Instagram Reels Algorithm: What Really Drives Reach

Three signals drive the distribution of a Reel: watch time, likes per reach and DM shares per reach. Adam Mosseri named them in January 2025, and they still hold in 2026. Shares weigh the most for reaching non-followers, three to five times a like according to Mosseri. On top of that, Instagram applies different ranking models depending on the surface. The direct consequence for your production: write the Reel to be sent to someone in a private message, not to be liked. The rest of this article turns those signals into specifications for shooting, editing and briefing. According to Social Media Today (January 2025), Adam Mosseri named watch time, likes and shares as the three signals that matter most for ranking, and encouraged creators to track average watch time, likes per reach and sends per reach in their analytics.

What are the three signals Meta actually named, and in what order?

Watch time, likes per reach, sends per reach. Three signals, publicly named by Adam Mosseri in January 2025 as those that weigh most across all Instagram surfaces, and picked up in 2026 by Hootsuite in its overview of social algorithms. Not four. Not "engagement" in the broad sense. Three.

The part almost every French article skips is the "per reach". It is not the number of likes that is looked at, it is the number of likes divided by the number of accounts reached. A Reel with 180 likes on 48,000 accounts reached (0.37%) sends a weaker signal than a Reel with 55 likes on 1,900 accounts (2.9%). The second keeps being pushed, the first fizzles out. If your monthly reporting lines up raw totals, it measures nothing of what the machine is looking at.

Shares weigh the most for reaching non-followers: Mosseri says they count slightly more than likes for unconnected reach, without giving a numerical weighting. This is a platform statement, not an independent audit. The exact weighting is not known, it varies by surface and probably by user. But the hierarchy itself has been stable for eighteen months.

What about saves and comments?

Used, but not in the trio. The inro.social analysis published in autumn 2025 is clear on this point: saves do feed the ranking models, without being officially placed above likes; comments are used by some surface models without being part of the top 3. Which invalidates the very widespread "saves are the new king" line. A save is a good indicator of a piece of content's lasting value. It is not a lever for cold reach.

As for views: they are not a signal. They are a display counter. Instagram counts them for you, not for itself.

"My business account is throttled." No. Instagram has no interest in degrading the accounts that buy its advertising. When you compare a business account and a personal account on equivalent content, the gap observed comes from the content: too corporate, too promotional, too polished. The shadowban exists in brands' imagination far more than in the logs.

Why does the same Reel take off in Explore and die in your followers' feed?

Because the same signals do not decide. Instagram does not have one algorithm, it has several, one per surface. The feed, Stories, the Reels tab and Explore answer to distinct prediction models, trained on different objectives. dataslayer.ai described this very well in late 2025: Instagram looks less like a single global feed than a swarm of personalised prediction engines, which reward depth of attention and active sharing rather than vanity metrics.

The operational distinction to remember comes down to two words: connected reach and unconnected reach.

For connected reach, your followers, the relationship history dominates. Has this person already watched your videos to the end, commented, replied to your Stories, visited your profile? The content matters, but it comes behind the relationship. For unconnected reach, Instagram has no relationship data to work with. All that is left is content signals: time watched, completion rate, share rate, save rate. Hence a complete reversal of priorities depending on the target.

Which means that a Reel "for the community" and a Reel "for acquisition" are not briefed the same way, not edited the same way and not judged on the same numbers. In 2025 we stopped delivering to our clients a single series of Reels supposed to do both. It does not work: the community Reel assumes a context that the cold audience does not have, and the acquisition Reel looks impersonal to those who already follow you.

A word for those who also produce for TikTok: the logic there is close but does not weight the same signals, and the testing rhythm is nothing alike. We went into this in our article on how TikTok chooses which videos to show. Duplicating a Reel to TikTok without reworking the hook remains the quickest way to lose both.

The DM share: why has it become the real engine of cold reach?

