Average Instagram Engagement Rate: Benchmarks From 245,559 Measured Posts

Ask what the average Instagram engagement rate is and the published answers on one search results page run from 0.36% to 4.42% for the same platform in the same year. We measured 245,559 Instagram posts against the people who actually saw them, the entire corpus buzzabout tracks rather than a sample of it: the median post gets 2.41%, the top quartile 4.25%, the top decile 6.77%. The mean is 3.52%.

The median barely moves between posts seen by 350 people and posts seen by 400 million. So the law everyone repeats, that engagement falls as you grow, is a property of dividing by followers rather than a property of Instagram. And it only exists on the 54.5% of posts where the platform publishes a view count at all, which is the asterisk none of the benchmarks carry.

What is the average Instagram engagement rate?#

Instagram engagement rate against published reach. Formula: (likes + comments + shares) / views. n = 245,559 posts from 154,393 accounts, published December 2, 2021 to July 23, 2026.

The median post in that set collected 116 engagements on 6,258 views. A further 4,042 posts, 1.65% of the set, collected nothing at all.

These 245,559 posts are 54.5% of the 450,324 we pulled. Of the rest, 173,592 carry a view count our pipeline modeled rather than one Instagram published, and 31,173 report zero views. The next section is about why that split matters more than anything else on this page.

The published benchmarks disagree by a factor of twelve because they disagree about the denominator, not about Instagram. Our best time to post study already worked through why studies of the same platform contradict each other, so one line here: divide by followers and the rate becomes a function of account size, divide by the people who saw the post and account size cancels out. The second version is what makes 245,559 posts from 154,393 accounts we neither own nor sold anything to comparable to each other at all.

Mean against median, with the numbers rather than the assertion. The mean is 3.52% and the median 2.41%. Underneath them the counts are far worse behaved.

A ratio survives outliers roughly. The counts do not survive averaging at all: the mean number of likes on a post in this corpus is 47 times its median. That is why every table below reports medians.

Where the Instagram engagement rate exists at all#

Not one of the 181,779 image posts in this corpus carries a view count Instagram published. The split:

Media format is therefore perfectly confounded with the instrument, with zero exceptions in 181,779 tries. The measured subset is, in effect, a reels benchmark: 245,462 of its 245,559 posts are videos and the handful left carry no media object at all. The image-versus-reel rate gap that every benchmark page publishes is not a platform fact that anyone outside the account could have measured.

The model leaves fingerprints, so this is not an inference from the flag alone. Across the 173,592 posts with a modeled view count, about 52,000, 30.0% of them, compute to exactly 3.0%. The band around that value tells the same story.

A real distribution of human behavior does not stack a third of itself within a tenth of a point of one value, and it does not produce zero zero-engagement posts in 173,592 tries while the published set produces 4,042. The estimator appears to back-solve views from engagement against roughly a 3% prior. One argument from our first draft is retired here: on the small pull the modeled rates stopped dead at 9.18%, and we read that ceiling as part of the signature. At full scale there is no ceiling, the highest modeled rate is 1,049.81%, so the pile-up and the missing zeros carry the finding, and they carry it easily.

The control that separates instrument from format. A control group of 5,358 video posts came through the same modeled path. If the signature belonged to images, those videos would look like the other 245,462 videos. They do not. They pile 18.9% of themselves into the 2.9% to 3.1% band, 14.6% at exactly 3.0%, and not one of them has zero engagement. They carry the image signature, so the signature belongs to the instrument.

Now the part that costs us something. This flag is ours, not Instagram's. It records whether the platform published a number to the surface we read, not whether Instagram knows the number. Instagram made views the primary metric across every format on August 7, 2024, photos and carousels included, replacing plays and impressions, so an account owner does see views on a feed image inside Insights. What nobody outside an account can see is reach on somebody else's feed image, which is exactly why every published image engagement benchmark, including the ones on the search results page above, is denominated in followers instead. Our own first draft of this article published an image-versus-video rate gap. It was an artifact of our own instrument, and we cut it.

Reels against photos: the comparison you can still make#

Whether reels beat photos is a question the rate cannot answer for anyone. The counts can, because likes, comments and shares are recorded for both formats even where views are not.

Median engagements per post, full cleaned corpus (n = 419,024 posts carrying a media object).

Video collects 2.85 times the engagement of an image. Our first pull, fifty times smaller, put the multiple at 2.8. A ratio that holds through a fifty-fold scale-up is the most stable number in this article.

Forty categories hold at least 40 posts in each format, and video leads in 38 of them. The table shows the extremes and both reversals.

