Every advertiser is looking for a unicorn.
By a unicorn ad, I don't just mean an ad that posts a great ROAS for a few days. I mean the rare ad that can absorb a significant amount of spend, remain profitable, and keep doing it for a long period of time. Its performance does not just edge out the typical ad. It can outperform an ordinary ad many times over and become responsible for a disproportionate share of an account's profitable growth.
These ads exist. And when you find one, they can transform an advertising account.
So every advertiser wants to know how to create their next one. Find the right hook. Test three headlines. Try two images. Compare the results. Identify what worked. Then use those lessons to manufacture another winner.
It sounds scientific. It sounds disciplined.
The problem is this does not work as well as you'd expect.
You probably can't predict your next unicorn ad.
Not because creative strategy doesn't matter. It absolutely does. The problem is that you don't control enough of what happens between creating the ad and seeing the result.
You Are Not Actually Testing What You Think You Are
A meaningful A/B test requires control. Ideally, you take two ads, expose them to comparable groups of people under comparable conditions, change one variable, and measure the difference.
That is not necessarily what happens when you test creative on Meta.
You control the creative you submit. Meta controls a tremendous amount of what happens next: who sees the ad, when they see it, where they see it, how often they see it, what other ads are competing for that impression, and how aggressively each ad is delivered based on Meta's prediction of who is likely to respond.
Don't Just Take My Word For It — Listen to Meta's Scientists
A 2025 paper titled Characterizing and Minimizing Divergent Delivery in Meta Advertising Experiments was written by six data scientists, including three from Meta. They examined 181,890 Meta A/B tests. The researchers documented what they call divergent delivery: different variants in an A/B test can wind up being delivered to meaningfully different groups of people.
That matters because the measured result can reflect two things at once — the effectiveness of the creative, and changes in the composition of the audience that actually received it.
Even when the researchers restricted their analysis to tests with the same target audience, budget, bid configuration, and optimization goal, meaningful audience imbalance remained. It was greater for conversion-optimized tests than for awareness-optimized tests. That makes intuitive sense: when the objective is conversion, Meta's delivery system is actively trying to find the users it believes are most likely to take the desired action.
It's virtually impossible to test creative while controlling for audience.
So when Creative A beats Creative B, what exactly did you learn?
Maybe Creative A really is better. Or maybe Meta found a better pocket of customers for it. Maybe Creative B was shown to a less receptive group. Maybe the two ads received different exposure patterns at different times of day.
The paper does not say A/B testing is useless. In fact, it shows that divergent delivery can be reduced under tightly controlled conditions. But those conditions are very different from the way most advertisers normally run performance campaigns.
The test may not be cleanly answering, "Is Creative A inherently better than Creative B?" It may be answering something closer to, "How did Creative A perform with the people Meta chose to show it to, compared with Creative B and the people Meta chose for that one?"
Those are not the same question.
Be Skeptical of Anyone Who Knows "What Converts Now"
You will often hear agencies, consultants, and creative platforms claim they know what converts now. Maybe it is UGC. Maybe it is founder-led video. Maybe it is a particular hook, visual style, opening, or ad format.
There may be useful patterns in those observations. Good advertisers should absolutely pay attention to them. But the confidence behind the claim is often much stronger than the evidence supports.
If Meta can deliver different creatives to different slices of the available audience, and if performance is also influenced by novelty, timing, competition, consumer mood, and changing market conditions, then yesterday's winning format is not a reliable recipe for tomorrow's winning ad.
The data does not support a high level of certainty on creative testing.
At best, "what converts now" gives you another informed hypothesis worth testing. It does not tell you what your next unicorn will be.
There Is More Noise in the System Than We Want to Admit
None of this means advertising is random. Better ads increase the probability of success. Better offers matter. Better positioning matters. Understanding the customer matters.
But the performance of any individual ad contains a significant amount of uncertainty. The advertiser does not control the auction, the surrounding competitive environment, the exact users selected for delivery, the order in which they encounter messages, or how the market feels on a particular day.