Because it is the only gesture for which the user pays a social cost. Liking takes a tenth of a second and commits nobody. Sending a video to a friend puts your credibility with them on the line. Instagram knows it, and values this gesture more than the like for opening up new audiences; the exact weighting has never been published, only blogs put forward an unsourced ratio. According to Médiamétrie's L'Année Internet 2025, relayed by CB News, 68% of French people visit their social networks and messaging apps daily, or 43.9 million individuals, up 2% on 2024.

The problem: the DM share is not won in the edit. It is won in the writing. A brilliant edit on a subject nobody feels like forwarding will produce good watch time and zero sends. We see this every month on very well-produced brand Reels: beautiful image, clean rhythm, 0.2% sends per reach. The video is pretty and does not circulate.

The five reasons that trigger a forward in a private message

• "This concerns you." Direct usefulness for an identifiable person: a tip, a warning, a deadline, a trade trick.
• "This is so us." Identity content, the group's running joke, the situation lived by two. The share works here as a wink.
• "Look what this does." Visual proof: before/after, product test, a demonstration that surprises. E-commerce brands massively underuse this register.
• "Shall we go?" Coordination: a place, an event, a product to buy together. Restaurants and physical venues have a structural advantage here.
• "I was right." The video that settles a debate. A clear opinion on a subject where people disagree becomes an exhibit.

A concrete instruction for your scripts: before approving a concept, write the sentence someone would type above the Reel when forwarding it. If you cannot write it, the concept is not ready. We have built it in as a mandatory field in our briefs since early 2025, and it is the line that does the most damage in approval meetings. Many ideas do not survive it.

What we do not know: whether Instagram distinguishes a share to a friend from a share to a group, or from a send to yourself. Nothing public says. We measure sends as a whole, and we take care not to invent an internal mechanism that nobody has documented.

Retention or completion: what should you really aim for in the edit?

Completion, then the loop, not duration. The model tries to predict whether the viewer will make it to the end, as Shopify sums it up in its 2026 guide to Reels. A 12-second video watched in full sends a better signal than a 50-second video abandoned at second 9. Lengthening to "tell the story properly" is the most costly editing decision we see go through client approval.

The production constraints we apply, which are debatable but defensible:

• Stay under three minutes. gyre.pro reminded readers in early 2026 that the recommendation window stays open below that bar. Beyond it, the format leaves the Reels logic. In practice, our organic deliverables rarely exceed 35 seconds. As a reminder, according to MediaNama, Adam Mosseri announced on 19 January 2025 that Reels could now run up to three minutes, against a previous limit of 90 seconds.
• A new visual idea every 2 to 3 seconds. A change of shot, scale, overlay or movement. The eye needs a reason to stay.
• Burned-in subtitles, every time. A large share of views happen with the sound off, especially in cold reach where the person does not know you and will not take the risk of turning the volume up.
• Loop. The last frame joins the first, or the last sentence restarts the first. A clean loop mechanically doubles the time watched for part of the audience.
• No opening logo, no jingle. Three seconds handed over to drop-off.

The first three seconds, in one sentence

The hook is the most direct lever on watch time, and the only element of the video that can be re-shot on its own for a few tens of euros. We will not re-explain it here: the definition, the measurement method and the real cost are covered in our article on the advertising video hook. Just remember that the 3-second view rate is the first metric to pull from Insights when a Reel does not take off. If it is low, the problem is neither the sound, nor the hashtags, nor the publishing time.

Are your Reels engagement benchmarks comparable? No, and here is why

The figures published in 2026 range from 0.30% to 5.53% average engagement on Reels, and they are all honest. They do not use the same denominators, the same samples, or the same account scopes. Setting two benchmarks side by side without normalising them is like comparing a price excluding VAT with a price including VAT in two currencies.

The gap between 0.48% and 5.53% is not an error. Socialinsider divides by followers, Metricool divides by views. Quid publishes a median, which flattens the hits. Add composition bias, since a panel full of creators does not behave like a panel of retail brands, and you get figures that cannot be compared with each other, all usable separately.