Our first pull had one reversal in ten cells, and it was an artifact: a single prolific wargaming account had flooded the cell. At this scale that artifact is gone and two real reversals stand. War and conflict is close to even. Books and literature is not: the median book image collects 164 engagements against 74 for a book video, on 3,088 images. Quote cards and page photos beat videos in the one niche where the thing being posted is itself a still. Which is the argument for reading n and category before reading a multiple.

What this corpus is and is not#

The corpus is 450,324 Instagram posts, every Instagram post tracked by buzzabout research runs, extracted from the production database on July 23, 2026. Of those, 419,151 carry a nonzero view count and 245,559 carry one Instagram published. The measured 245,559 come from 154,393 accounts across 4,212 topic-seeded datasets and were posted between December 2, 2021 and July 23, 2026, nearly all of it recent: 109,824 posts from 2026, 124,904 from 2025, 10,551 from 2024, 280 from everything earlier.

The first version of this article ran on a curated slice of about 2% of this corpus, pulled through the API. This is the same population measured whole. Where a number moved between the two versions, scale and composition moved it, and this page is the corrected record.

Labeling is two-tier, and every table on this page states which tier it stands on. Content intention is labeled on 92% of the corpus and category on 99%, so those tables describe the whole thing. The deep content labels, call to action, hook, tone, narrative structure and language, exist on about 54,000 posts, the 12% that went through a full content-analysis run. That subset is not a random sample. Customers research commercial niches, so it leans promotional, and the unlabeled remainder runs above the corpus median, not below it.

Language shows the tier problem plainly. A language label exists on 30,497 measured posts: English 19,213, French 7,738, Hindi 596, nothing else above 600. The other 215,062 measured posts, 88%, carry no language label, and they are unremarkable in every way we can check: median rate 2.49% against the corpus 2.41%, median views 6,330 against 6,258. We can tell you the labeled slice leans English and French. We cannot tell you what language Instagram speaks.

Carousels cannot be separated. Every post in the corpus carries exactly one media object or none, so a carousel arrives looking like an image.

Concentration is negligible. The busiest account supplies 0.09% of the measured posts, 222 of them, the top ten supply 0.77% together, and 122,959 of the 154,393 accounts, 79.6%, appear exactly once. No single publisher can move a headline number here.

Posts enter this corpus because somebody ran a research query on a niche. It is a sample of topic-driven content, not a panel of Instagram, and every claim here describes the sample.

One reconciliation before a reader finds it. Our best time to post study ran on this same tracked posts table six days earlier, so the corpora are near-identical and there is nothing left to reconcile about size. The difference is what each study needs from a post. Timing needed a timestamp and a view count, which nearly every post carries. This study needs reach Instagram actually published, which exists on 54.5% of posts, and its craft tables need the deep content labels, which exist on 12%. The labels are the constraint: they exist because the instrument reads the post and its comment thread and returns what the post was for, which is a different operation from counting how often a keyword appears.

Does Instagram engagement rate fall as your account grows?#

Not in this data. Split the measured posts into ten equal groups by how many people saw them and the median rate stays inside a band less than half a point wide, from a few hundred views to four hundred million.

Reach deciles, measured subset. Each group holds 24,555 or 24,556 posts.

Median engagements climb from 3 to 11,049 across those same ten groups. The rate moves 0.46 points from its lowest decile to its highest, and the lowest is not where the folk rule needs it to be: decile five, around six thousand views, has the lowest median, not the smallest posts.

The finding is not carried by a few prolific accounts. Take one median per account first, then the median of those 154,393 numbers, and the headline moves from 2.41% to 2.46%.

Two consequences. The by-follower tier table is a picture of its own denominator, not of Instagram. And the folk rule it produces finds no support at either end: the smallest-reach decile sits below the corpus median, so a small audience does not flatter you, and the highest-reach decile has the highest median in the table, so a large one does not punish you.

What the post asked for#

One class of variable moves the number, and it is not craft. It is what the post asked the viewer to do. Intention is labeled on 90% of the measured posts, so this chart describes the whole corpus, not a slice of it.

Sample sizes: emotional 6,557, persuasive 6,499, entertainment 8,604, self-presentation 7,427, engagement 17,017, informational 46,959, promotional 128,717. A further 23,779 measured posts carry no intention label and run at 3.16%, above the corpus median, because the labeling effort concentrates on commercial niches. Promotional is the largest labeled group by a factor of nearly three, 58% of every labeled post, and the worst performing: 1.16 points below persuasive and 1.63 below emotional.

