Creative Analysis: How to Read Ad Performance by Creative
Creative analysis reads ad results at the level of the creative: tag each ad, pair the tags with platform metrics, spot patterns and avoid false lessons.
02 / In this article
Creative analysis is the practice of reading ad performance at the level of the creative: which angle, format, hook and proof each ad used, and how that relates to its results. You tag every ad, pair the tags with the metrics the platform documents, look for patterns across several ads, and write down what you learned and how sure you are. This guide shows the tagging table, the metrics to read, the traps that produce false lessons and a weekly routine.
Key Takeaways
- Analyse the creative, not only the campaign. Campaign reports tell you what a budget did. Tags tell you why one ad may have behaved differently from another.
- Tag before you read. Angle, format, hook, length and proof type, assigned from the ad itself, turn a list of ads into data you can group.
- Read metrics as a chain. Impressions, then attention, then clicks, then conversions. Each link answers a different question, and none of them proves the next one.
- Distrust one ad, one week, one number. Small samples, delivery effects and attribution gaps create lessons that disappear on the next test.
- Write the lesson down with its confidence. A learning library with a status (idea, signal, confirmed) is worth more than a dashboard nobody reopens.
What is creative analysis?
Creative analysis is the structured reading of ad results by creative attribute, so that each result can be connected to a choice someone made: the message, the format, the opening seconds or the kind of proof shown. It is the "read" step in the loop of brief, produce, test, read and decide.
It differs from regular campaign reporting in the unit of analysis. A campaign report asks whether spend produced results. Creative analysis asks which creative choices were present when results were better or worse. If you are new to the role that owns this work, our guide to what a creative strategist does explains where it sits. For the structure that decides what you tag in the first place, see the creative strategy framework.
It also differs from testing. A test answers one question designed in advance. Analysis reads everything you run, including ads outside any test, and decides what to test next. Our guide to creative testing covers the design side.
Why do platform reports not answer the creative question by default?
Platform reports group results by campaign, ad set and ad, not by the ideas inside the ads. An ad name such as "Spring_V3_Final" says nothing about the angle, the opening line or whether a person appears on screen, so you cannot group ads that share a trait.
So you add the missing structure yourself, with tags. And you accept that the platform controls delivery: Google's guidance on asset reporting warns that ratio metrics for single assets, such as CTR, are influenced by which assets appear together and should be treated as directional only. A reading that ignores delivery will mislabel good ads as bad and the reverse.
How do you tag creatives so they can be compared?
You tag each ad with a small, fixed set of attributes, assigned by a person who watches or reads the ad, and you record them in one table that you keep for every ad you launch. Fixed means the same words every time: five angles do not become fifteen because a new colleague joined.
Keep the list short, because each extra attribute multiplies the ads you need before a group has enough members to say anything.
| Attribute | Question it answers | Example values (yours will differ) |
|---|---|---|
| Angle | What is the main reason to care? | Price, speed, convenience, social proof, problem-solution, status |
| Format | What kind of asset is it? | Static, short video, carousel, user-style video, motion graphic |
| Hook | What happens in the first seconds or the first line? | Question, bold claim, demonstration, pain point, pattern interrupt |
| Length | How long is the video, or how many words is the copy? | Under 15 seconds, 15 to 30 seconds, over 30 seconds |
| Proof type | What makes the claim believable? | Demonstration, customer quote, numbers from your own data, expert, none |
| Opening visual | What is on screen first? | Product, person, text only, screen recording, result |
| Offer and call to action | What is the viewer asked to do? | Buy, book, sign up, learn more |
| Placement context | Where was it designed to run? | Feed, stories and reels, search, video feed |
Three rules keep the table useful.
- Tag from the ad, not from memory. Tag at launch, from the file, with the ad name stored in the table. Memory rewrites history after the results arrive.
- One value per attribute. If an ad mixes two angles, pick the dominant one and note it. A column that allows lists becomes impossible to group.
- Put the tags in the ad name. A naming convention such as angle, format, hook and version separated by underscores lets you filter inside the ad platform and in exports. Keep the full table elsewhere.
Need a steady supply of tagged variants?
We make static ads, short video and variants for testing, with the angle and hook of each one written down, so your team can compare them in its ad accounts.
See AI ad creativesWhich metrics should you read, and what does each one tell you?