What we do with our clients: we drop the external benchmark as a contractual target. We fix a single denominator, accounts reached, available in Insights, we calculate likes/reach, sends/reach and saves/reach on the last thirty Reels, we take the median, and that is the baseline. Three months later, the only question worth asking is: have we moved our own median? The rest is slide decoration.

One figure is still worth knowing, because it explains the general gloom: Metricool measured in 2026 a 24% rise in the volume published by brands across all formats, to which Instagram responded with 27% more views and 19% more engagement, but with an engagement rate down 6.61%, from 3.79% to 3.54%. More content, more absolute views, less yield per post. If your rates are falling while your volumes are rising, you are in line with the market average, not being penalised.

How to test a Reel without burning media budget: the 14-day Trial Reels protocol?

Trial Reels are a free test bench on cold audience, and almost nobody uses them as one. Launched in December 2024, they are activated at the time of publishing via the "Audience" setting and then the "Trial" option: the Reel is then shown only to non-followers. In other words, Instagram lends you exactly the surface you are interested in, cold reach, without you paying a cent of media.

The insights come back within 24 hours according to flowgent.ai (January 2026): views, likes, comments, shares, plus a comparison with your previous trials. Meta also notes that a Trial Reel can take longer to accumulate views, since it is served to people who do not know you. Fewer views therefore does not mean failure. If the metrics cross a threshold, the Reel graduates automatically and moves to your full audience, a mechanism documented by postfa.st in early 2026. According to Meta's announcement on Trial Reels (December 2024), this automatic sharing with your followers is an option you choose when creating the Reel, and Instagram triggers it if it judges the Reel to be performing well based on the views received during the first 72 hours.

The protocol, day by day

• Days 1-2, fix the variable. Only one thing changes between versions. The hook, or the sound, or the subtitle format. Never two. Buffer documented a clean test of this kind: with identical content, trending audio generated 24% more views than a voice-over. One figure, one variable, one decision.
• Days 3-9, publish 2 to 3 trials a day. Volume is not a luxury: the creator Millie Adrian reports around 2,600 followers gained after publishing nearly 75 Trial Reels in 15 days, or five a day. It is a single case, not a law, but it gives the order of magnitude of the throughput needed for the differences to be readable.
• Days 4-12, read at 48-72 h. The collection window generally runs between 24 and 72 hours according to Infinity Likes. Before 48 h, the rankings still flip.
• Day 13, decide on sends/reach, not on views. The winner is the one with the best share ratio, not the one with the biggest counter.
• Day 14, industrialise. The winning concept becomes a template: same structure, three variations, and it goes out as a normal post. The best two or three then go into paid media.

A firm position: a creative concept that did not hold up as a Trial Reel has no reason to hold up in paid. You would be spending budget to buy information the platform gives you for free. The question of the monthly throughput needed to feed this set-up can be calculated; it depends on your media tier, and we detail the method in the creatives budget calculation by media tier.

Does a Trial Reel penalise the account? Nothing, neither from Meta nor in the serious analyses published since December 2024, points to a penalty. The Reel is simply served to a different audience. The worry comes from a misreading of the view counter, which is mechanically lower on cold audience.

What does this change concretely in your brief and your production quote?

Six lines to add to the brief, and one line to remove from the contract. The algorithm's signals are not reporting metrics, they are specifications. As long as they do not make it down into the document the editor reads, they change nothing.

The six lines:

• Target surface. Community or acquisition. Only one of the two per Reel, decided before the shoot.
• The sharing sentence. What someone would write when forwarding the video. If it is empty, the concept goes back to writing.
• Maximum length. Expressed in seconds, not "short". The editor needs a number.
• Loop point. How the end joins the beginning, described in one line.
• Variable tested. The element that changes compared with the previous version, and only that.
• Acceptance threshold. The median sends/reach of your last thirty Reels. Below it, the deliverable is not approved.

These six lines slot into an existing brief structure; if you do not have one, take the 9-field creative brief template and graft them into it. Twenty extra minutes of writing per series, and a good part of the approval back-and-forth disappears.

