An overseas marketing team asks its Chinese production partner to repeat the opening of video B. B finished the campaign with the higher overall click-through rate. But the placement breakdown tells a different story: video A had the higher rate in each placement.
That is possible without an arithmetic error. The videos received different mixes of impressions. Before a report becomes a production instruction, distinguish what happened to the delivered campaign from what you have learned about the creative.
A wider distribution system changes the interpretation of a score
TikTok For Business announced on October 5, 2026 that TikTok Ad Network was generally available for U.S. targeting. The announcement describes optimization across eligible TikTok and network inventory, shifting budget toward placements with stronger results. This describes the vendor's delivery mechanism, not a forecast for a client or a promise of identical account settings.
ONCE's inference is straightforward: when distribution can change, an aggregate creative score reflects more than the video. Audience selection, placement and delivery arrangements can also affect it. A valid total is not automatically an answer to “Would this opening work better if we filmed it again?”
A hypothetical reversal, calculated in full
Every number below is invented to explain the arithmetic. These are not account results, benchmarks or an ONCE test. Placement X and placement Y are generic labels; neither represents a claim about a real format's typical performance. Assume a consistent definition of impressions and clicks.
| Video | Placement | Impressions | Clicks | Clicks / impressions |
|---|---|---|---|---|
| A | X | 9,000 | 180 | 2% |
| A | Y | 1,000 | 50 | 5% |
| B | X | 1,000 | 10 | 1% |
| B | Y | 9,000 | 360 | 4% |
| A total | All | 10,000 | 230 | 2.3% |
| B total | All | 10,000 | 370 | 3.7% |
A leads in X and Y, yet B leads overall because most of B's impressions occur in the higher-rate placement. Aggregate CTR is total clicks divided by total impressions, equivalent to weighting each placement's rate by its share of impressions. It is not the simple average of placement rates.
A reversal between grouped and aggregate patterns is a form of Simpson's paradox. It does not establish a platform error or prove that A has a causal creative advantage. It shows why the delivery composition needs explaining.
ONCE's three-view review: actual, segmented and common-weight
Keep the actual campaign result
Retain total impressions, clicks, spending and business outcomes. In this illustration, B really produced more clicks within its delivered combination. Do not erase that result because another view favors A.
The actual view answers what the campaign delivered. Choosing what to film next is a different question and needs more context than an aggregate leaderboard.
Inspect shared placements and retain the unknowns
Group results by stable creative ID and placement, preserving impressions, consistently defined clicks, spending and available business outcomes. Check dates, country, audience, optimization objective, product and offer for other differences.
TikTok's current custom-report documentation lists dimensions such as ads and placements and metrics including CTR. Available combinations and coverage still depend on what the account actually returns. Missing data is not zero. Unclassified inventory belongs in the reconciliation rather than disappearing from a preferred view.
If one video ran only in X and the other only in Y, these records do not provide a within-placement comparison. Do not fabricate the missing cells to complete a neat table.
Use one declared distribution for a diagnostic comparison
Where both videos have data, choose and state a common set of placement weights. Calculate the sum of each placement's CTR multiplied by its common weight. Weights must sum to 100 percent and have an explicit rationale, such as a planned delivery mix or an agreed reference distribution. Do not select them after seeing which creative they favor.
For illustration only, equal weights for X and Y yield 3.5 percent for A and 2.5 percent for B. Those are hypothetical common-mix figures, not actual delivery, a forecast or replacement values for the platform report.
This removes only the difference in placement weighting. It does not remove audience selection, timing, auction conditions or other differences within placements. Calling it the “true creative lift” would overstate what was calculated. With little shared coverage, sparse data or unstable rates, preserve uncertainty instead of manufacturing a ranking.
A planned experiment requires its planned interpretation
Optimizely's official discussion of Simpson's paradox warns about changing variant distribution during an experiment and recommends using segments to develop further testing ideas rather than making release decisions from selected subgroup results. These are experiment-interpretation principles, not TikTok account instructions.
If you have a properly designed and executed randomized experiment, interpret it for its predeclared population, primary metric and analysis plan. Do not overturn the overall result because a post hoc segment or new weighting looks more attractive. The method here diagnoses mixture in routine, nonrandom delivery reports; it does not replace a causal experiment.
Write the uncertainty into the next production brief
For this example, a defensible note is: “The ranking changes with placement composition; the creative explanation needs testing.” It is not “B's opening always wins.” A further comparison could focus on a shared placement and the intended creative variable while holding audience, product, offer and destination as stable as the design requires.
Determine sample and analysis needs from actual traffic and the question, not a universal impression threshold. Higher CTR also does not establish more purchases or qualified leads. Evaluate the outcomes and costs that match the business objective.
For teams commissioning product footage from China, this prevents a media-distribution advantage from becoming an unsupported filming rule. ONCE's video marketing articles offer related campaign-planning perspectives.
A durable creative record should retain version identity, delivery mix, comparison conditions, observed results and unresolved explanations, rather than only a winner badge. For ONCE, this is a proposed direction for reusable creative knowledge, not a claim of achieved client lift or higher business valuation.
Sources and scope
Checked October 30, 2026: TikTok For Business, Now available for U.S. targeting: TikTok Ad Network (October 5, 2026); TikTok Business Help Center, How to create a Custom report (updated December 2025); and Optimizely, Simpson's Paradox: Discover possibilities with your segments, not shipping decisions (no publication date shown).
The tables, weights and review method are ONCE analysis using hypothetical data. No advertising account was accessed or changed. The original AI cover is a production illustration, not a client scene.