Read the metrics in the order a viewer moves through an ad, and use only the definitions the platform documents. Each metric answers one question, and each has a blind spot.
Click-through rate. Google defines CTR as clicks divided by impressions. Meta separates two versions: CTR (all), which counts all clicks, and CTR (link click-through rate), which divides link clicks by impressions. Use the link version when you care about traffic to a destination, and never compare one version with the other. A high link CTR suggests the message and visual earned attention. It does not tell you whether the people who clicked were the right people.
Early video attention. Meta documents "3-second video plays" as the number of times a video played for at least three seconds, or for 97% of its length if it is shorter than three seconds, with replays within a single impression excluded. TikTok documents "2-second video views" and "6-second video views" as the number of times a video starts playing for at least two or six seconds in an impression session, with replays excluded. These counts need a denominator before they can be compared, usually impressions.
Sustained attention. Meta defines a ThruPlay as a play to completion or for at least 15 seconds, whichever comes first. That makes it a measure of how many plays reached a meaningful point, not how many people watched in full.
Cost per result. Meta reports the average cost of each result for the optimization goal you chose. The goal, the audience, the competition and the creative all shape it, so use it to ask a question about a creative rather than to settle one.
Conversions and revenue. These come from your tracking, not from the ad platform alone. A platform-reported conversion depends on its attribution settings, and your analytics tool may count differently. We come back to this below.
Frequency. Meta describes frequency as the average number of times each person saw an ad, calculated as impressions divided by reach. It is a context metric. A rise in frequency alongside falling results is one of the signals discussed in our guide to creative fatigue.
How do hook rate and hold rate fit in?
"Hook rate" and "hold rate" are industry shorthand, not official platform metrics. A common way to define them from documented numbers is 3-second video plays divided by impressions for hook rate, and ThruPlays divided by 3-second video plays for hold rate. Define them in your own report in exactly those terms, because other teams use other formulas. Our dedicated article on hook rate walks through the calculation and its limits.
How do you turn tags and metrics into a readable pattern?
You group ads by one tag at a time, compare the metric you chose in advance, and look for a difference that repeats across several ads. A pattern is a difference that survives a change of audience, week and version. A single winning ad is an anecdote.
Work in this order.
- Fix the question before the data. "Do demonstration hooks hold attention better than question hooks in our short videos?" is a question. "What worked?" invites you to find a story in noise.
- Pick one metric for that question. For a hook question, use early video attention per impression. For an angle question, use link CTR and then the cost per result. Matching the metric to the link in the chain stops you crediting an angle for a landing page effect.
- Group by one tag, holding the others as steady as you can. If every demonstration hook was also a short video and every question hook a static, you have compared formats, not hooks.
- Count the members of each group. Three ads per group with wide gaps between them is a weak signal. Many ads per group with a consistent gap is a stronger one.
- Check the spread, not only the average. One outlier can carry a group.
- Label the output. Mark the finding as an idea, a signal or confirmed, and name what would change your mind.
The figure is a reading aid, not a statistical test. A large gap on a small sample is a reason to retest. A small gap on a large sample is usually a tie, and a tie is a result too: the attribute you changed did not matter much.
What does a generic reading look like?
The table below is fictional and illustrative. The ads, tags and relative results are invented to show the method, not to suggest what works. The scenario is a hypothetical local service business running short videos on one platform.
| Ad | Angle | Hook | Length | Early attention vs. account average | Link CTR vs. account average | Cost per result vs. ad set |
|---|---|---|---|---|---|---|
| A1 | Speed | Demonstration | Under 15 s | Higher | Similar | Lower |
| A2 | Speed | Question | Under 15 s | Similar | Similar | Similar |
| A3 | Convenience | Demonstration | 15 to 30 s | Higher | Lower | Higher |
| A4 | Convenience | Pain point | 15 to 30 s | Lower | Lower | Higher |
| A5 | Social proof | Demonstration | Under 15 s | Higher | Higher | Lower |
| A6 | Social proof | Question | 15 to 30 s | Similar | Higher | Similar |
A careless reading says "demonstration hooks win". A careful reading says more.
- Early attention is higher for all three demonstration hooks (A1, A3, A5), which is a signal about the opening, across three different angles.
