TL;DR
AI creative analysis tags each ad against a fixed creative taxonomy, scores it across attention, retention, interest, conversion, durability and economics, then reads the pattern of those scores to explain what won and why. Score against your own account baseline rather than published benchmarks. The output should be your next creative brief, not a ranking.
AI creative analysis reads your ad creative against its performance to find what wins, what repeats, and what breaks. It works at the creative level rather than the account level: instead of asking which campaign to cut, it asks which hook, angle, format and offer earned the result, and whether that pattern is reproducible. Vendors often sell the same capability as creative intelligence.
This article is for advertisers running enough Meta creative that manual review has stopped scaling, roughly 20 live creatives or more. It is tool-agnostic.
Two boundaries keep the job clear. This is not creative testing, which is designing and running a valid experiment. Testing generates the evidence; analysis explains the pattern. And it is not account-level analysis, which looks at campaigns, budgets and audiences. To identify winning ads with AI, the useful unit is the creative, because the creative is what you can brief more of.
Meta now exposes more of this natively. A Creative breakdown in Ads Reporting, rolling out from July 11, 2025, combines with other metrics to show delivery status, reach, impressions, cost per result and amount spent for individual creative elements in Flexible format campaigns. It is table view only, since bar and trend charts are not supported while it is applied, and its results exclude dynamic creative ads, which report element-level data separately (PPC Land). On August 20, 2026 Meta announced that Meta AI can connect to ad accounts to analyze performance and recommend optimizations, saying it can "identify patterns in successful content and help explain why certain creative may have stopped resonating" (Search Engine Land). That analysis is read-only: the changes are still made in Ads Manager. Neither gives you a consistent taxonomy across your whole library or a profitability-aware rubric. Those are the parts you own.
Drowning in Meta Ads?
Put your campaign on autopilot with Nova.
Read moreThe Creative Signal Framework
Five steps, built on 7 structural dimensions and 6 performance layers
Tag structure
Label every creative against a fixed taxonomy of seven structural dimensions, before you look at any performance data.
Score outcomes
Score each creative across six performance layers, measured against your own account's trailing baseline for the same placement.
Read the shape
Diagnose from the pattern across layers rather than the total score, because each shape points at a different fix.
Extract the repeatable pattern
Group top scorers by taxonomy dimension and find where winners concentrate, then check the pattern has enough delivery behind it to trust.
Brief the next creative
Turn the pattern into a production brief, run it, and feed the results back into step one.
Step 1: Tag structure
A taxonomy turns a folder of ads into a dataset. Without one, "the UGC one did well" is an anecdote. With one, it is a row you can group by.
The seven structural dimensions
| Dimension | What it captures | Example values |
|---|---|---|
| Hook | The first attention-capturing element: the opening 1 to 3 seconds in video, or the dominant visual or headline in a static | Question, bold claim, statistic, social proof, problem, demo, pattern interrupt |
| Format | Media structure and placement fit | Static, carousel, UGC video, founder video, product demo, 9:16 Reel, AI-generated asset |
| Angle | The underlying reason to buy | Problem-solution, social proof, comparison, mechanism, aspiration, education, skeptic |
| Offer | The commercial frame | No offer, percentage off, bundle, free trial, urgency, guarantee |
| Visual style | Production aesthetic | Lo-fi UGC, polished brand, screenshot, lifestyle, product-on-white, before and after |
| Duration or density | Video runtime, or information density for statics | 6s, 15s, 30s, 45 to 60s; sparse or dense for statics |
| CTA | The explicit ask and its framing | Shop Now, Learn More, Sign Up, Get Offer, and whether the ad carries one ask or several |
Some practitioners add Proof (review, rating, ingredient shot, before and after) and Concept (the central idea, such as "unboxing") as separate fields. Others fold proof into Angle. Either convention works provided you pick one and hold it, because the value comes entirely from applying the same labels across every asset.
Tag before you look
Tag the creative before you look at performance, or the labels drift toward the result you already know.
