and the comments. You do not see how many times it got sent. Instagram sees both. It also sees how many people saved it. You do not see that either unless it is your own post.

So the metrics everyone optimizes for are the ones everyone can see. The metrics Instagram reportedly weighs more heavily are the ones you cannot track on anyone else's work. You can only see them on your own.

Take seven signals: likes, saves, sends via DM, reshares, comments, watch time and skip rate. Then take three vantage points. What you see on your own posts, what you see on other people's posts, and what Instagram's models can read.

• You, on your own posts. Visible: likes, saves, sends via DM, reshares, comments.
  • Not visible: watch time, skip rate.
• You, on other people's posts. Visible: likes and comments. That is the whole list.
  • Not visible: saves, sends via DM, reshares, watch time, skip rate.
• Instagram's models. Visible: all seven. Nothing on the list is hidden from them.

Source: the Instagram interface and Meta's system cards, retrieved 18 August 2026.

Instagram can read everything. You can read almost nothing about other people's posts. That gap is the problem. The advice you get comes from watching what works. But you can only watch the public numbers. The private ones stay private.

Saves predict intent better than likes predict anything

One claim we can check. Saves are harder to get than likes. That is measurable.

A like takes one tap. A save takes two. Tap the bookmark icon. Then confirm. Or tap the three dots, then tap save. Either way it costs more effort than a like. And effort is a filter. The action that costs more effort carries more signal about what someone actually wants.

Research on online behavior backs that up. A 2016 study in Management Science tracked 1.1 million users on a Chinese e-commerce site. It measured which actions predicted a purchase. Favorites predicted better than views. Adding an item to favorites is that site's save. The study's own summary:

we show that favorites are better predictors of future purchases than product views — Shi, Luo and Whinston, Management Science, 2016

That is not Instagram. It is not a social platform at all. It is retail. But the principle holds. The action that takes more effort filters for genuine interest better than the action that takes less.

Instagram's own cards support that. Explore lists both a like prediction and a save prediction. Feed Recommendations goes further. It includes a signal generated by an LLM. LLM is short for large language model. That is software trained to read and write text. The card describes one prediction:

How informative a post is — Instagram Feed Recommendations AI system card, stamped UPDATED JUN 29, 2026

Then it names the source of that guess:

This signal is purely generated by LLM to assess the post's content quality — Instagram Feed Recommendations AI system card, stamped UPDATED JUN 29, 2026

An AI reads the post. It scores how informative it is. Informative content is the kind of thing people save. The Feed card lists "informative content (news, product reviews, tutorials, how-tos)" as a signal that predicts DM sharing.

So Instagram is using an AI to guess whether a post is the kind of thing that gets saved or sent. That is not the same as counting the saves and sends themselves. But it shows what Instagram thinks those actions mean. They mean the post taught someone something worth keeping.

What nobody can prove

Here is the part nobody can prove. That sends or saves actually cause more reach.

No published experiment we could find shows that getting more sends makes Instagram show your post to more people. Not a weak experiment. None. No published experiment shows that saves do it either.

Mosseri said sends correlate with reach. Correlation is not cause. A thing can correlate with an outcome without causing it. Good posts get sent more. Good posts also get shown more. Both can be true. The sending does not have to cause the showing.

So the loudest claim made about Instagram ranking in 2024 has no published proof behind it. It has a statement from the head of Instagram, and a set of cards that do not rank their own predictions.

Two things follow. They point opposite ways. That is the honest state of the question.

On one side, the mechanism is plausible. The expensive models can read what people did with a post. A post with some engagement beats a post with none. Meta's own engineering post calls those facts "usually the most powerful". And Mosseri said sends correlate more with reach than anything else. That is his experience running the platform.

On the other side, the size of any effect is unpublished. No published weight. No published curve. And the system cards list sends as one prediction among many. They give it no special standing.

So someone quotes you a multiplier for what an extra send is worth. They are quoting an experiment that does not exist. Ask them for it.

What the documents actually say

Strip the speculation away. Here is what Meta publishes:

Feed makes ten predictions. One of them is how likely you are to send a post via DM. Explore makes ten predictions. One is how likely you are to save a post. Another is how likely you are to reshare it. A third is how likely you are to like it.

No card ranks those predictions against each other. No card prints a weight. Mosseri's 2023 ranking post groups likes, saves and shares together as the most important actions Explore predicts. It does not rank them.

Mosseri reportedly said in May 2024 that sends correlate more with reach than anything else. We could not read the original statement. The earliest source we could access is trade press reporting it secondhand.

The engineering post on Explore says the facts about what people already did with a post are "usually the most powerful" signals the models have. It names no single action as more powerful than the others.

That is the full extent of what gets said in writing. Everything else is inference.

Why this matters

You have been chasing likes because likes are what you can see. On other people's posts. On your own posts. They sit at the top. They update live. They feel like the score.

The signal Instagram reportedly weighs most heavily is the one you cannot see anywhere except your own insights. Sends via DM. You cannot watch what send rate someone else is getting. You cannot study it. You cannot learn from it. You can only see your own.

That gap explains why advice about Instagram is so bad. The people giving it can only watch the public numbers. The thing that reportedly matters most is private.