- A3 has strong early attention but weak clicks and a higher cost per result. The hook held people and the rest of the ad did not convert interest into action. The lesson is about the middle of the video or the offer, not the hook.
- The two best cost-per-result readings (A1 and A5) are both short and both demonstrations, so length and hook are mixed together. You cannot separate them from this table.
- With six ads and one account, nothing here is confirmed. The output is a test brief: run a demonstration hook at two lengths, with the same angle.
That last step is the real product of analysis. A good reading ends in a decision about what to make next.
How do you avoid false lessons?
You avoid false lessons by knowing four ways creative data misleads: small samples, delivery effects, attribution gaps and confounded variables.
Small samples. A rate computed from few events moves a lot with each additional event. Statistics gives a method for this: the NIST/SEMATECH e-Handbook sets the minimum sample size for a proportion test from the significance level, the risk of missing a real difference and, above all, the smallest change you care to detect. The smaller the difference, the larger the sample required. You do not need the formula for every ad. Ask how many events sit behind a rate before reading anything into it, and decide the stopping point before you look.
Delivery effects. Platforms do not give every ad an equal, independent chance. Meta's documentation describes a learning phase after a new ad set or a significant edit, during which delivery is still being optimized and performance is less stable, and it lists a creative change among the significant edits. Google says an asset can receive few impressions because other assets are predicted to do better. In both cases, an ad that is shown less may simply have been chosen less. For a clean comparison of two ads, use a structure built for it, such as the platform's A/B test feature, rather than reading two ads that share an ad set. Our creative testing guide explains how.
Attribution gaps. The ad platform reports the conversions it attributes to ads under its own settings. Your analytics tool reports sessions and events under its own rules. Neither is the whole truth, and the two rarely match. Choose the system of record for the decision, write down its attribution settings in the report, and look at the creative question in a second source when the stakes are high. If your analytics tracking is incomplete, fix that before reading creative results. Our analytics and tracking service covers setup, and our guide to the Google Analytics MCP shows one way to query your own analytics data with an AI assistant, read-only and under human review.
Confounded variables. A new ad often differs from the old one in angle, format, hook, audience, week and offer at once. Then you learned nothing about any of them. When two attributes always move together in your table, you cannot separate them with the data you have.
What does each platform give you at the creative level?
Each platform reports at the ad level and states limits. Meta Ads Manager lets you read every ad with the metrics above, and our Meta Ads service covers day-to-day management. Google's asset reporting advises evaluating performance at the asset group or campaign level, and treats asset-level ratio metrics as directional. TikTok documents video views at two and six seconds as counts, which you convert into rates with impressions. Because its thresholds differ from Meta's three seconds, never compare attention numbers across platforms without naming the definition each uses.
Our overview of paid social covers how the channels fit together, and the video ads page covers the production side of the formats you tag.
Want analysis to feed your next creative round?
We build variants from what your tagged data suggests, so every round starts from a written hypothesis instead of a guess.
Talk to us about your creativeHow do you build a weekly creative report?
A weekly creative report is a one-page document with the same structure every week: what ran, what the data suggests, what you will make next, and what you are unsure about. Consistency matters more than depth, because the value comes from comparing weeks.
Use this structure.
- Scope. The date range, the platform, the attribution setting and the objective. Put them at the top so nobody misreads a number later.
- What ran. A short table of new and active ads with their tags and spend ranges. Keep the full ad names.
- Headline movements. The two or three changes that matter, each with the metric, the comparison and the sample behind it.
- Pattern check. One question from the pattern step, with its answer and its label (idea, signal or confirmed).
- Warnings. Delivery changes, tracking problems, edits to ad sets, and any ad whose rate rests on few events.
- Next round. The two to four variants to make, each with its hypothesis, tags and the single metric that will judge it.
- Decisions needed. What the team or client must decide this week, such as pausing, scaling or briefing new work.
How do you build a learning library?
A learning library is a searchable list of findings, each tied to the ads and tags that produced it and to a confidence label. It answers the question every new teammate and every new brief starts with: what do we already know?
Keep each entry short and structured.
- Statement. One sentence, specific: "Demonstration hooks held early attention better than question hooks in short video for this account."
- Evidence. The ad names, the date range, the metric and the sample.
- Status. Idea (a hypothesis), signal (seen more than once, not yet tested directly) or confirmed (tested with a controlled comparison and repeated).
