Hands are in 45% of live Meta ads, which is awkward if you generate them

We keep a crawl of the Meta Ad Library and label every ad in it — what the layout is doing, what is in the frame, how much text it carries. This is a look at what 4,487 of those ads actually contained on one day in September 2026.

Three things stood out. Hands appear in 45% of them, which matters more than it sounds if any part of your creative is generated. A quarter of tech ads paint a CTA button into the image, on top of the button Meta already renders. And dense text has practically disappeared — only 1.2% of ads land in the heaviest text band.

Before any of that: what this sample is, and the four things it cannot tell you.

What the sample is, and what it isn't

• 4,487 ads from 89 advertisers. Median 51 ads per advertiser. The five largest advertisers are only 7.7% of the total, so no single brand dominates the corpus — but an industry slice here is 5 to 28 advertisers, which is a statement about those advertisers' house styles rather than about an industry. We've left out every industry whose whole slice rests on fewer than ten advertisers. Where a cut inside a larger industry is thinner than that, the advertiser count is printed next to it.
• One day. Every ad was last seen in a single crawl on 6–7 September 2026. This is a snapshot, not a trend line, and nothing here should be read as "currently running".
• One still per ad. This is the big one. For a video ad we classify the poster frame, and for a carousel the first card. So "no text on the image" for a video means "no text burned into the poster frame" — a different claim from "no text in the video". Every number below is split by format for this reason, and where we don't split it, we say so.
• One classifier, one pass. Where two of our fields agree with each other, that is one model being consistent, not two sources confirming something.

And the constraint that governs the whole exercise: the Ad Library shows active ads only and ranks them by impressions, so this corpus skews towards ads that have been running a while at scale. Nothing here says anything about what performs. These are prevalence figures — what advertisers are doing, not what is working for them. Anyone who tells you those are the same thing is selling something.

Hands, in 45% of everything

45.3% of the ads show hands. Split by format: 29.4% of static images, 30.2% of carousels, 62.6% of video poster frames.

The interesting part is the spread between layouts, and it holds inside every format — which means it is a real compositional convention and not an artefact of video being different:

Lifestyle In-Use is the most common layout in the corpus at 22.4% of it. Nearly three quarters of those ads have hands in frame. Product-hero shots — the clean studio product shot, whatever layout surrounds it — are under 10%. The direction holds in all three formats; the size does not. The lift is 2.4× for statics, 2.1× for carousels and 1.2× for video posters, where the base rate is already high. The drop is steadier: 0.31×, 0.21× and 0.14×.

Which is worth sitting with if you generate creative. Hands are the best-known failure mode in image generation, the thing that gives a render away at a glance. And the most common layout in this corpus is the one that puts a hand in frame holding, using or pointing at the product, three times out of four.

The practical consequence is a review one, not a strategic one: if you are generating a product-in-use shot, the hands are not an incidental detail you can skim past — they are the composition, and they need checking every time. They're the third item on our pre-flight list for exactly this reason.

It is also worth noticing which way generated creative tends to drift. The clean product-hero shot is the layout where the hard part can be avoided, and a tool that finds hands difficult will quietly push you towards it. That may be the right call for your product. It should be a call you make, rather than one the renderer makes for you.

A quarter of tech ads draw their own CTA button

99.6% of the ads in the corpus already carry a native Meta call-to-action button — the one the platform renders below the creative. On top of that, 9.8% also have a CTA button painted inside the image itself.

Tech does it more than seven times as often as the lightest categories, and it stays ahead of the all-ads base rate in every format:

• 40.1% of tech static image ads, against 13.7% of all static image ads
• 30.7% of tech carousels, against 12.8%
• 10.0% of tech video posters, against 5.8%
• Across all formats: tech 24.5%, food and beverage 3.4%, fashion 3.5%

Tech's format mix explains almost none of this; it is a genuine difference in how those advertisers compose. (The within-format cells rest on nine to eleven advertisers each, so the strength of the finding is that the same pattern turns up three times independently, not the advertiser count in any one cell.)

