To tell which part of your ad is driving installs and sales, you have to measure the ad in pieces, not as a whole. Tag the same creative elements (hook, CTA, character, music, visual style) on every ad, group spend and results by element, then read the right metric at each funnel stage, from hook rate down to ROAS. Segwise is an AI creative analytics platform built for exactly this: it tags each element with multimodal AI and joins it to your MMP's install and revenue data, so the answer shows up without a spreadsheet.
Ask a user acquisition team what is working and you often get some version of this: no way to tell which hook style is actually driving the best hook rates. It is a problem most teams live with. You can see that an ad is working. You just cannot see why it is working, or which piece of it to put in the next batch.
This matters more than the org chart suggests. Circana and NCSolutions, in their Five Keys to Advertising Effectiveness analysis of nearly 450 CPG campaigns, found that creative drives nearly half (49%) of advertising's incremental sales, the single biggest factor, ahead of brand (21%), reach (14%) and targeting (11%), according to Circana. Plenty of articles still quote a twenty-year-old 65% figure, but even the current number makes the point: if half your result comes from the creative and you cannot say which element did the work, you are flying blind on the thing that matters most.
The good news: the method is not complicated. It is just tedious by hand, which is why most teams skip it. Before you start, you need four things: your install and revenue data from your MMP, per-ad spend from each network, your full creative library, and a consistent way to tag it. With those in place, the rest is five steps, and this guide walks through each one with a worked example and a checklist you can run this week.
TL;DR
Per-ad reports tell you which ad won, not which element won. The two are different questions, and the second one is the one you can act on.
Tag the same elements the same way across every ad, or the comparison falls apart before it starts. A shared tag taxonomy is the foundation.
Group spend and results by element, not by ad, and always control for spend. Raw install totals mostly reflect where the algorithm sent budget.
Read the metric that fits the funnel stage: hook rate for the first few seconds, CTR for the click, install rate for the store, then trial and ROAS for revenue.
Confirm your finding with one clean test that changes only the element you suspect. If the edge survives an even-budget test, it is real.
Why your ad reports can't answer this
Open any ad network dashboard and you get a row per ad: spend, impressions, installs, cost per install, ROAS. Useful, but it answers "which ad performed," not "which element performed." Those are not the same question.
Here is the catch. One hook appears in five different ads. One CTA appears in ten. A character shows up in half your library. The elements are smeared across dozens of ads, so a per-ad view can never isolate them. You can stare at a leaderboard all day and never see that your best three ads share the same opening three seconds.
Spend makes it worse. Ad networks concentrate budget on a handful of ads very quickly, so your install totals mostly describe where the algorithm sent money, not which creative choice was strongest. The ad with the most installs is often just the ad that got the most budget. Raw totals reward spend, not quality.
So the fix is not a better leaderboard. It is to stop reporting on ads and start reporting on elements.
A five-step way to find the element that's working
1. Tag the same elements on every ad
You cannot compare elements you have not labelled. Step one is to break each creative into its parts and tag them consistently: the hook style, the CTA, the main character or presenter, the visual style, the music or audio, the core message. For a mobile game you might add the gameplay shown and the first-level difficulty. For DTC you might add the product, the offer and the social proof.
The word "consistently" is doing the heavy lifting. If one person tags a video "UGC hook" and another tags a near-identical one "testimonial opener," your analysis is dead on arrival. Agree on a fixed vocabulary and apply it the same way every time. This is where a written tag taxonomy pays for itself, and it is the step teams most often rush.
2. Group spend and results by element
Once every ad carries the same tags, flip the unit of analysis. Instead of a row per ad, build a row per element value. One row for each hook style, one for each CTA, one for each character, and so on. Sum the spend, impressions, installs and revenue of every ad that carries that tag.
Now you can ask real questions. Which hook style has the lowest cost per install across everything it touched? Which CTA shows up in your top-ROAS ads and never in the bottom? Which character quietly drags down every ad it appears in? These answers are invisible in a per-ad view and obvious in a per-element view.
3. Read the right metric for each funnel stage
An element does its job at a specific point in the funnel, so judge it with the metric for that point. Grading a hook on ROAS is unfair to the hook, and grading a CTA on install rate misses where it actually works.
