Marketing is a search problem

In 2026, Optimizely looked back over 173,000 experiments run on its testing platform. Only 12% produced a significant improvement.

None of these were random guesses. Each was an idea a team believed in enough to build and ship.

That one number changes the job. If the people closest to the work are wrong seven times in eight, winning depends less on choosing well and more on how many ideas get a real test.

So how many ads should a team test? As many different ideas as its traffic can judge. In Motion's 2026 benchmark, mid-sized advertisers test about seven new ads a week and the largest about nineteen.

Nobody can pick the winner

The 12% is no outlier. Published estimates put the share of experiments that work at about one in ten at Booking.com, Google Ads and Netflix.

Figure 1

A median of one idea in ten works

Share of experiments that succeeded

Microsoft33%

Bing15%

Booking.com10%

Google Ads10%

Netflix10%

Airbnb Search8%

If the best-instrumented teams in the world cannot call a winner in advance, a creative review meeting cannot either.

And the few winners carry almost everything

Low odds would matter less if every win were the same size. They are not. In AppsFlyer's 2025 study of 1.1 million video ads, the top 2% of ads took 43% to 53% of the spend.

Rare hits, and a very large prize for each one. That is what a search looks like. In a search, the result depends on how much ground gets covered.

So the accounts with the most winners test the most

Motion's 2026 benchmark shows it inside ad accounts: 578,750 Meta ads from 6,015 advertisers.

Figure 2

The hit rate stalls. Testing keeps climbing.

Meta advertisers, by monthly ad spend

New ads tested a weekShare that win

Under $10K2.84.0%

$10K to $50K4.16.4%

$50K to $200K6.68.1%

$200K to $1M11.28.6%

Over $1M18.88.8%

Look at the right side first. The hit rate, the share of new ads that become winners, climbs to about 8% and then stops. A budget twenty times larger buys almost no improvement in the odds of any single ad.

Now the left side. Bigger accounts simply test more, up to nineteen new ads a week. With the odds per ad nearly flat, more winners can only come from more attempts.

The platforms' own guidance points the other way. LinkedIn suggests four or five ads in a campaign.

The ceiling was the hours behind each ad

Why do teams run five ads when the evidence says twenty? Because each one costs a brief, a designer, a round of review and a build. The limit on velocity was never ideas. It was the hours behind each ad.

AI is removing that cap. An extra version no longer needs an extra afternoon.

What AI changes, and what it does not

If making were the only limit, the advice would be to launch a thousand ads. Meta's help centre says otherwise: "too many ads can result in worse performance."

The reason is evidence. A budget buys a fixed number of results, and every extra ad takes a share of them. Meta's own researchers described the new situation in July 2026: "Ad creative optimization is increasingly constrained by evaluation rather than generation."

Making ads has become cheap. Finding out which one is good has not.

There is a second catch. In an MIT field experiment that produced 11,024 real ads, the people paired with AI turned out, in the authors' words, "more homogeneous outputs." Twenty rewordings of one idea are a single test in twenty outfits.

So velocity is not volume. It is the number of different bets an account can make and still judge. That is a job for a loop: quality supplies ideas that deserve a test, velocity gets enough of them into the market, and optimization separates the real winners from the lucky ones.

What to do on Monday

1. Count the new ads launched in a typical week. New ideas only. A resize or a new button colour does not count.
2. Work out the evidence budget. Divide the month's test conversions by the number of new ads. A handful per ad means too many ads. Hundreds per ad means most of the search is not being run.
3. Make the ads different from each other. Before a batch goes live, write one line per ad saying what it bets on. If two lines read the same, that is one test.

More tests find more winners. They also produce more ads that only look like winners. When a dashboard shows a winner, how can anyone know it won? That is optimization, and the next article.