The promotional gap is mostly a property of accounts, not posts. Promotional runs below the rest in 35 of the 37 categories with enough posts on both sides to compare. But set each account against itself, across the 2,519 accounts with at least two promotional and two non-promotional posts, and the median within-account difference is 0.05 points, with the non-promotional posts winning 53.3% of the time (p = 0.001): real, and nearly nothing. Accounts that mostly post promotions run 2.08% against 2.70% for the rest, and that account-level gap of 0.62 points is the whole cross-post gap of 0.59. Selling does not sink a post much. Being a seller sinks a feed. Our TikTok engagement rate benchmarks study runs the same test on nearly three times as many accounts and lands on the same shape.

The call to action says the same thing at higher resolution. This is a deep-label table, 30,760 measured posts, and the useful comparison is against posts explicitly labeled as having no call to action, which is a different set from posts we could not label.

A further 214,799 measured posts carry no call-to-action label at all and run at 2.49%, above the corpus median, for the same reason the unlabeled intention rows do. Below the reporting floor of 40 posts, 1,157 labels covering 2,931 posts are suppressed.

Read the table against its own baseline, and read the n column before the rate: the two cells at the very top hold 41 posts each. Three things survive the scale-up from our first pull. Asking the viewer to do something on the platform, comment, share, tag, follow, still beats asking for nothing, by up to about 0.6 points. Asking the viewer to leave still costs: visit profile 1.63%, click link 1.58%, sign up 1.57%, visit website 1.51%, visit store 1.33%, all below the baseline and below every on-platform ask with real volume. And every ask for money sits at or near the bottom, purchase itself at 1.38% on 1,072 posts; the one purchase-adjacent cell above the baseline, visit shop at 2.37%, holds 44 posts and we flag it rather than lean on it. One thing did not survive: save beat the baseline in our first pull and sits exactly on it now, 2.00% on 240 posts. The stay-or-leave split is the finding. Asking people to stay is free. Asking them to leave is not.

These fields are not Instagram metadata. No native dashboard exports them, and no tool that counts keyword frequency can produce them, because they come from reading the post and its comment thread rather than counting words in it.

The levers that do nothing#

The null results matter more than usual here, because they are what leaves purpose standing. Everything in this section runs on the deep-labeled subset, and each figure carries its n.

Hooks. Forty-three hook labels clear 40 posts. They span 2.60% (bold claim / contrarian, n = 62) to 1.52% (relatability identity, n = 133), a spread of 1.08 points. Posts explicitly labeled as having no hook sit at 1.97% (n = 2,642), below the corpus median rather than on it, which is what the promotional lean of this subset predicts. The two biggest hook families, question (1.84%, n = 4,971) and curiosity / information gap (1.91%, n = 6,735), sit below even the no-hook baseline. A further 901 labels covering 2,431 posts fall below the floor. Across 43 labels, which hook a post used tells you almost nothing about where it lands.

Tone of voice. Our first pull could not build this table at all: no tone label cleared the floor, and the fragmentation was the finding. At full scale 109 labels clear it, and the table that now exists points one coherent way. The top is playful: playful and confident 4.41% (n = 60), playful 3.39% (n = 274), humorous, casual 3.08% (n = 60). The bottom is commercial: professional 1.22% (n = 439), enthusiastic promotional 1.10% (n = 87), promotional 0.87% (n = 49), the worst cell in the section. The two largest cells, enthusiastic (1.69%, n = 2,750) and enthusiastic professional (1.58%, n = 2,239), sit below the median. The fragmentation is still real, 1,912 labels covering 7,265 posts remain below the floor, and most cells above it are small, so read this as a direction rather than a ladder: tone points the same way intention does, away from selling.

Narrative structure. The problem-agitation-solution family arrives under 23 different spellings in the raw labels, from "problem-agitation-solution" through "pas" to "before-after-bridge". Merged, that family covers 5,022 posts and sits at 1.82%, below posts with no detectable structure at all (1.98%, n = 5,146) and well below story or anecdote (2.46%, n = 1,329). That ordering is unchanged from our first pull on six times fewer PAS posts: the most-taught copywriting formula underperforms having no formula, and telling a story beats both. The largest labeled structure, the list, sits at 1.80% on 8,791 posts, down with PAS. The merge rule is a judgment call and the count depends on it, so it is stated here rather than buried: any label containing the problem-agitate stem or the before-after-bridge form.

The best hook beats the worst by 1.08 points. Content intention spans 1.63 points, and the call to action 1.20 between its top cell and purchase. Tone spans more on paper, but its extreme cells hold 49 to 60 posts. Craft, as these labels measure it, moves the number less than purpose does.