This is where AI earns its place first. Tagging 200 creatives by hand is a day of work nobody does twice. A multimodal model reading each asset and returning structured tags can make that fast and repeatable, though tag definitions should be fixed in advance and spot-checked; reviewing 5 to 10% of tags by hand is usually enough to catch drift. If you are building this yourself, the assets and their performance can be pulled together through the Marketing API's Insights endpoint.
Step 2: Score outcomes against your own baseline
Published creative benchmarks disagree with each other. Hook-rate and hold-rate figures vary widely between practitioner sources, partly because they do not agree on the denominator, and reported fatigue thresholds differ too. Meta does not publish universal benchmarks for these creative-level rates, and placements behave differently from each other.
Your account is the benchmark
Use your own placement-matched account baseline before relying on generic creative benchmarks.
Score each creative against your account's trailing baseline for the same placement. A 60 to 90 day window is a reasonable starting point: long enough to be representative, short enough to reflect the current account. Fast-moving accounts may need shorter. Bands: 2 points at or above the 75th percentile of that baseline, 1 point between the 25th and 74th, 0 below the 25th.
The six performance layers
| Layer | Question | Metric | Scoring |
|---|---|---|---|
| Attention | Did it stop the scroll? | Thumbstop rate | Percentile bands |
| Retention | Did it hold once stopped? | Hold rate | Percentile bands |
| Interest | Did it earn the click? | Link CTR | Percentile bands |
| Conversion | Did the click convert? | Post-click CVR | Percentile bands |
| Durability | How much delivery before it decayed? | Delivery accumulated before a material CTR decline against the creative's own early baseline | Percentile bands; more delivery before decay scores higher |
| Economics | Did it clear the profitability floor? | Actual ROAS against break-even ROAS, or contribution profit per acquisition | 2 above target, 1 above break-even, 0 below break-even |
Fix your definitions once
These metrics are not standardized across the industry, so treat the following as this framework's operational definitions rather than universal ones. Neither thumbstop nor hold rate ships as a native column, so both have to be built as custom metrics in Ads Reporting from Meta's own fields.
- Thumbstop rate = 3-second video plays ÷ impressions
- Hold rate = ThruPlays ÷ 3-second video plays
- Durability decline = an illustrative 20% drop in link CTR against the creative's own stabilized early-period level. This is a framework heuristic rather than a Meta standard, and our creative fatigue guide covers alternative decay signals
One caution on the denominator: Meta counts a ThruPlay when a video of 15 seconds or shorter is watched to completion, or when at least 15 seconds of a longer video is watched. It therefore behaves as a completion event on short creative and a retention event on longer creative. Mixing hold rate defined on ThruPlays with hold rate defined on 15-second views inside one account is the most common way a scoring system quietly stops working.
Not every layer applies to every format
Statics and carousels have no video-view metrics, so attention and retention cannot be scored the same way.
Metric availability by format
| Layer | Video | Static and carousel |
|---|---|---|
| Attention | Thumbstop rate | Not available |
| Retention | Hold rate | Not available |
| Interest | Link CTR | Link CTR |
| Conversion | Post-click CVR | Post-click CVR |
| Durability | CTR decay vs own baseline | CTR decay vs own baseline |
| Economics | Against break-even | Against break-even |
Video scores out of 12, statics and carousels out of 8. Do not compare raw totals across formats. Compare within format, or compare each creative's share of its available points.
Step 3: Read the shape, not the total
The core idea
Creative diagnosis comes from the shape of the score, not the total score.
A creative scoring 2 on attention, 2 on retention and 0 on conversion is telling you something specific, and it is a different message from 0, 0, 0 at a similar total.
Reading the score shape
| Score shape | Consistent with | Next move |
|---|---|---|
| High attention, low retention | The hook works and the body does not | Keep the hook, rebuild what follows it |
| High attention and retention, low conversion | A promise the offer or landing page does not meet | Investigate offer and page alignment first, then downstream conversion friction |
| Low attention, strong lower funnel | A good asset almost nobody sees | Re-cut with a new opening, but only if the lower-funnel sample is large enough to trust |
| Strong scores, low durability | A winner with a short shelf life | Budget replacements before it decays rather than after |
A score pattern narrows the investigation. It does not prove the cause.