The line to remove from the contract, now. View commitments. We still see commercial proposals go by that promise a cumulative volume of views over the quarter. It is an indicator that the algorithm itself does not treat as a signal, and which is inflated by publishing more mediocre videos. Billing on that basis aligns the provider's interests with volume, not with circulation. We have turned down this type of clause several times, including when it cost us the contract.

On the budget side, the order of magnitude we observe on the French market in 2026 for a produced organic Reel, writing, shoot, edit, variations, is between €150 and €600 per unit depending on the level of production and the volume committed, with a clear threshold effect beyond around ten videos a month. The line-by-line, item-by-item detail is in our annotated Reels production quote.

Originality, reposts, AI: what costs you reach in 2026?

The repost is the quietest way to sabotage your own distribution. Instagram prioritises original content and its creator over the accounts that republish it. The rule has been in place since 2024, has been reinforced since, and also applies to content taken from other platforms with a visible watermark. According to Engadget (April 2024), Instagram announced that it would stop recommending Reels altogether from accounts that had posted unoriginal content 10 or more times in the previous 30 days.

Which does not mean recycling is dead. Taking the structure of a Reel that worked and reshooting it with your products, your face and your voice: that is original. Downloading a creator's video and reposting it with a credit in the caption: that is a repost, and your account will be served behind the source.

On generative AI, we prefer to say what we do not know. Meta has communicated on the share of recommended content in the feed, and a figure circulates in several French articles about the fourth quarter of 2025. We found it truncated in the sources we consulted, with no access to the original full sentence. So we are not publishing it. What we do see on our own productions, however: fully generated sequences hold completion poorly beyond 8 seconds as soon as a human face speaks, and share rates collapse. For product shots, transitions or backgrounds, AI passes without problem.

And what about hashtags?

Marginal. Mosseri has said publicly that they do not do much for reach. They help qualify the subject of a video for the models, that is all. Three to five descriptive hashtags, not thirty. The time saved on hunting for hashtags is better invested in rewriting the hook.

One last thing, and it is the one that costs brands we take over midway the most: the Instagram Reels algorithm does not reward consistency in itself. Publishing five times a week content that nobody forwards does not produce cold reach, it produces team fatigue and reporting that climbs in volume and falls in rate. Three Reels a week written to be forwarded beat ten Reels written to fill a calendar. Start with the sharing sentence. Everything else follows from it.

Frequently asked questions

How many Reels should you publish per week?

Three to five a week are enough to feed the models, provided each one is written to be shared. During a testing phase with Trial Reels, the throughput temporarily goes up: Millie Adrian published about 75 trials in 15 days. Outside a testing phase, consistency matters less than the shares-to-reach ratio of each video.

Why have my Reels been getting fewer views for several weeks?

In most cases, it is a drop in rate, not a penalty. Metricool measured in 2026 a fall in the Instagram engagement rate from 3.79% to 3.54% despite 24% more content published by brands. First check your 3-second view rate and your sends/reach: if those two ratios hold, it is the market that is moving.

Are hashtags still useful on Reels?

Very little for reach. Adam Mosseri has stated publicly that they have no significant effect on distribution. They help the models qualify the subject of a video. Three to five descriptive hashtags are more than enough, and the time saved is better spent rewriting the first three seconds.

Does publishing as a Trial Reel penalise the account's reach?

Nothing indicates it in Meta's communications or in the analyses published since the December 2024 launch. The Reel is shown only to non-followers, so the view counter climbs more slowly. Meta even states this explicitly. If the metrics cross the threshold, the Reel graduates automatically to the full audience.

Is my business account throttled by Instagram?

No. Instagram has no interest in degrading the accounts that buy its advertising. The gaps observed between business and personal accounts are explained by the content: too promotional, too corporate, not shareable enough. Before talking about a shadowban, compare your sends/reach with that of your ten best videos.

Is it better to use a trending sound or a voice-over?

It can be tested, and the test is free. Buffer documented a like-for-like comparison where trending audio generated 24% more views than a voice-over. One test, on one account: redo it on yours with Trial Reels, one variable at a time, and read at 48-72 hours.