“This ad has the best ROAS in the account. Make more like it.”

So you pull the platform report. Sure enough, the product-and-promo ads get the clicks. They get the sales.

But a platform can only tell you how an ad was delivered and what it could track afterward. It can’t tell you whether that ad created the order or just happened to be standing next to it.

Creative performance analytics connects each ad creative to its delivery, what people did on your site, and the orders that followed. What you get out of it are creative performance insights: the patterns worth testing next. It shows correlation, though. So when the question is whether a creative caused sales, a holdout or geo test decides.

AI makes this a lot more practical. It can classify a big creative library. It can flag a pattern that deserves a closer look. And it can make a dashboard easier for a media buyer to dig into.

What it can’t do is turn a click path, an engagement metric, or a visual pattern into proof that an ad caused an order.

What is creative performance analytics?

It’s tying each ad (an image, a video, a headline) to everything that happened after it ran. Delivery. Site behavior. Orders.

And the point is the next decision. Which hook should we make more of? Which angle is worth a test? Which ad is quietly getting credit for sales that were going to happen anyway?

A creative dashboard earns its place when it gives you a better next question, not a ranking of yesterday’s ads.

So where does that record start?

How do you build a creative record that reaches the order?

You’d think Google Ads would be one place to analyze creative. It isn’t.

Search, Performance Max, YouTube, and Demand Gen each give creative a different role. And each keeps a different record. Search-query and ad-copy analysis can be useful, but that’s a different job from creative analysis. For video-led work, I usually start with the YouTube assets, including video that runs inside other Google campaign types.

So start with a durable ID for each creative asset. Then keep a record of every ad it appeared in.

From there, a data environment the brand owns can connect creative and ad IDs to platform delivery, session behavior, and first-party orders. If your setup supports it, you can also look at the same record through a documented MTA (multi-touch attribution) view.

Now you can ask better questions:

Those beat a platform ranking every time. They connect the ad to the rest of the business.

Do you need a perfect naming convention?

No. It’s probably going to break anyway. Here’s how:

A sturdier start is a simple serial ID for every image and video.

Then AI can build a richer classification record tied to that ID. It can transcribe the video and sample frames at a low rate. Then it can rate the asset against context you control: the product, the use cases, the customer, and the comment themes you already know about.

Your team should be able to review that classification and fix it. Treat it as a working label. It’s not an objective verdict.

Which categories are useful?

Sort by format first, because the useful labels depend on the creative.

For images, you might separate a lifestyle scene, a white-background product shot, a headline-led ad, a comparison, or a feature explainer.

For video, the useful labels might be the hook (problem/solution, narrative, offer, explainer, or how-to), the product or offer shown, and the angle.

The angle is the thesis behind the ad. It could be a new technology. A news hook. Or a specific problem-and-person combo the viewer is meant to recognize. Sometimes the hook states it. Sometimes it doesn’t.

Persona labels help when the persona is explicit in the ad. Forcing a persona onto every product or offer video just adds noise.

What should the weekly view actually do?

Think of it like a scouting report. Its job is to tell you who to watch next.

A useful view can refresh nightly. It lets you sort creative by the measures that matter to the business, including directional trends.

And it should show the creative itself. A thumbnail gives people outside the data team enough context to jump in. Playable video is even better when it’s practical.

The recurring review can stay focused. At least weekly, an automated check can flag:

You don’t need to fire every small wiggle into Slack. Watch the details closely enough that your media buyer gets a short list of questions worth their time.

If you run Meta, here’s how to build the Meta version, table by table.

Which engagement metric should you watch for upper-funnel video?

Keep business outcomes at the center.

Then back-test whether a proxy, like seconds viewed per impression, completion rate, or click-through rate, actually moves with those outcomes in your data.

My view: I prefer time spent with the message, measured as seconds viewed per impression, over both CPV (cost per view) and CPC (cost per click).

Why? Seconds viewed per impression normalizes the engagement action. CPV is heavily a function of CPMs and bidding, as well as engagement. And CPC has the same problem. It mixes CPMs and the auction into what should be a read on whether people engaged.

So when a cost metric moves, you can’t tell how much of that is people watching and how much is the auction.

Still, no proxy works everywhere. Check yours against your own business data before you trust it.

How do you analyze creative performance, and where does AI fit?

