Creative Analytics: The Four-Rung Metric Ladder That Tells You Which Ad Actually Won

Same Budget, Same Audience, 56% Cheaper

When Google's own team ran its Pixel 7a campaign in Germany, it did not change the targeting or the budget to move the number. It changed the creative. By the four-week mark, the creator-led skippable ads were showing 49% higher watch time, 128% higher consideration, and a cost per lifted user 56% lower than the control. One variable moved. It was the one inside the video.

That result is unremarkable to anyone who buys media. What is remarkable is how few teams can reproduce the analysis. Ad platforms report by campaign, ad set, and ad ID. They do not report by hook, by actor, by opening shot, or by angle, because they do not know what those are. So most advertisers can tell you which ad spent the money, and almost none can tell you which idea earned it.

Creative analytics is the discipline of closing that gap: measuring the advertisement rather than the channel it happened to run in. This post is about which metrics actually read at that level, in what order, and how long you have to wait before each one is telling you the truth.

Why channel dashboards cannot answer creative questions

The unit of analysis problem

Every reporting surface you have is organized around the thing the platform bills for. That is the ad ID. It is a container, and the container knows nothing about its contents.

Two ads in the same ad set might share a hook and differ only in the actor. Two more might share the actor and differ in the offer. To the dashboard, these are four unrelated rows sorted by spend. The one comparison you actually want to make (does this hook beat that hook, across every ad we ever put it in) is not a query the platform can serve, because nobody told it what a hook is.

This is why creative analytics starts before the reporting and not after it. If the attributes are not recorded at the moment the variant is produced, there is nothing to group by later, and every test result dies with the ad that produced it.

What automated delivery quietly removes

Modern campaigns make this harder, not easier. Automated bidding and broad targeting reallocate budget toward whatever is converting, sometimes within hours. That is good for your account and terrible for your read: by the time you open the report, the platform has already starved the variants it disliked and pushed spend into the ones it liked, so the sample sizes are wildly unequal and the comparison is contaminated.

Anyone reading raw performance under automated delivery is partly reading the algorithm's opinion rather than the audience's. That is exactly why controlled split tests still matter, and why Reddit's self-serve split testing and its equivalents are worth the setup cost when a decision is expensive.

The three questions worth building for

Strip away the dashboards and creative analytics answers three questions campaign reporting structurally cannot:

• Which attributes of our creative correlate with performance, across every ad that carried them?
• How long does a winning creative stay a winner before it decays?
• What share of everything we produce ever clears the bar at all?

The rest of this post is the measurement apparatus for those three.

The Creative Read Ladder

Here is the frame I use, and the only genuinely new idea in this post: the Creative Read Ladder. Four rungs of metric, ordered by how decisive each one is. The higher rungs arrive fast and diagnose. The lower rungs arrive slowly and decide. Almost every bad creative call comes from reading a rung before it had the volume to discriminate, or from treating a diagnostic rung as a verdict.

Rung	Metrics	Arrives	What it is for
1. Delivery	Impressions, CPM, frequency	Hours	Test validity
2. Attention	2s and 6s views, 25/50/75/100% completion	Same day	Diagnosing the hook
3. Intent	Click-through rate, landing page views	1 to 3 days	Diagnosing the promise
4. Outcome	Cost per acquisition, ROAS	1 to 4 weeks	Deciding

Rung 1: delivery, the sanity check nobody runs

Before comparing anything, confirm the comparison is legal. Did both variants get comparable impressions? Is frequency similar? Did one variant get an unlucky placement mix?

This rung never tells you a creative is good. It tells you whether the next three rungs mean anything. Skipping it is how teams end up confidently ranking a variant that received one fifth of the delivery.

Rung 2: attention, the fastest honest signal

Attention metrics are the closest thing to a pure creative read you get, because they are almost entirely determined by the asset and barely at all by your offer, price, or landing page.

