How to Measure Creative Performance: The 4-Tier Protocol

Measuring a creative means comparing variants under identical conditions, not reading an absolute number. You stack four tiers: attention (hook rate = 3-second views ÷ impressions on Meta, 2 seconds on TikTok), retention (hold rate at 15 seconds), action (outbound CTR, CVR) and cost (CPA per creative). Each figure is only valid at the same placement, audience temperature and time window, and above a minimum volume. Below that, you are reading noise. The rest of this article gives the floors, the comparability conditions and the decision sheet that turns these four tiers into a call: kill, iterate or scale. On the TikTok side, the TikTok Ads Manager help centre defines 2-second views as the number of times your video is played for at least 2 seconds during an impression session, not counting replays.

Why does a 35% hook rate prove nothing on its own?

Because a hook rate is not a property of the video: it is the product of a video, a placement and an audience. Change any one of the three and the figure moves by ten points without a single frame having been touched.

The most useful data we have read this year comes from AdSights (2025 analysis, picked up by Sepia Lab in 2026): in cold prospecting, thumbstop sits between 18 and 28%, with a median around 22%. In warm retargeting, the same measure climbs to 30-45%, with a median of 36%. A fourteen-point gap, caused by the audience, not the creative.

Set that next to the threshold everyone repeats, “above 30%, it's a good creative”, and you see the problem. A mediocre creative served to an audience that already knows the brand clears the bar. A solid creative served cold, on an account that is scaling, falls below it. Both are judged wrongly. In the account audits we run before taking over a production, this is the most frequent mistake: creative rankings built on lines that ran neither in the same place nor in front of the same people.

The three most common contaminations

• Audience temperature. Retargeting, 1% lookalike, broad: three different thumbstop regimes. Comparing a remarketing creative with a prospecting creative is like comparing a downhill sprint with an uphill one.
• Mixed placements. Feed, Reels, Stories and Audience Network have neither the same autoplay behaviour nor the same cost of attention. A creative with 60% of its impressions going to Audience Network will show a hook rate nobody can interpret. According to Médiamétrie's L'Année Internet 2025, relayed by CB News, mobile accounts for 80% of the time the French spend on the Internet, up from 66% in 2020.
• Advantage+ and uncontrolled distribution. When the algorithm alone decides how budget is split between assets, the creative receiving 70% of the spend is not necessarily the best one: it is the one that won the first hours of exploration. That is an allocation bias, not a verdict.

Add one last detail that many people overlook: Meta publishes no official hook rate benchmark. The 25%, 30% and 40% figures in circulation are a practitioners' convention, aggregated by third-party tools (AdSights, Sepia Lab, Benly, Hawky) across their own pools of accounts. That is useful. It is not a standard.

If you want the detail of the ratios themselves and how to calculate them in the ads manager, we have already written about the three ratios to measure on Meta Ads video. This article deals with the tier above: under what conditions a number is entitled to enter a decision.

An absolute threshold with no comparability condition is an opinion with a decimal point.

What are the 4 tiers of creative measurement, and in what order should you read them?

Attention, retention, action, cost, in that order and never in any other. The order is not aesthetic: each tier filters the volume available to the next. If 100 people see the impression and 20 stay for 3 seconds, your hold rate is calculated on 20 people, not 100. Reading CTR before the hook is diagnosing a leak by looking at the water bill.

Tier 1: attention

Hook rate measures one thing: how many people did not scroll. Nothing else. A high hook rate says nothing about sales, and a low hook rate condemns everything else, because it shrinks the base on which the next three tiers are calculated. We will not go back over how to build a hook here; that is the subject of our article on the definition and cost of a video ad hook. In Kantar Media's TGI 2025 study, reported by The Media Leader, 47% of respondents say they never watch ads shown on social media.

Tier 2: retention

Hold rate = 15-second views ÷ 3-second views. Always state which formula you are using, because three definitions are in circulation: relative to 3-second views, relative to impressions, or expressed as a completion percentage. The figures are not comparable across the three. We work with the first (AdSights, 2026 guide), and we write it out in full in every report.