Creative performance analysis connects each ad to its delivery, what people did on your site, and the orders that followed. Then you read the patterns against a metric you’ve back-tested on your own sales. AI helps by tagging creative at scale and flagging patterns to check. It doesn’t decide which ad caused sales.

What does AI actually do here?

It builds and maintains the classification record tied to each creative ID: the format, the hook, the angle, the offer. A person reviews and corrects it, because it can be wrong. Then it looks across those tags and your metrics for patterns. Do problem-solution hooks hold attention longer?

Which video metrics do I use?

These are my own definitions. Use views at 3 seconds and 15 seconds where your platform reports them. Not every platform does.

MetricMy formulaWhat it suggests
Thumbstop rate3-second views ÷ impressionswhether the opening stops the scroll
Hook rate15-second views ÷ 3-second viewswhether people keep watching
CTA strengthlink clicks ÷ 3-second viewswhether viewers act

My first pick is still time spent, measured as seconds viewed per impression (see above). None of these guarantees success. Check whether each one moves with your own sales before you trust it. And in-app website previews may count as link clicks. So once GA4 is joined, use sessions instead of link clicks for CTA strength. Here’s how to build a Meta creative dashboard with them, on a first-party order record.

What can’t AI do?

It can’t prove an ad caused sales. It can’t replace a holdout. And it shouldn’t pause an ad on its own. Every pattern is a hypothesis. Your media buyer checks the other explanations and owns the action. That’s an AI reporting layer with context and limits.

Last click undervalues the upper funnel, too. An upper-funnel video can look weak on last-click ROAS and still be adding sales. Before anyone pauses it, a credible test result should be in the room. So test the pattern before you scale it.

If you want to know which creative patterns are worth a test, bring the channels you run on to a first conversation. Or start with the MMM or incrementality quiz.

How do you evaluate a creative analytics tool or build?

Run it through these five checks. They work the same whether you’re looking at a creative analytics platform or thinking about building your own.

  1. Does it connect each creative to first-party orders (not just clicks/views)? Platform conversions are the platform’s story. Your order record is the business’s.
  2. Can you compare creatives across channels in one place? If YouTube lives in one report and Demand Gen lives in another, you’re comparing exports. You aren’t comparing ideas.
  3. Does it separate correlation from what needs a test? A good view tells you what it saw. It also tells you which ideas it can’t settle without a test.
  4. Does the brand own the underlying data? If you can’t get to the raw record, you can’t ask it a question the dashboard didn’t plan for.
  5. Does it account for delivery and exposure? An ad that runs more will rack up more of everything. So you want to know heavier-spend ads don’t look better just for running more. Think of baseball. The player with the most at-bats isn’t automatically your best hitter.

When is a custom build worth it?

The catch comes first: if you build it, you own it.

A standard creative-reporting product can be a fine choice. That’s true if its workflow already answers your question and you’d rather have a managed service.

A custom setup that the brand owns starts to make sense when you need particular metrics, joins across your own sources, branded reporting, unusual investigations, or specific alert rules.

The tradeoff is ownership. A custom system keeps you close to the underlying record and handles questions a preset dashboard won’t. But you (or your team) also own maintenance, access, uptime, and improvements. Good documentation makes a future handoff easier. It doesn’t make the responsibility go away.

So decide based on the sources, joins, and reporting you need, and on your capacity to run the thing. No one setup fits every team.

Can a creative dashboard prove an ad caused a sale?

No. A creative can do all of these without creating any incremental demand (demand that wouldn’t have happened without the ad):

The record can have gaps, too. Delivery shifts. Identity breaks across sessions. Natural demand, promotions, landing-page changes, and other media all shape the outcome.

That doesn’t make the dashboard useless. It makes it a hypothesis tool.

A good recurring view can point you to a creative pattern worth repeating. Or a landing-page mismatch worth looking into. Or a tracking gap worth fixing. Or a question that needs a different kind of evidence.

And your media buyer still owns the action. An automation shouldn’t pause ads, change budgets, swap offers, or make targeting decisions. It should show the evidence and say what context is missing. Then it should help the team think through other explanations first.

How do you test a creative idea?

What happened when Javvy Coffee tested its top- and bottom-of-funnel creative?

First, a caution. This isn’t my client. It’s a published case study from Haus, a testing company, about Javvy Coffee, a coffee concentrate brand. I tell a version of it in the video above because it shows the creative problem so clearly.