One ruler, two platforms#

The incumbents cannot run this comparison. They measure their own customers' accounts, so their TikTok panel and their Instagram panel are different populations of different companies. We measured posts we do not own on both platforms, in a single extract of a single database made on July 23, 2026, with the same formula.

Three findings come out of running both.

The instrument gap. TikTok publishes a view count on 99.98% of posts, Instagram on 54.5%. Instagram benchmarking is broken in a way TikTok benchmarking is not: on TikTok you can measure the wrong thing, on Instagram you frequently cannot measure at all.

The decay law is absent on both. On neither platform does the rate fall as reach grows. Instagram sits inside a 0.46-point band from a few hundred views to four hundred million, highest in the top decile. TikTok is a shallow U: 3.9% in its lowest-reach decile, a trough of 3.5% in the mid-thousands of views, 4.9% in its highest, with one twist at the extreme end: TikTok posts above 10 million views convert reach to engagement worse than any other named view band, 3.26%. Two different platforms, one ruler, and the pattern that every follower-denominated table teaches appears on neither.

Purpose replicates. TikTok medians by intention: emotional 7.20% (n = 18,233), entertainment 6.44% (n = 31,695), self-presentation 6.19% (n = 19,155), persuasive 5.70% (n = 13,621), engagement 5.20% (n = 39,037), informational 4.05% (n = 107,058), promotional 2.86% (n = 213,266). Promotional is last on both platforms, more than a point below the platform median on TikTok and three tenths below it here, and on both platforms the account-level split reproduces the gap.

Our TikTok engagement rate benchmarks study goes deeper on that corpus, including the finding that the account explains roughly six times more of a post's rate than any measurable property of the post. It reports a median of 3.95% on 460,483 posts. Both articles read from the same July 23 extract of the same database, so where they quote each other the numbers agree exactly, with nothing left to reconcile.

What to do with this#

Find your line first.

Then compute your own the same way, or the comparison means nothing:

1. Divide by views, not by followers. A follower-denominated rate describes your account; a reach-denominated one describes your post.
2. Compute it per post, then take the median of your posts. Do not compute it on a total.
3. Use medians everywhere. The mean like count in this corpus is 47 times its median, and a single post that traveled will drag any average you compute.
4. Draw no conclusion from fewer than about 40 posts. That is the floor we used here and it is not generous.

Then stop optimizing the number. The two things it responds to are what you were asking for and who was standing there, and only one of those is craft. The rate tells you whether the right room was shown the post. It cannot tell you who was in the room.

Answering that is a different operation: profiling the accounts that liked and commented, reading what they reshare rather than only what they publish, and grouping them by psychographic similarity instead of by age, on the reasoning that two people the same age are not persuaded the same way. That is the work buzzabout does, on the same six networks these numbers came from.

Frequently asked questions#

What is a good engagement rate on Instagram?#

Against published reach, above 2.41% puts a post in the better half of this corpus, above 4.25% in the top quartile, above 6.77% in the top decile, above 18.81% in the top percent (n = 245,559). Below 1.13% is the bottom quartile. A rate of exactly zero happens to 1.65% of posts, so it is rare rather than shameful.

Why does every Instagram engagement rate benchmark say something different?#

Because they mostly divide by followers, and once you do that the answer depends on how many followers the accounts in your sample had. Here is what the first page of results said on July 22, 2026.

The averages on one page span 0.36% to 4.42%, a factor of twelve, and the same page calls a nano account "high" above 7.97%. Most of that spread is the denominator.

Why do the follower-tier tables show engagement falling with account size?#

Because a rate with followers on the bottom has to fall when the top does not keep up. The tier table above runs from 3.38% to 4.42% at 1k to 5k followers down to 0.75% to 1.01% at 50k to 100k. Two things are worth noticing. First, that same table then goes back up: 0.87% to 1.14% at 100k to 500k, 0.88% to 1.13% at 500k to 1M. A law that reverses inside its own table is a description of a sample. Second, when we split by actual reach instead of followers, the rate stays inside a band less than half a point wide from a few hundred views to four hundred million, and it is highest at the top. The tier tables are measuring account size, which is a real thing to know, and it is not the same thing as how a post performed.

Do reels beat photos?#

On engagement counts in this corpus, yes: a video post collects a median 117 engagements against 41 for an image, 2.85 times as many (n = 250,820 and 168,204). On engagement rate, nobody outside an account can answer it, ours included, because Instagram does not publish reach on a feed image to anyone else: zero of the 181,779 image posts in this corpus carry a published view count. Any page that gives you an image engagement rate next to a reel engagement rate is either using a follower denominator for both or modeling one of them.

