Blog

Original research · Published 2026-10-30 · Updated 2026-10-30 · 4 min read

Does Carousel Design Affect Reach? We Tested Nine Variables

Within a consistent template, we found no evidence that visual design predicts which carousel wins. We compared nine design variables between 33 outlier posts and 37 ordinary ones using 5,000 shuffles, and none reached p below 0.05; the closest was p = 0.091. Consistency and the idea itself are better bets.

TL;DR: Raw gaps looked suggestive until we ran the test. What did hold up is descriptive: slides are either dark or light, most carry no loud accent colour, and covers carry about three times the type of inside slides.
A permutation test asks whether a gap between two groups is bigger than random reshuffling of the same data would produce. A result below p = 0.05 would conventionally count as significant.

What did we test?

We measured 1,064 slides from 127 carousels by 18 accounts in AI, tech, business and design niches, then compared nine design variables between posts that beat their own account's median by 1.5 times or more (33 posts) and posts near the median (37 posts). We shuffled the labels 5,000 times per variable. Not one variable was significant at p below 0.05, and the closest, a split-field layout measure, reached p = 0.091.

The raw gaps had looked meaningful, and we would have reported them as findings had we skipped the test. That is the useful lesson: within an established template, visual variation did not predict which post won. This is correlational, small, and from one niche cluster, so it does not prove design never matters. It says that tuning small design details is a poor use of time once your template is competent.

What does hold up in the data?

These findings are descriptive, not claims about what wins.

Slide brightness in 1,064 measured slidesShare of slides by brightness: 31.2 percent dark, 10.6 percent mid-tone, 58.2 percent light. Few slides sit in the middle.31.2%Dark slides10.6%Mid-tone58.2%Light slides

Share of slides by brightness class (dark below 0.35 average luminance, light above 0.65). 332 dark, 113 mid, 619 light.

What did we get wrong first?

We believed our own decks were not dense enough. Measured with the same code, our detail swing was 26.7% against the market's 14.4%, and our luminance swing was 42.6 to 73.8 against 18.5, so we over-swung by 2.3 to 4 times. We softened the inverted ground in one style from a very dark brown to a lighter one for that reason. A second near-miss: one account hides its like counts, and our parser read the missing counter as zero, producing a false median of 0 until we excluded those 15 posts.

Is slide 2 treated as a second cover?

Mostly not. Across 117 decks slide 2 was darker than the cover in 52% and denser in 42%, with median differences near zero. This measures execution, not whether it would perform better. See size and slide count for the other corpus numbers.

So what should you spend time on?

  1. The idea and the hook. The test says design tweaks are not where the variance lives.
  2. One locked template with a contrast check.
  3. Volume with variation in ideas, not in decoration.

That is how our carousel workflow is built.

Frequently asked questions

Does carousel design affect Instagram reach?

In our test of nine design variables on 70 posts from 18 accounts, none was significant at p below 0.05, so we found no evidence that small design differences predict winners within a template.

Should carousels be light or dark?

Real slides split: 31.2% dark, 58.2% light, 10.6% in between. The advice we draw is to pick one and commit, not to aim for the middle.

How is a carousel cover different from inside slides?

Mainly in the amount of type: covers had 27.7% ink share against 9.0% for inside slides. Differences in texture and brightness were close to coin flips.

How big was your sample?

1,064 slides from 127 carousels and 18 accounts for the descriptive findings, and 33 outlier plus 37 ordinary posts for the significance test. It is small and limited to AI, tech, business and design accounts.

The Content Engine researches a niche, writes post ideas and renders carousels (PNG) and reels (MP4) on your own computer. USD 67 one time, 14-day refund. It needs a paid Claude plan (Pro or higher), Node.js and FFmpeg, and we have tested it on Windows only.

See the Content Engine

Sources and our data

Keep reading

Results vary. No promise of views, followers or income. The Content Engine is an independent product, not made or endorsed by Anthropic. Claude is named only as the tool it runs on.