They are also precisely defined, which helps. TikTok counts a 6-second video view as a play of at least 6 seconds, a full play if the video is shorter, or at least one engagement inside the first 6 seconds, and it reports completion at the 25%, 50%, 75% and 100% marks. Read those thresholds as a hook autopsy:

• Drops before 25%: a hook problem. The opening did not earn the next second.
• Holds to 50%, collapses before 75%: a middle problem, usually a claim that arrived without proof.
• Holds to 100%, converts poorly: not an attention problem at all. Move down the ladder.

Attention is diagnostic, not decisive. Plenty of ads hold attention beautifully and sell nothing.

Rung 3: intent, the most misread rung on the ladder

Click-through rate sits in the awkward middle: fast enough to feel actionable, far enough from revenue to mislead. It is genuinely useful as a paired read with attention (high hold and low CTR usually means the video entertained without ever making an offer), and genuinely dangerous as a standalone verdict, which is a case worth reading in full in does a high CTR mean a good ad. If you need the mechanics and the levers, what CTR is and how to improve it covers them properly.

Treat rung 3 as a diagnosis of the promise your creative made, not of the business it produced.

Rung 4: outcome, the only rung that pays rent

Cost per acquisition and ROAS are where creative decisions should actually be made. They are also the slowest and noisiest, because they depend on conversions, and conversions are rare events. That tension is the entire problem of creative measurement, and it is what the next section is about.

The sample-size gate: when a rung is allowed to talk

Why early reads flip

Conversions are rare, so early conversion data is mostly noise. With three purchases on variant A and one on variant B, the "300% better" variant is a coin flip wearing a suit. The uncomfortable arithmetic is that the metric you most want to decide on is the metric that takes longest to become trustworthy, and the metric that arrives first is the one you least want to decide on.

The discipline is not to decide faster. It is to decide on the highest rung that has enough data, and to be explicit that you did.

Decision latency, rung by rung

Give each rung an honest waiting period before it is allowed into a decision:

• Delivery: valid within hours. Check it immediately, then stop looking.
• Attention: stable after a few thousand impressions per variant, usually the same day.
• Intent: stable after a few hundred clicks per variant, typically one to three days.
• Outcome: stable only after enough conversions per variant to survive the noise, often one to four weeks.

Google's Pixel 7a creative experiment was read at the four-week mark, and that is a team with more traffic than almost any advertiser reading this. If a global campaign waits four weeks for a clean creative verdict, a brand spending a few hundred dollars a day is not getting one in 48 hours.

The practical compromise

Most advertisers cannot fund a statistically clean outcome read on every variant, and pretending otherwise produces confident nonsense. The workable pattern is a two-stage decision:

• Cut on attention. Kill the bottom of the pack on hold and completion, where data is cheap and abundant, within a day or two.
• Decide on outcome. Spend the volume required for a conversion read only on the handful of survivors that already earned it.

That is a real methodological concession, and it is better than the alternative, which is a portfolio of decisions made on four conversions each.

Attribute-level reporting: making results compound

The five tags that do most of the work

Attribute tagging is the difference between a company that learns and one that just spends. Record these at production time, not at reporting time, because after the fact nobody remembers:

• Hook type: problem-first, result-first, question, demonstration, unboxing.
• Actor profile: the persona, not the person. Age band, presentation, delivery style.
• Format: talking head, voiceover over product, split screen, screen recording.
• Angle: the specific claim or benefit the ad argues.
• Length band: under 10 seconds, 10 to 20, 20 to 40.

Five fields, recorded consistently, turn a pile of unrelated ad rows into a queryable dataset. After thirty or forty creatives you can finally ask the question that matters: not "did this ad win" but "do our problem-first hooks beat our result-first hooks, and by how much."

Why the one-variable rule bends on UGC

Classic testing doctrine says change one thing at a time. It is good advice that partly breaks on user-generated-style video, because a different actor inevitably delivers the same script differently, with different pacing, framing, and energy. There is no such thing as an actor swap that holds everything else constant.

The honest response is not to abandon control. It is to accept that individual UGC tests measure bundles, and to recover precision at the portfolio level instead: with enough tagged creatives, an attribute that genuinely works shows up repeatedly across many different bundles. One test cannot isolate the hook. Forty tagged tests can.