The economics layer is the one most scoring systems skip. Platform ROAS does not know your cost of goods, shipping or the discount you ran, so a creative that looks like a winner on reported ROAS can be losing money per order, and scaling one below your profitability floor can increase total loss if the economics do not improve. Score against break-even ROAS and the ranking frequently reorders.
Step 4: Extract the repeatable pattern
Patterns over ads
A single winning ad is a result. A pattern is an asset.
Group the top scorers by each taxonomy dimension in turn and look for concentration:
- Which hook types concentrate at the top? Five of seven best performers opening on a problem statement is a briefable instruction.
- Which angle carries across formats? An angle that wins as a static and as a UGC video is likely a real insight about the buyer rather than a production artifact.
- Which combinations outperform their parts? Problem hook plus mechanism angle is a different ad from problem hook plus social proof, and they rarely perform alike.
- What do winners share that you did not intend? Duration clustering, a recurring visual motif, a CTA phrasing that crosses unrelated concepts.
Check the volume first
Do not promote a pattern into a creative rule until the underlying creatives have enough delivery to make the comparison credible.
A pattern found across five creatives with 30 conversions between them is noise wearing the costume of an insight. Treat thin-volume patterns as hypotheses to test. Our creative testing guide covers the volume thresholds that make a comparison valid.
Step 5: Diagnose the failure modes
Symptom to cause to fix
| Symptom | Consistent with | What to change |
|---|---|---|
| Bottom-band thumbstop, later layers untested | Weak hook. The ad never earned an audience | Rebuild the opening. Keep the body if it scored well elsewhere |
| Healthy thumbstop, bottom-band hold | Body or pacing failure. The opening promised what the middle did not pay off | Tighten the first 10 seconds, move proof earlier |
| Good CTR, bottom-band post-click CVR | Angle or offer mismatch | Align the creative's promise to the landing page, or change the offer |
| Rising frequency with CTR decaying against its own early level, CPM drifting up | Creative fatigue | Rotate the creative, and see the fatigue guide below for detection signals |
Each of these narrows where to look. Confirming the cause generally still needs a test. Detection and remedies for the last row are covered in our creative fatigue guide.
A worked example
Illustrative figures
The following uses illustrative numbers to demonstrate the method. It is not client data.
A store with a 42% contribution margin runs six creatives in a prospecting campaign over three weeks. Break-even ROAS is 1 ÷ 0.42, or about 2.38. Trailing 90-day baseline for this placement: thumbstop median 31% and 75th percentile 36%, hold median 24% and 75th 28%, link CTR median 1.1% and 75th 1.35%, post-click CVR median 3.4% and 75th 4.0%.
Six creatives, tagged and measured
| Creative | Hook | Angle | Format | Thumbstop | Hold | CTR | CVR | ROAS |
|---|---|---|---|---|---|---|---|---|
| A | Problem | Mechanism | UGC video | 39% | 29% | 1.4% | 4.1% | 3.1 |
| B | Bold claim | Aspiration | Polished video | 41% | 16% | 0.8% | 3.2% | 1.9 |
| C | Problem | Social proof | UGC video | 36% | 27% | 1.3% | 3.9% | 2.9 |
| D | Statistic | Education | Founder video | 22% | 26% | 1.2% | 4.0% | 2.6 |
| E | Demo | Mechanism | Product demo | 34% | 26% | 1.5% | 1.8% | 1.4 |
| F | Social proof | Comparison | Carousel | n/a | n/a | 0.9% | 3.1% | 2.2 |
Scored against the baseline. Durability is omitted because three weeks does not give a reliable decay curve, so video scores out of 10 and the carousel out of 6.
