Four ads in the ad set. One is the hero shot on white. One is the same hero shot on a soft beige. One is that shot with the price badge moved left. And one is a customer filming herself in a kitchen.

You already know which of the four is going to be interesting. The other three slots can only ever tell you about backgrounds.

Short answer: Ad variations are different enough when the winner would change what you do next. If beige beats white, your next brief is identical either way, so that slot bought you nothing. Vary the hook, the offer framing, the format or the proof, and the result has somewhere to go.

The takeaways

A variation earns its slot when the result changes your next brief. Background tints, badge position and font weight fail that test at any budget.
Two near-identical ads cannot pull apart. Each one takes a thin slice of the ad set's weekly conversions, and both get dragged toward the same format baseline.
Meta publishes no similarity threshold and no duplicate detection. There is no measurable distance between two creatives, so "different enough" stays a judgment about what you would do with the answer.

What counts as different enough?

A variation is different enough when the two possible outcomes lead to two different next moves. That is the whole test. Ask what you would brief on Monday if version A wins. Then the same for version B. If the answer is the same brief either way, you built one ad twice and paid for it four times.

Four things reliably pass. The hook, meaning the first thing a scrolling person sees or hears. The offer framing, meaning what the ad says you get. The format: static against video against a person talking to camera. And the proof: a review quote against a demo against a before and after.

Tint, crop, font weight and badge placement fail it. Not because they never matter, but because whichever one wins, you go and make the same next ad.

Why do near-identical ads never separate?

Because each row only gets a thin slice of the same weekly conversions, and two results built on a handful of events sit comfortably inside each other's noise. Meta's own documentation puts the learning phase at roughly 50 optimization events per ad set per week. Split that across four ads and no single row carries enough weight to be told apart from its neighbour.

There is a second reason, and it lives inside the scoring rather than the auction. Any scoring system worth using pulls a thin early result back toward what that format normally does, so a lucky first day does not crown a winner. Adscalr does this with Bayesian shrinkage against format-specific priors.

Two near-identical statics share that baseline. They get pulled toward the same place, and they stay there. That is the correct answer, because nothing about them differs.

What does delivery do with a wall of duplicates?

It picks one and funds it. Meta concentrates spend on whichever ad looks most promising in the first hours, which means a set of near-duplicates hands you one row with data and three that starve before they say anything. That was not a four-way test. That was one ad plus three ghosts you paid for.

It gets worse the more similar the ads are. When the four creatives sit far apart, early concentration at least tells you which direction the audience leans. When they differ by a background colour, the system is picking at random and the report will not say so. I went through those mechanics in why Meta only spends on one ad, and the fix is the same here: give delivery fewer, further-apart things to choose between.

Is there a number for how similar is too similar?

No, and anyone who gives you one made it up. Meta publishes no similarity threshold and documents no deduplication of near-identical creatives, so there is no official line you are crossing. There is also no accepted way to measure the distance between two images in units that predict performance.

Adscalr does not close that gap either. It has no image diff, no similarity score, no duplicate detection. What it does instead is upstream: concepts are generated from competitor angles, audience quotes sorted by awareness stage, your own performance data and fatigue flags, which produces different arguments rather than different tints. The composite score then reads the results across six metrics.

So "different enough" stays a judgment call. The good news: it is a judgment about your own workflow, answerable in ten seconds without a tool.

When is a small variation worth a slot?

When the small element is something you intend to standardise across the entire account. A logo lockup, a caption style, a price badge treatment: if the answer becomes a default that every future ad inherits, the tiny difference is worth buying, because you are amortising it over a hundred creatives rather than four.

Run that as its own deliberate test though. Two rows, not four. Put it on your highest-volume campaign, expect it to take weeks rather than days, and write down in advance what you will change if it wins. I have run maybe three of these in a year, and two of them came back flat, which was still worth knowing because it ended an internal argument. A standardisation test is slow, boring and worth doing. It is not the same activity as filling empty ad slots on a Tuesday because the ad set looked sparse.

Treat the slots as a budget

Every ad slot spends conversions you cannot get back that week. Four slots means four ways to be wrong, and three of them are usually variations of the same idea. Cutting to two properly distinct concepts feels like doing less. It reads faster, delivers cleaner, and leaves you with an answer you can brief against.

Which makes this a creative problem before it is a testing problem: if all you have is one concept and a colour picker, no test design will save the ad set. Building concepts that are far enough apart to be worth testing is what the ad creation side of Adscalr is for, and it is also the part you can do by hand tonight with a competitor's ad library open and your own reviews in a second tab.

How different should ad variations be?