Tier 3: action

Outbound CTR first, CVR second. Outbound CTR (not the “all clicks” CTR, which counts caption expansions and clicks on the page name) measures whether the promise created movement. CVR measures whether that promise could be kept. When CTR is good and CVR has collapsed, the problem is almost never the video: it is the gap between what it promises and what the page shows.

Tier 4: cost

CPA per creative is the only figure you really care about, and it is also the slowest to become reliable. ROAS, for its part, is not an asset metric. It is an account metric, possibly a campaign one. Attributing it to a video in an Advantage+ environment is like crediting one player with the team's score. We cover realistic ranges in what ROAS to expect from creative ads; here, just remember that “this creative's ROAS: 2.4” in a report is a sentence to challenge.

Which thresholds should you use in 2026 for each platform, and where do they really come from?

Market thresholds are for detecting an anomaly, not validating a creative. Use them like a smoke detector: below your platform's median, there is probably a structural problem. Above it, you still know nothing; you need to compare against your own baseline.

For retention, Adlibrary (2025/2026) gives a healthy zone for cold traffic of 15 to 25% at 15 seconds, or 8 to 15% at 75% completion. Below 10%, the middle of the video has a structural problem: too long before the proof, too much context, not enough reasons to stay. Best-in-class DTC runs between 25 and 30%.

The TikTok correction almost nobody applies

Same creative, same edit, same voice: TikTok retention is structurally lower than on Meta (Adlibrary, 2026), because swiping there is almost a reflex and the cost of leaving is zero: the video loops and the user moves on to the next one by themselves, very early. More than once we have seen teams conclude that “UGC doesn't work on TikTok” by comparing a TikTok hold rate with a Meta threshold. Wrong diagnosis, wrong decision, a whole batch thrown away.

Two serious sources, two truths: the apparel CTR example

Sepia Lab, in its Video Ad Benchmarks by Industry 2026, gives for apparel in Meta video: CTR 2.25%, CVR 1.46%, CPA $36.76, ROAS 2.18. Trendtrack, on the same vertical in 2025, gives CTR 1.24%, CVR 4.11%, CPA $10.98, ROAS 2.30. A factor of two on CTR, a factor of three on CPA. According to Kantar Media's Total Video review for the first quarter of 2026, fashion puts 52% of its video advertising investment into Paid Social in France, a share well above the market average.

Neither of them is lying. They do not share the same panel of accounts, the same average basket or the same definition of a “click”. Operational conclusion: an external benchmark never replaces your internal baseline. It is there to frame an order of magnitude when you start, and that is all. As market reference points: WordStream/LocaliQ (Facebook Ads Benchmarks 2025) publishes, for traffic campaigns across all sectors, an average CTR of around 1.7% and an average CPC of around $0.70; e-commerce panels such as Triple Whale, for their part, put Meta CPM at around $15 over the last rolling year, an order of magnitude from a private panel that should be checked in the source report. These amounts are in US dollars, as published by the sources.

At what volume can a creative be judged?

Each tier has its floor, and a tier that has not reached its floor is not read at all, not read “with caution”. This is the hardest discipline to maintain internally, because a number displayed in the ads manager always gives the illusion of being information.

The floors below are the ones we apply on our client accounts. They are not thresholds published by Meta or TikTok: they are a working convention, calibrated so that the ranking of creatives stops reshuffling from one day to the next. Your own thresholds may differ if your conversion rate is very high or your basket very low.

The anti-premature-kill rule fits in one line: never cut a creative on tier 4 before you have spent 3 times the target CPA on that asset. Below that, you are deciding on chance. You can, however, cut it very early on tier 1: an 11% hook rate over 10,000 cold impressions is not going to recover on its own, and there is no reason to pay to find out.

This D+3 / D+7 / D+30 rhythm also depends on the production throughput behind it. A brand that releases 4 creatives a month cannot keep the same decision cadence as one that releases 25: we detailed the calculation in how long it takes to see results with creative ads.

“Do we have enough data?” The right answer is not a feeling. It is a floor written down before the test launches, and respected even when the number is tempting.