It would be easy to call the drawn button redundant. It isn't, necessarily. In Stories and Reels the native button is small and sits in a part of the frame the interface crowds, and burning a button into the composition is a deliberate response to that. What we can say is that it is a house convention in one vertical rather than a general practice — and that if you are generating a tech static and your layout includes a drawn CTA, you are doing what 40% of tech statics do, not something unusual. Across all of tech's formats the figure is 24.5%.

Statics carry burned-in copy. Poster frames don't.

The largest single split in the dataset:

• 3.5% of static image ads have no text in the image at all
• 1.4% of carousels
• 29.2% of video poster frames

So for statics and carousels, text in the image is effectively universal — 98.6% of carousels have some. For video, nearly a third of poster frames are clean.

This is a production-practice difference, and it's the reason we split everything by format. A corpus-wide "15% of ads have no on-image text" number would be mostly a statement about video poster frames, and it would be misleading in both directions.

Dense text is nearly extinct

Sorting ads by how much text they carry, 1.2% land in the heaviest band — 52 ads out of 4,487. Just over two thirds are light, 15.2% are middling and 15.1% carry none at all. The heaviest band holds in every format: 2.2% of statics, 0.8% of video, 0.8% of carousels.

The honest caveat is that "heaviest" here is a judgement against no published rubric, so treat it as a rough band rather than a measurement. But the direction is unambiguous and it survives the format split. Whatever the old text-heavy ad looked like, almost nobody is running it now.

Worth noting what this is not evidence of. There is no text limit on Meta ad images any more. Advertisers are running light text because light text is the house style, not because a rule requires it.

One rarity worth a line

Before / After ads are 0.85% of the corpus — 38 ads. Given how prominent the layout is in ad-craft writing, we expected more.

The tempting explanation is that Meta's health and appearance policies suppress it. The data doesn't support that: health, fitness and beauty advertisers run before/after at 1.37%, roughly double the 0.69% rate everywhere else. It is rarer in the categories with no policy exposure at all. So it is a genuine rarity we have no mechanism to explain, on 38 ads, which is why it is a line rather than a section. (And for what it's worth, Meta's current policy explicitly permits it.)

What we're not going to tell you

Most of what this kind of classification produces is circular. Ads our classifier calls UGC-styled also get called low-polish 92% of the time; ads labelled "Us vs. Them" contain a comparison 100% of the time. Those look like findings and are just one model agreeing with itself. We pre-registered a bar for what counted as surprising — minimum sample sizes, a minimum lift, a minimum number of distinct advertisers, and a cap on how much of any cell could come from one brand — and then threw out everything that cleared it on the numbers but was a tautology on inspection. What's above is what was left.

We also cut several findings that cleared the statistics and failed the sniff test: a serif-font preference among home and garden advertisers (seven advertisers, one house style), prices in finance ads (six advertisers), faces in education ads (course ads contain people). And the language distribution, which measures nothing about Meta and everything about which advertisers we chose to crawl.

The recurring temptation with a dataset like this is to convert prevalence into advice — 71% of top ads use hands, so use hands. That inference doesn't hold. The library shows active ads ranked by impressions, so "common" here means "being run by advertisers who have been running it a while", which is a fact about the industry's habits and not a result. What the numbers are good for is knowing what normal looks like, and noticing when your own output has drifted away from it for reasons nobody chose.

Method

Every ad in a crawl of the Meta Ad Library, classified against a fixed schema: 4,504 rows, of which 4,487 classified successfully (14 skipped, 3 failed). All 4,487 were last seen in one crawl window of about 23 hours on 6–7 September 2026, across 89 advertiser pages. Percentages are of classified ads within the stated format unless labelled otherwise. Four industries whose entire slice rested on fewer than ten advertisers were excluded rather than published; where a cut inside a larger industry is thinner than ten, its advertiser count is stated beside it. No ad images, advertiser names or ad copy are reproduced here or anywhere public — these are aggregate counts only.

---

A snapshot, not a trend, from a sample of advertisers we chose. Every figure is a prevalence count, and none of them is a performance measurement; the Ad Library carries no performance data for commercial ads, so no such measurement was available to make. We'll re-run this on a later crawl and say if it moved.

Terrads keeps the ads it crawls searchable, so finding the ones that look like the brief is a query rather than an afternoon of scrolling.