This is why post-hoc explanations can become dangerous. An ad wins and everyone explains why: "It was the hook." "It was the UGC format." "It was the first three seconds." "Customers want founder-led videos now."
Perhaps. But if those explanations were reliably predictive, advertisers would be able to take the lesson and consistently manufacture the next winner. Anyone who has run enough creative knows how hard that is. Sometimes the ad everyone expects to win dies immediately. Sometimes the quick variation somebody almost did not upload becomes the account's best performer.
Think About Rolling a Seven
Suppose your job is to roll a seven with a pair of dice.
You could spend hours studying the physics of dice rolling. How hard should you throw them? Should they bounce off the wall? Should you shake them first? From what height should you release them? Is there an ideal motion of the wrist?
Maybe technique can affect the outcome at the margins. But there is a much simpler way to increase the number of sevens you roll:
Roll the dice more often.
Advertising works much the same way. There is value in studying results and improving the quality of your attempts. But at some point, another hour spent explaining yesterday's losing ad may be less valuable than using that hour to create tomorrow's new one.
The objective is not to stop thinking. It is to recognize when analysis has reached the limits of what the data can actually tell you.
This Does Not Mean "Throw Stuff at the Wall"
There is an important distinction here. Saying you cannot reliably predict which ad will become a unicorn does not mean creative quality does not matter. It certainly does not mean you should crank out random ads and hope something works.
Every ad still has to clear a basic standard:
- It needs to stop the scroll.
- It should call out your ideal customer.
- It must make the pain point or dream outcome clear.
Those standards improve your odds. They eliminate plenty of bad ideas before you spend money on them. But they do not let you look at five competent ads and reliably identify which one will absorb six figures of spend over the next six months.
Your Best Idea Might Fail the First Time
This leads to another conclusion advertisers often resist: sometimes you should retry an idea that failed.
That is especially true when you have a strong reason to believe the underlying concept is good. Maybe customers repeatedly mention the problem in reviews. Maybe it is one of the strongest reasons people buy the product. Maybe the demonstration makes the value proposition unusually clear. Maybe the same message works extremely well in email, on the product page, or in sales conversations.
And yet the ad gets a poor first run on Facebook.
The usual conclusion is: "We tested that. It didn't work."
But what did you actually test? You tested one execution, during one period of time, with one particular delivery outcome from Meta. That is useful evidence, but it is not necessarily a final verdict on the underlying idea.
Try a different visual. Change the opening. Turn the static image into a video. Rework the execution. Or, if you still believe in the idea, simply give it another run later. Sometimes the creative was wrong. Sometimes the timing was wrong. Sometimes Meta's first attempt at finding an audience for it simply did not produce a winner.
Change the Question
The biggest implication is not about how you analyze ads. It is about how you organize creative production.
Stop asking:
"How do we figure out what our next winning ad will be?"
Instead ask:
"How do we build a system that gives us enough good attempts to regularly discover winners?"
That changes advertising from a prediction problem into a production-and-discovery problem.
Your creative workflow should consistently produce legitimate new attempts: new hooks, new visual treatments, new formats, new demonstrations, new pain points, new dream outcomes, new objections, new customer stories, and new versions of ideas that deserved another chance.
Some will fail. Most may fail. That is not necessarily evidence that the process is broken. The goal is not to eliminate losing ads. The goal is to make losing experiments relatively inexpensive while giving the winners enough room to scale. A true unicorn can pay for a lot of unsuccessful rolls along the way.
Stop Trying to Predict the Unicorn
You should learn from your results. Develop hypotheses. Understand your customers. Improve your creative. Look for patterns. Use data to make better decisions. Just do not confuse patterns with certainty.
Your next unicorn ad is probably impossible to predict in advance. And that means you should not build a creative process that depends on predicting it. Build one that keeps rolling the dice.