How do I calculate my Instagram engagement rate?#

Take one post. Add its likes, comments and shares. Divide by its views, from your own Insights. Multiply by 100. Repeat for every post you want to judge, then take the median of those rates, not the rate of the totals. Compare that median to the percentile table above.

Does engagement rate differ by niche?#

Less than you would expect. The full spread from the highest niche in this corpus to the lowest is a factor of 2.9, still smaller than the factor-of-twelve spread between published benchmarks for the same platform, and the ends carry it: religion and politics at the top, shopping and home and garden at the bottom. The table is in the appendix below.

Why is your number higher than the 0.36% everyone quotes?#

Different denominator, not a different Instagram. That figure divides engagement by followers. Ours divides it by the people who saw the post. On any account whose posts reach fewer than 100% of its followers, and that is nearly all of them, the follower-denominated number is smaller by construction. Neither number is wrong. They are not the same quantity and they should not carry the same name.

Appendix: your niche's line#

Median engagement rate by category, measured subset. Forty labels clear the 40-post floor; two labels covering 34 posts fall below it and are not shown.

Top to bottom, that is a factor of 2.9, wider than the factor of two our first pull showed, and still narrower than the spread the denominator alone creates between published benchmarks.

Method and limits#

Formula. engagement_rate = (likes + comments + shares) / views, computed per post from the raw counts. The database stores the counts and no engagement-rate field, so there is no rounded rate to disagree with: the number is the arithmetic.

Cleaning. 31,173 of the 450,324 posts report zero views. Their rate is undefined, and keeping them as zeros would print a 7.82% zero rate against the true 1.65%, so we dropped them. The 31,173 split into 13,435 modeled images with no number produced, 140 images not flagged as modeled, 17,552 videos not flagged as modeled, 40 modeled videos, and 6 posts with no media object.

The split. The remaining 419,151 posts divide into 245,559 with a view count Instagram published and 173,592 with one our pipeline modeled. Every table above except the format-count tables uses the 245,559. The 173,592 appear only as the distribution shown in section two.

Reporting floor. No cell below 40 posts is published. Suppressed labels and their post volume are disclosed under each table.

The labels. Content intention, call to action, hook, tone of voice, narrative structure and category are produced by reading the post and the comment thread attached to it, not by the Instagram API, which returns none of them. Intention and category cover nearly the whole corpus; call to action, hook, tone, narrative and language cover the 12% of posts that went through a full content-analysis run.

Known distortions, all of them:

• Topic-seeded sampling. Posts enter through research queries on niches. Food and drink is the largest labeled category at 26,168 measured posts, followed by sports and shopping. This is a distribution, not a census.
• No follower count on the post record. The record carries an author name, title and URL and nothing else, so we cannot rebuild the follower-tier table on our own data even to refute it directly.
• Carousels are indistinguishable from images. No post in the corpus carries more than one media object.
• Label fragmentation. 23 spellings of one narrative structure, 1,912 tone labels below the floor, 901 hook labels below the floor. Merges are judgment calls and the only one we made is stated in the visible text.
• The deep-labeled subset leans promotional. 58% of its labeled intentions are promotional, because customers research commercial niches. Every call-to-action, hook, tone and narrative table describes that subset, not the platform.
• Unlabeled does not mean average. The 23,779 measured posts with no intention label run at 3.16% and the 214,799 with no call-to-action label at 2.49%, both above the corpus median of 2.41%. Our first pull had this backwards, its unlabeled rows ran below the median, which was a fact about that slice rather than about Instagram.
• 215,062 measured posts carry no language label. On rate and reach they are indistinguishable from the corpus (median 2.49%, median views 6,330), so language coverage is a labeling gap rather than a hidden population.
• The modeled-views flag is documented as a bare boolean. Its meaning was established empirically, by checking it against sources whose view counts we can verify independently, not from documentation.
• The snapshot is live. The database keeps collecting; this extract is frozen at July 23, 2026. The youngest posts are still accumulating engagement, and the July 2026 monthly cell is the lowest of the recent months at 1.91%, plausibly recency rather than decline.
• Our own first draft was wrong in the way section two describes, and published an image-versus-video rate gap that measured our instrument rather than Instagram.

Data extracted July 23, 2026, in the same pass as our TikTok study's. What these numbers are: population patterns in a topic-driven sample, useful for placing your own work against a distribution, and not a controlled experiment on your account.