How do you normalise two creatives so they are genuinely comparable?

Five conditions must be identical: placement, audience temperature, time window, budget regime and campaign objective. If even one differs, the declared winner is a statistical illusion.

The 5 comparability conditions

• Placement. One placement per test, or at the very least a breakdown by placement in the reporting. Reels and Feed do not get blended into an average.
• Temperature. Cold against cold. AdSights' 14-point gap is enough to flip a ranking.
• Window. Same calendar period, same days of the week. A test launched on a Friday and another on a Tuesday do not face the same auction competition.
• Budget regime. Either all variants at equal ad set budget, or all in the same CBO campaign. Never a mix of the two.
• Objective and optimisation event. A creative optimised for “add to cart” and one optimised for “purchase” do not receive the same type of traffic.

The test protocol we apply

One ad set dedicated to the test, separate from the structure that scales. Between 3 and 6 variants at most, because beyond that the budget gets diluted and none reaches its conversion floor. Budget calculated backwards: number of variants × 3 times the target CPA = test budget. Duration fixed in advance, 7 or 14 days. No intervention during the first 72 hours, the period when the algorithm explores and produces unstable figures.

And above all: a single variable changed between variants. Different hook, identical body. Or different body, identical hook. If you change the hook, the voice and the CTA at the same time, you will know which one wins but never why, and you will not be able to replicate anything on the next batch.

The same internal-baseline logic applies to media costs: building your own reference is always better than importing someone else's market. The how-to is in building your own CPM benchmark in 14 days.

What should you do when the tiers contradict each other?

A contradiction between two tiers is the most useful diagnosis you will get: it pinpoints the flaw to the second. A creative that is “average everywhere” teaches you nothing. A creative with a 42% hook rate and a 7% hold rate tells you exactly what to redo.

The “weak hook, strong hold” profile is the one we look for most with our e-commerce clients, because it is the cheapest to fix. The body of the video has already been validated by the viewers who stayed 15 seconds. All that remains is to shoot four 3-second openings. Marginal cost close to zero, potential gain across the whole volume of impressions.

One case that does not appear in the table, because it calls for a different diagnosis: all tiers degrade gradually after two weeks with nothing having changed. There, you are no longer dealing with a creative quality problem but with audience wear-out, covered in our article on video ad fatigue and its false diagnoses.

How do you go from measurement to decision: kill, iterate or scale?

Three possible outcomes, and one rule: the decision must be written down before looking at the numbers. Otherwise you will always find a way of reading the data that saves the creative you put the most energy into.

The decision sheet

• Kill at D+3. Hook rate below 15% on cold traffic with 10,000 impressions reached. No discussion, no “let's give it a chance”.
• Iterate at D+7. Hook is fine, the next tier is failing. Send it back to production with a precise instruction taken from the diagnostic table above. One variable changes.
• Scale at D+14 or D+30. CPA below target over at least 25 to 30 conversions. Increase the budget in steps of 20 to 30%, and prepare the variations already, because a winner wears out.
• Keep in reserve. The fourth outcome, rarely formalised: a creative that is average in acquisition may well perform in retargeting. Do not throw the file away, move it to a different audience tier.

The creative scorecard: 5 to 7 metrics, not one

The scorecards that last over time weight between 5 and 7 metrics (Hawky.ai, 2026). A typical weighting given by Benly.ai in 2026: hook rate 25%, hold rate 20%, the rest split between completion rate, outbound CTR and CVR. We systematically add a volume line: a creative whose score rests on 4,000 impressions is marked “not judgeable”, whatever its score.

Beware the opposite trap. A single score reassures the management committee and hides exactly the information that is useful for production. A creative at 72/100 tells nobody what to reshoot next week. Always keep the per-tier detail alongside the score.

Winner rate, the metric we actually watch

At the level of the whole set-up, the individual performance of a creative matters less than the yield of the batch. How many creatives do you have to produce to get one that falls below the target CPA and stays there for 30 days? That is the cost per winning creative, and it is the only line that links your production budget to your profit and loss account. We detail the calculation in the real cost per winning creative in e-commerce.

