Evidence status: Vendor-published case study; experiment structure disclosed, full statistical appendix not public
The important result was not “Meta works”
Mejuri already knew that Meta could generate attributed revenue.
The problem was that attributed revenue was not answering the question the brand cared about.
According to a case study published by incrementality vendor Haus, Mejuri found that much of the revenue credited to Meta was coming from people who already knew the jewelry brand. That created a harder question than whether ads were producing purchases:
Was Meta creating enough new demand to justify the way the budget was allocated?
The distinction is central to the case.
If a large share of paid social revenue comes from people who were likely to buy anyway, a strong platform ROAS can coexist with weak incremental economics. The answer is not automatically to cut the channel. It is to change what the channel is being asked to do and test whether the new strategy creates more causal lift.
That is what Mejuri did.
The intervention: move part of the budget up-funnel
Haus says Mejuri reallocated 25% of its Meta budget to upper-funnel campaigns and then tested the impact with a geographic holdout design.
The stated hypothesis was specific: upper-funnel Meta activity might bring in more incremental revenue from new customers.
That is a better experiment question than “Does upper funnel work?”
It defines:
- a channel;
- a budget change;
- a customer segment;
- a business outcome;
- and a counterfactual.
The primary KPI in the public case was new revenue.
That matters because Mejuri was not simply trying to increase total attributed sales. It was trying to understand whether the new allocation created revenue that would not otherwise have happened.
The public case identifies the test period as February 22 to March 14, 2025 and describes a two-cell design with a holdout. Haus says Mejuri used DMAs as the geographic unit.
What Haus reported
Haus reports that the initial experiment produced an 11% lift in incremental revenue.
After a two-week post-treatment window, the vendor says the measured lift increased to 12.9%, while iROAS improved by 57%.
Haus also states that overall performance was about 1.5 times better than Mejuri’s original iROAS goal.
Those are material figures.
They should also be read for what they are: results reported by the measurement vendor that designed and hosted the experiment.
The public case page does not expose a complete statistical appendix. It does not provide the exact media spend, market-level outcome table, confidence interval, power calculation or all pre-test balance diagnostics.
That does not invalidate the result.
It limits what an outside reader can independently verify.
The correct editorial interpretation is therefore:
Haus reports that Mejuri’s reallocation toward upper-funnel Meta activity produced a measurable lift in new-customer revenue and a higher incremental return under the tested conditions.
That is stronger than treating the result as a universal benchmark.
Why the post-treatment window matters
The difference between the initial 11% lift and the later 12.9% figure is operationally interesting.
Some advertising effects can occur after the active exposure period.
A person may see an ad, delay the purchase, return through another channel and convert days later. If the experiment stops observing outcomes the instant media exposure ends, part of the effect can be missed.
A post-treatment window attempts to capture that delayed behavior.
But longer windows create their own problems.
The further the measurement moves away from the controlled exposure period, the greater the chance that other events influence the outcome: promotions, seasonality, competitor activity, organic demand or changes in other media.
There is no universally correct post-treatment period.
The important lesson is that the conversion cycle should influence the experiment window before the test begins.
The real media-buying lesson is allocation, not attribution
The strongest part of the case is not the reported lift.
It is the decision process.
Mejuri did not respond to disappointing incrementality by simply lowering Meta spend.
It changed the role of the budget.
If lower-funnel Meta was capturing demand that already existed, shifting part of spend toward upper-funnel activity created a chance for the platform to reach people who were less likely to convert without advertising.
That is a portfolio decision.
A simplified operating sequence looks like this:
- Measure the incrementality of the current mix.
- Identify where attributed performance and causal performance diverge.
- Form a hypothesis about why.
- Reallocate budget toward a different job.
- Run a controlled test.
- Compare the incremental economics.
- Keep, reverse or iterate the allocation.
That process is much more transferable than Mejuri’s exact percentage.
Why 25% should not become a benchmark
The public case makes the 25% reallocation memorable.
It should not become a rule.
A different advertiser may need to move 10%, 40% or none at all.
The correct size depends on:
- current channel mix;
- existing upper-funnel spend;
- geographic scale;
- brand awareness;
- conversion volume;
- margin;
- available test markets;
- experiment power;
- customer purchase cycle.
The treatment must be large enough to create a detectable difference without becoming economically reckless.
That is an experiment-design problem, not a media-planning slogan.
Failure mode: using platform ROAS as the experiment KPI
If Mejuri had evaluated the test only through Ads Manager, the experiment would have answered a different question.
Platform ROAS tells the operator how the platform assigned conversion credit under its attribution rules.
A geo holdout attempts to estimate the difference between what happened with the treatment and what would have happened without it.
Those numbers can diverge.
A platform can report strong ROAS while the holdout shows limited lift.
A platform can also under-credit upper-funnel activity that generates conversions later through Search, direct traffic or another channel.
For a test designed around incrementality, the experiment KPI must sit outside the platform attribution model.
Failure mode: changing creative, budget and audience logic at once
Media-mix experiments become harder to interpret when the treatment changes too many things.
If an advertiser simultaneously introduces new creative, raises total spend, changes audiences and adds an upper-funnel objective, a positive result may be real but difficult to attribute to one decision.
In Mejuri’s public case, the central intervention is described as reallocating part of Meta spend toward upper-funnel campaigns.
That is a relatively legible treatment.
Operators should protect that clarity.
The more strategic the question, the more disciplined the treatment should be.
What the case does not prove
The result does not prove that:
- upper-funnel Meta always outperforms lower-funnel Meta;
- 25% is the correct allocation for other brands;
- a 57% iROAS improvement is reproducible;
- GeoLift is the only valid measurement approach;
- platform attribution is useless;
- a single experiment permanently determines the correct media mix.
Incrementality is conditional.
The result can change as creative, competition, brand awareness, seasonality and channel mix change.
That is why Haus itself frames incrementality as a continuous practice.
On that point, the vendor’s commercial interest and the operational reality happen to align: one experiment is a decision input, not a permanent truth.
How a media team could apply the method
A team does not need Mejuri’s scale to adopt the logic.
The minimum framework is:
Question: Which part of paid media may be receiving more attribution credit than causal credit?
Business KPI: New revenue, contribution margin, qualified customers or another outcome outside platform reporting.
Treatment: One clear media change.
Counterfactual: A valid control or holdout.
Window: Long enough to capture the relevant purchase cycle.
Decision threshold: Define in advance what result would justify keeping the change.
Follow-up: Retest after meaningful changes in spend, creative or market conditions.
The most useful outcome is not a beautiful lift chart.
It is a better rule for where the next dollar should go.
Evidence note
This analysis relies primarily on a case study published by Haus, the measurement vendor used in the experiment. The public page discloses the treatment concept, KPI, dates and headline results, but does not provide a full statistical appendix. Radar therefore treats the reported 12.9% incremental-sales lift and 57% iROAS improvement as vendor-reported case results, not independently audited findings.