- Scope. Where it holds: platform, audience, format, season, offer. Findings do not travel for free between accounts.
- Next action. The test that would confirm or overturn it.
- Review date. Findings age, and a lesson from last year may be false now.
Re-read the library before each brief. It stops two failures: repeating a test you already ran, and treating a one-off result as a rule.
How does creative analysis connect to making better ads?
Creative analysis connects to production through hypotheses. Each finding becomes a brief line, such as "keep the demonstration opening, change the offer in the middle", and each brief produces variants that test it. Without that link, analysis becomes a report that describes the past.
This is where fast variant production matters. Versai's AI ad creatives service covers static ads, short video and variants, with AI-assisted production under human direction. Analysis gives those variants their reasons: which angle to alter, which hook to swap, which proof to add. The service is designed so a team that reads its data can ask for the next round quickly and compare it in its own ad accounts. For creator-style formats, see our UGC ads page, and for the planning layer above all of this, our growth strategy work. If you test landing pages alongside ads, A/B testing is the matching service.
Frequently Asked Questions
What is creative analysis in advertising?
Creative analysis is the structured reading of ad results by creative attribute, such as angle, format, hook and proof type. It connects performance to the choices made in the ad, so that the next round of creative starts from evidence instead of taste.
What is the difference between creative analysis and creative testing?
Analysis proposes hypotheses from everything you run. Testing checks one of them under controlled conditions. See our guide to creative testing.
Which metrics should I use to compare creatives?
Use the platform's documented metrics in the order a viewer moves: early video attention, link click-through rate, cost per result and then conversions from your tracking. Choose one metric per question, and write down its definition. For a video opening, hook rate is a common shorthand built from documented counts.
How many ads do I need before I trust a pattern?
There is no universal number. The sample size you need depends on how small a difference you want to detect and how much risk of a wrong call you accept, as the NIST/SEMATECH e-Handbook explains for proportions. As a working rule, trust a pattern more when it repeats across several ads, weeks and angles, and label thin evidence as an idea.
Can I trust the conversions my ad platform reports?
Trust them as the platform's view under its attribution settings, not as the whole truth. Compare with your analytics tool, record the attribution settings in your report and keep one system of record for decisions.
How often should I review creative performance?
Weekly works for the report. The reading itself should wait for enough data per ad and for the learning state to settle. Decide your minimum before you look, and update the learning library monthly.
How does fatigue show up in creative analysis?
As results that weaken for the same ad while frequency rises, easiest to see by comparing an ad with its own earlier weeks. Our article on creative fatigue covers the signals.
Next steps
Start small. Write your tagging table with five attributes, tag every ad that is live today, and write one question to answer next week. Then add the weekly report and a first page of the learning library. If you want a team to build the variants that your analysis calls for, see our AI ad creatives service, and for the paid channels around them, paid media. When you are ready to talk, get in touch with the channels you run and the creative you already have.
Sources
- Google Ads Help, "Clickthrough rate (CTR): Definition": https://support.google.com/google-ads/answer/2615875
- Google Ads Help, "About the ad level asset report" and asset reporting guidance: https://support.google.com/google-ads/answer/9564897
- Meta Business Help Center, "CTR (link click-through rate)": https://en-gb.facebook.com/business/help/877711998984611
- Meta Business Help Center, "CTR (all)": https://en-gb.facebook.com/business/help/928745330472862
- Meta Business Help Center, "3-second video plays": https://en-gb.facebook.com/business/help/743427195703387
- Meta Business Help Center, "ThruPlays": https://en-gb.facebook.com/business/help/471190536725647
- Meta Business Help Center, "Frequency": https://en-gb.facebook.com/business/help/1546570362238584
- Meta Business Help Center, "About the Learning Phase": https://www.facebook.com/business/help/112167992830700/
- Meta Business Help Center, "About learning limited" (significant edits): https://en-gb.facebook.com/business/help/269269737396981
- TikTok for Business Help Center, "video play" (2-second and 6-second video views): https://ads.tiktok.com/help/article/video-play
- NIST/SEMATECH e-Handbook of Statistical Methods, "Sample sizes required" for a proportion test: https://itl.nist.gov/div898/handbook/prc/section2/prc242.htm
Newsletter
One short email when we publish
Notes on AI, image, video and brand work, from the people who do it. No spam.