The same six creatives, scored
| Creative | Attention | Retention | Interest | Conversion | Economics | Total | Shape |
|---|---|---|---|---|---|---|---|
| A | 2 | 2 | 2 | 2 | 2 | 10/10 | Strong throughout |
| B | 2 | 0 | 0 | 1 | 0 | 3/10 | Hook works, body fails |
| C | 2 | 1 | 1 | 1 | 1 | 6/10 | Solid throughout |
| D | 0 | 1 | 1 | 2 | 1 | 5/10 | Low attention, strong lower funnel |
| E | 1 | 1 | 2 | 0 | 0 | 4/10 | Earns clicks, does not convert |
| F | n/a | n/a | 0 | 1 | 0 | 1/6 | Below the profitability floor |
Reading the shapes:
- A and C score well across every available layer, and both are problem-hook UGC video. Two of the top performers share a hook type and a format.
- B has the strongest attention score in the set and the weakest retention. Consistent with a hook that earns the stop and a body that loses the viewer. The asset worth keeping is the opening.
- E earns top-band clicks and converts at roughly half the baseline median. Consistent with an offer or landing-page mismatch rather than a creative-quality problem.
- D underperforms on attention alone. Everything after the stop is at or above median, so it is worth re-cutting with a stronger opening rather than retiring.
- F cannot be compared to the videos on total. Scored on its available layers it is below break-even at 2.2 against 2.38, which moves it from the maybe list to the cut list.
The pattern: problem hook plus UGC video concentrates at the top, with mechanism and social proof both working underneath it. The brief: three new UGC videos opening on a problem statement, testing mechanism against social proof in the body, plus one re-cut of D with a problem-led opening.
From analysis to new creative
The output that matters
AI creative analysis should end in a production brief, not a dashboard.
The loop runs continuously: analyze, extract the pattern, brief it, produce variations, test, re-analyze. The friction is almost always the handoff, where analysis lives in one place and the brief gets written by hand in another.
Systems that hold both halves close that gap. Nova reads account performance against your unit economics rather than platform ROAS alone, and Iris, its creative specialist, generates new creative from your product imagery and brand voice using the angles that score in the analysis. Because scoring and production sit in the same system, the pattern and the brief do not have to travel between tools. On the aggregation side, our Creatives Hub consolidates each creative's performance across every ad it appears in.
Manual versus AI creative analysis
Where each one is stronger
| Task | AI | Human |
|---|---|---|
| Tagging against a taxonomy | Scalable and repeatable when the taxonomy is fixed | Slower at volume, and labels drift between sessions |
| Scoring against a rolling baseline | Straightforward once defined | Error-prone in spreadsheets |
| Spotting cross-dimension patterns | Strong at library scale | Good on small sets, misses combinations at scale |
| Flagging fatigue | Useful for continuous monitoring | Tends to catch it later |
| Judging brand fit | No view on what the brand should not say | The decision |
| Interpreting why a pattern exists | Proposes hypotheses | Supplies the causal reading and the context |
| Big creative bets | Recombines what already worked | The decision |
AI is strongest at recombining observed winning patterns. Genuinely novel brand bets still benefit from human creative direction, which is exactly the input a system trained on your past cannot supply. Tool selection for this sits in our creative analysis tools roundup, and for the broader account view, see how to audit your Meta ads account using AI.
Frequently asked questions
AI creative analysis for Meta ads: common questions
Summary
Tag every creative against a fixed taxonomy of hook, format, angle, offer, visual style, duration and CTA. Score it across attention, retention, interest, conversion, durability and economics, against your own account baseline rather than a published benchmark, and only on the layers its format actually supports. Read the shape of the score rather than the total, because high attention with low retention, good CTR with weak conversion, and decaying CTR at rising frequency each point somewhere different. Then turn the winning pattern into the next brief, which is the only output that changes what happens next.
Sources
- Meta unveils Creative breakdown for flexible formats and AI-generated image ads, PPC Land, July 11, 2025
- Meta AI can now analyze and optimize Meta Ads campaigns, Search Engine Land, August 20, 2026
- Insights API, Meta for Developers
- About ThruPlay, Meta Business Help Center
Performance Optimization
Meta Ad Image Sizes & Specs: The Complete 2026 Guide
Oversized image files slow your ads down and cost you conversions. Here's what Meta actually requires, why it matters for performance, and how to fix it in seconds
Read more