Keep the ledger separate from the workflow

Recording attributes is a measurement decision, not a process one. The brief templates, review gates, and versioning that surround it belong to running creative as an operation. What belongs here is narrower and non-negotiable: whatever your workflow looks like, the attributes must land in a field you can group by, not in a filename or somebody's memory.

Winner Rate: measuring the system instead of the ad

The definition

Winner Rate is the share of creatives you launch that clear your cost-per-acquisition bar. It is a portfolio metric, and it is the one number that describes your creative operation rather than any individual ad.

It matters because the alternative framing quietly misleads. "Our best ad has a 4.2 ROAS" tells you nothing about whether you can produce another one. Winner Rate does.

A worked example

A skincare brand launches 24 creative variants in a quarter. Three of them beat the CPA bar and get scaled. That is a 12.5% Winner Rate, which means the real production question is not "how much does a video cost" but "how much does a winner cost."

The cost-per-winner equation: cost per winner = cost per creative × creatives launched per winner.

At 12.5%, every winner costs eight creatives. If a creative costs $200 to produce, a winner costs $1,600 and you will ration your tests. If a generated variant costs around $2, the same winner costs about $16, and rationing stops being the constraint. The Winner Rate did not change. The economics of discovering it did, which is the whole argument for cheap variation in improving ROAS with UGC.

Reading the trend, not the number

There is no benchmark Winner Rate worth chasing, because yours is a function of your own bar and your own category. Compare it only against your own history. Rising volume with a flat Winner Rate means you got cheaper, not smarter. A climbing Winner Rate means your analytics are actually feeding your briefs.

Measuring decay: every creative has a half-life

Fatigue is a measurement problem before it is a creative one

A winning ad is not a permanent asset. Frequency climbs, the addressable pocket of the audience saturates, and performance slides. The analytics question is how you detect that slide early enough to act, and the answer is that you watch a trend line rather than a threshold.

TikTok's own guidance points at exactly this: refresh an ad group's creatives when delivery results show a consistently declining trend, or when daily new users go low. Notice what that guidance does not say. It does not name a day.

The timescale is days, not months

Platforms now publish real numbers on how fast decay begins. TikTok's Smart Creative pauses initial videos that show signs of fatigue within the first 3 to 5 days. That is the platform's own operating assumption about a creative's early life, and it is a useful anchor: if your creative review cadence is monthly, you are reading a curve at roughly one tenth of the resolution the platform uses.

The same guidance suggests keeping 3 to 5 different creatives per ad group and 3 to 5 diversified ad groups per campaign, which is a supply requirement disguised as a best practice. A decay-aware measurement system is only actionable if there is something in the queue to replace the fading creative with.

Retire on the curve, not the calendar

The practical read is a standing weekly check on two lines per creative, from launch:

• Frequency climbing, CPA flat: healthy. The creative is still finding new people.
• Frequency climbing, CPA rising two periods running: finished. Retire and promote the next variant.
• Frequency flat, CPA rising: not fatigue. Look at the offer, the landing page, or the auction.

The middle case is the one teams misread. A creative that is finishing looks exactly like a creative having a bad week, and attributing it to the bad week is the most common way a good ad gets ridden into the ground.

How Novoads solves the cost of a readable test

A measurement system this granular only pays off if variants are cheap enough to produce in the volume it requires. Novoads generates UGC-style video ads from a script and an AI actor, or from an uploaded product image, so producing eight tagged variants of one angle is a working session rather than a production cycle. A generated video runs roughly $2 to $11 depending on the model, with a five-second Seedance or Kling clip around $2.07, which is what makes an eight-creatives-per-winner Winner Rate affordable instead of theoretical.

You can try it for $1 for 3 days. Cancel any time.

The ad is the variable, so make it the unit

Every other lever in a modern ad account has been automated away from you. Bidding is machine-tuned, targeting is broad by design, placements are chosen for you. The creative is the last input you fully control, which makes it the only place a real advantage can still be built, and that is exactly the layer your reporting is worst at describing.

Creative analytics is not a nicer dashboard. It is the decision to make the ad itself the unit of analysis, to read each metric only when it has earned the right to speak, and to judge your system by how often it produces a winner rather than by how good your best ad happened to be.