What we do not know, and refuse to pretend we do: there is no universal winner rate. On our accounts, it varies enormously with the maturity of the account, the breadth of the target and the nature of the product. A brand with a strong offer and a broad audience finds winners much faster than a niche, high-basket product. Measure yours over three consecutive batches before making it a target.

Which metrics should you demand in a creative report, and which should you refuse?

Demand metrics broken down by asset, by placement and by temperature; refuse any account average presented as creative performance. A report that gives you the overall ROAS and three screenshots of videos is not a creative report, it is a presentation.

The four lines to impose in the reporting table: asset ID, placement, audience temperature, volume reached at each tier. Without these four columns, the performance figures that follow cannot be interpreted, even if they are accurate.

One figure we cite with reservations: Motion, picked up by Jetfuel Agency in 2026, claims that 73% of e-commerce video ads fail in the first three seconds. The source itself sets methodological limits on this measure. Take it as an order of magnitude on the importance of tier 1, not as a statistic to quote in committee.

A creative report without a “volume” column is a report of opinion. We stopped sending them in 2024.

Where do you start if you measure poorly today?

Take your last thirty creatives and reclassify them by placement and temperature before looking at a single number. It is half a day's work, and in most of the accounts we take over, that sort alone reverses two or three rankings that had been drawn up.

Then write down your floors. Not your performance thresholds: your volume floors. How many impressions before you talk about the hook, how many conversions before you talk about CPA. Display them in the same document as the reporting. Until those two numbers are set, knowing how to measure a creative's performance remains a theoretical exercise, because every meeting will reopen the question “do we have enough data?”

One last point, and it is the one that stings: an impeccable measurement protocol on too low a production throughput is useless. If you release three creatives a quarter, you will never have enough variants for the comparison to say anything at all. Measurement is a function of the volume tested. We put a number on how many variations to produce per angle in the 3 then 5 to 8 variants rule, and it is probably the lever that will change your numbers most this year, before any improvement to your dashboard.

Frequently asked questions

What is the exact difference between hook rate and thumbstop?

In everyday usage, the two mean the same thing: the percentage of impressions that generate a 3-second view on Meta, 2 seconds on TikTok. “Thumbstop” comes from English-language vocabulary; “taux d'accroche” is the French equivalent. Always check the duration your tool uses: some dashboards calculate thumbstop at 2 seconds even on Meta.

Can you attribute a ROAS to a specific creative?

No, not reliably in an Advantage+ or CBO environment, where the algorithm splits spend between assets according to signals you do not control. ROAS is a campaign or account metric. At asset level, work with CPA and the share of spend received, always stating the volume of conversions behind the figure.

How many impressions do you need before judging a hook rate?

We apply a floor of 8,000 to 10,000 impressions per asset before reading the hook rate at all, and 25 to 30 conversions before reading a CPA per creative. These are working conventions, not thresholds published by the platforms. Below that, the ranking of creatives reshuffles from one day to the next and the decision becomes random.

Why is my hold rate lower on TikTok than on Meta with the same video?

According to Adlibrary in 2026, TikTok retention runs structurally 5 to 10 points below Meta for an identical creative, because the app moves on to the next video around the fifth second. Comparing a TikTok hold rate with a Meta threshold produces a false diagnosis. Build a separate baseline for each platform before drawing conclusions.

Does Meta publish an official hook rate benchmark?

No. The 25%, 30% or 40% thresholds in circulation come from third-party tools and practitioner aggregations (AdSights, Sepia Lab, Benly, Hawky, Billo), each on its own pool of accounts. They are for spotting an anomaly, not validating a creative. Your internal baseline, measured on your placements and your audiences, remains the reference.

Do you need a single score to rank your creatives?

A single score makes reading easier in committee but destroys the information that is useful for production. Operational scorecards weight 5 to 7 metrics according to Hawky.ai in 2026, with for example 25% for hook rate and 20% for hold rate at Benly.ai. Always keep the per-tier detail alongside the aggregate score.