Javvy’s Meta account had two main campaigns. One ran top-of-funnel creative: mostly video, introducing the brand through problem-solution stories. The other ran bottom-of-funnel creative: mostly still images and product shots, often with testimonials or promo banners.

The budget was split evenly between them. So the question was simple. Is each campaign actually adding sales?

So they tested each one. First the bottom-of-funnel campaign, in a two-cell experiment with a holdout (a group held back from the ads). It ran for two weeks plus a one-week window afterward to catch delayed sales. Then the top-of-funnel campaign, the same way.

They didn’t just count Meta conversions. The main measure was new orders on Shopify. Amazon new orders were the second.

Guess what. Haus reports that the top-of-funnel campaign was over 13X more incremental than the bottom-of-funnel campaign.

And click-based MTA had it backwards. Per Haus, the top-of-funnel campaign’s cost per incremental acquisition came in lower than MTA reported. The bottom-of-funnel campaign’s came in much higher.

Javvy then moved the majority of its Meta budget to the top-of-funnel campaign. Haus reports that the brand’s overall CAC (customer acquisition cost) and MER (marketing efficiency ratio) improved right away.

My take: that doesn’t make promo creative useless. Demand capture closes the momentum you already have. Demand creation builds the momentum, so customers can be captured later. They have different jobs. And click attribution tends to give all the credit to the one standing closest to the sale.

It’s also one brand, one platform, and a test run by the vendor that published it. Treat it as a good example of test design. Don’t plan your budget around the number.

Source: Haus, “Top-of-Funnel Creative Drives 13X the Lift of Lower-Funnel Creative for Javvy Coffee” (published Nov 4, 2024).

The dashboard can tell you which ads got the click. Only a test can tell you which ads grew the business.

So which test fits?

If you can split people into groups, a holdout by audience can work. Some people see the creative. A comparable group doesn’t.

But a lot of creative runs where clicks can’t follow it, like upper-funnel video. People watch, don’t click, and buy later through search or another ad. No click, no tracking. That’s usually where a geo test comes in. You change the media in one set of regions and compare it with a set that stays the same.

Before you commit budget, check whether a geo test is feasible and choose the comparison for a geo test. My incrementality testing framework walks through the rest of the design. If you’d like help running one, see incrementality testing.

Creative Measurement Brief

Write one page before you build a dashboard or sit through a product demo. Fill in these five lines:

Can’t fill in that last line? Then you may be collecting creative data before you have a decision for it to support. If the record and the review process aren’t ready yet, start with the marketing data foundation or the reporting-system design.

FAQs

What is creative performance analytics?

It connects each ad creative to its delivery, on-site behavior, and orders, so you can see which creative patterns are worth testing next. It shows correlation, not cause. When you need to know whether a creative caused sales, a holdout or geo test decides.

What is creative performance analysis?

It’s reading your creative performance analytics for patterns. You tag each ad, connect it to delivery, site behavior and orders, then check which patterns move with sales. AI can speed up the tagging and the pattern-finding. Each pattern is still a hypothesis until a holdout or geo test backs it up.

How do you know if a creative caused sales?

Test it. Run the creative against a holdout group that doesn’t see it. Or use a geo test when the creative runs on channels clicks can’t track, like upper-funnel video. Then compare the outcomes. Click attribution alone can make demand-capture ads look better than they are, and upper-funnel ads look worse.

Can AI tell us which ad to scale?

It can surface patterns, comparisons, and counterarguments. A person should decide whether the evidence is strong enough to scale, pause, revise, or test an asset.

Do we need a perfect creative taxonomy first?

No. Start with durable asset IDs and a small classification approach your team can review. Improve it as you learn which distinctions actually help you make decisions.

Can creative-level MTA prove that a video caused an order?

No. It can make an observed path and the crediting rule easier to inspect. Proving cause takes a credible test designed for the decision.

The big picture

Platform reports weren’t built to answer creative questions. But creative is where a lot of the real growth decisions hide.

So give every ad an ID. Connect it to your orders.

Let AI do the sorting, and let a person do the deciding. Then test the idea that matters!

Next, whether a geo test is feasible to run three checks on your own sales history before launch, with free code before launch, the incrementality testing framework to plan a decision-useful DTC test, five decisions before building a marketing data foundation to keep a warehouse project connected to a real question, or how to build a Meta creative dashboard to join each Meta ad to its post, GA4 sessions and first-party orders by ad ID, table by table.