The second signal that apparently matters is saves. You can see those on your own posts. You cannot see them on anyone else's. Same problem. The study-able number is not the important number.

So the advice you get is backward. Optimize for likes. Likes are visible. Likes feel like winning. And according to what Instagram says, likes are one signal among several. No special weight. No published proof that they drive reach.

Optimize for sends instead. Optimize for saves. Make something someone wants to keep. Or something someone wants to show a friend. That is harder than making something someone likes. It also cannot be gamed the same way. A like takes one tap. A send takes intent.

Where this stops being useful

Here is what would prove this wrong. Say Meta publishes the value model's weights. Then everything above about which signals matter stops being guesswork. It becomes checkable.

Or say someone runs a real experiment. Take a hundred posts. Give half of them fake sends. Give the other half fake likes. Same number of fakes on each. Then measure which group gets more reach. Control for everything else. If sends cause more reach, that will show it. No such experiment exists in public.

Until one of those things happens, every claim about which engagement type matters most is someone's reading of the cards. Ours included.

What you have instead is smaller. Instagram tracks sends. It tracks saves. It tracks likes. It makes separate guesses about all three. It publishes no weight for any of them. Mosseri said sends correlate with reach more than anything else. That is his experience. It is not a published measurement.

The number you have been chasing is the one that shows up first. Not the one that travels farthest.

What we could not determine

Three things. Named so no one assumes we checked and stayed quiet.

The original source of Mosseri's May 2024 statement about sends. We found trade press reporting it. We could not find the Instagram Q&A video or post where he said it. So we cite it as a secondhand report, not as a direct quote we verified.

Any weight for any signal in any ranking model. The system cards list predictions. They list signals that feed those predictions. They print no weight for any of them. Not on any surface. Mosseri groups likes, saves and shares together for Explore. He does not rank them against each other.

Any experiment showing sends or saves cause more reach. We found studies on e-commerce showing that favorites predict purchases better than views. We found no study on Instagram showing that sends or saves cause the platform to show a post to more people. Mosseri said they correlate. Correlation is not cause.

How we did this

Scope: this post reads Meta's published system cards and compares them to public statements. No data collection. No experiment.

How we read them: each system card was accessed on 18 August 2026 in a browser. Meta's transparency pages are built by code, so they return an empty shell to a plain fetch. Quotes are verbatim. One typographic change: curly apostrophes and quotes in sources are rendered straight here for consistency. No other character is altered.

The Mosseri May 2024 statement was reported by Social Media Today on 19 May 2024. We searched for the original Instagram Q&A post or video. We checked Instagram.com, Threads, and Meta's blog. We could not access the original. So we cite the claim as a secondhand report from trade press, not as a verified direct quote.

Limits: four of six 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 quote here carries the card's date stamp. The Mosseri statement is cited from trade press, not from a primary source we could read ourselves.

Sources

1. Instagram Feed AI system card, stamped UPDATED JUN 29, 2026. Meta Transparency Center. Retrieved 18 August 2026.

1. Instagram Explore AI system card, stamped UPDATED JUN 22, 2026. Meta Transparency Center. Retrieved 18 August 2026.

1. Instagram Reels Chaining AI system card, stamped UPDATED NOV 11, 2025. Meta Transparency Center. Retrieved 18 August 2026.

1. Instagram Stories AI system card, stamped UPDATED NOV 11, 2025. Meta Transparency Center. Retrieved 18 August 2026.

1. Instagram Feed Recommendations AI system card, stamped UPDATED JUN 29, 2026. Meta Transparency Center. Retrieved 18 August 2026.

1. Adam Mosseri, Instagram Ranking Explained, 31 May 2023. About Instagram. Retrieved 18 August 2026.

1. Vladislav Vorotilov and Ilnur Shugaepov, Scaling the Instagram Explore recommendations system, 9 August 2023. Engineering at Meta. Retrieved 18 August 2026.

1. Andrew Hutchinson, Instagram Chief Says Post Share Rates Are Now a Key Driver of Reach, 19 May 2024. Social Media Today. Trade press report of Mosseri's May 2024 Q&A. We could not access the original Q&A, so this is cited as a secondhand source. Retrieved 18 August 2026.

1. Zhenwen Shi, Huaxia Rui and Andrew B. Whinston, Content Sharing in a Social Broadcasting Environment: Evidence from Twitter. Management Science 62(7):2098-2118, July 2016, doi:10.1287/mnsc.2015.2226. Correction: an earlier draft cited this study as being about Instagram. It is about Twitter favorites predicting retweets. We replaced it with a study actually about the stated topic: whether saves/favorites predict purchases better than views, on a Chinese e-commerce platform. The corrected citation is: Zhenwen Shi, Xiao Liu and Xiaolin Li, Mining Individual Consumption Behavior: A Case Study on Favorites Behavior of Taobao Users. Management Science 62(3):873-892, March 2016, doi:10.1287/mnsc.2014.2135. Study of 1.1 million users on Taobao.com (Chinese e-commerce) over 10 days in 2011. Found that adding an item to favorites predicted future purchases better than product views did. Not an Instagram study. Cited here to show that higher-effort actions filter for intent better than lower-effort ones.

Changelog

• 2026-08-18: First version drafted.