Hook (first few seconds): hook rate, or the share of impressions that watch past three seconds. The opening carries outsized weight. A 2021 TikTok study with the neuromarketing firm MediaScience found its ads draw attention in the first few seconds, and an ad on screen for six seconds or less still delivered 38% of the recall of ads watched for 20 seconds or more, as Social Media Today reported.
Body and offer: CTR and click-through. This is where your message, offer and CTA earn the tap.
Store and onboarding: install rate and conversion rate. A strong click that does not convert points at a mismatch between the ad's promise and the store page.
Revenue: trial starts, cost per action, ROAS and LTV. This is where the whole creative, and the quality of the user it brought in, finally shows up.
Map each element to the stage it influences, then read it there. A hook with a great hook rate but weak ROAS is not a failed hook. It is a great hook attached to a weak back half, which is a different fix.
4. Control for spend
This is the step that separates a real finding from a coincidence, and it is the one most people skip.
Because budget is never split evenly, you cannot compare raw install counts between elements. Compare rates instead: cost per install, conversion rate, ROAS, cost per action. A hook that drove 2,000 installs on $2,000 beat a hook that drove 5,000 installs on $9,000, even though the second number looks bigger.
Set a floor, too. A common rule of thumb is to wait until an element has run on at least $1,000 to $2,000 in spend, or roughly 50 or more installs, before you rank it. An element sitting on $300 and a handful of installs has not earned a verdict yet, no matter how good its rate looks, so park anything below your floor instead of ranking it. That way one lucky ad does not crown an element that never really got tested.
5. Confirm with one clean test
Grouping by element tells you where to look. A clean test tells you if you are right.
Pick the element you believe is winning and build two or three new ads that change only that element. Same everything else: same product shots, same CTA, same music, swap only the hook. Give them even budget caps so the algorithm cannot decide the race by itself in the first hour. If your element's advantage holds up under equal footing, it is real and you can scale it. If it disappears, the edge belonged to something else, usually spend or a lucky pairing, and you just saved yourself from betting the next quarter on noise.
A worked example
Here is a simplified example. The numbers are invented, but the pattern is one every UA team has seen.
A studio runs eight video ads for a mobile game. Five open with a gameplay hook, three open with a character hook. On the per-ad leaderboard, the top ad by installs is a character-hook ad with 7,800 installs on $18,000 of spend, a $2.31 cost per install. No single gameplay ad comes close on raw installs, so the obvious call is "character hooks win, make more of those."
Now group by hook style and compare rates instead of totals:
The character hook has more installs because it got more than twice the budget. On the metrics that control for spend, the gameplay hook wins clearly: a 23% lower cost per install ($1.75 against $2.27) and a D7 ROAS above breakeven while the character hook sits below it. The gameplay hook was the stronger element all along. It just never got the budget to prove it in a per-ad view.
The action falls out of the table: shift budget toward gameplay-hook ads, brief more variants that open on gameplay, and run a clean test that swaps only the hook to confirm the edge before you scale it.
A quick checklist
Run this against your own account:
Every ad is tagged for hook, CTA, character, visual style, audio and core message, using one fixed vocabulary.
Your analysis has a row per element value, not a row per ad.
You are comparing rates (CPI, CVR, ROAS, CPA), never raw install or spend totals.
Each element is judged on the metric for its funnel stage, from hook rate to ROAS.
Elements below a minimum spend threshold are parked, not ranked.
Your top finding has been, or is about to be, confirmed with an even-budget test that changes one element.
The winning elements are written into the next creative brief so the insight actually ships.
If you can tick all seven, you can name the element driving your installs and your sales, not just the ad.
Where this breaks down by hand, and what to automate
Everything above works in a spreadsheet. It also falls apart in one, for three reasons. Tagging hundreds of videos by hand is slow and drifts over time. The performance numbers live in your MMP while the live creative lives in each ad network, and nothing joins them for you. And by the time you have finished cross-referencing last month, the next batch is already live. This is the exact bind WGT's user acquisition team described before they automated it: "To identify top-performing assets, we had to pull data from our MMP and then manually cross-reference it against the live creative sitting in each individual ad network. There were no shortcuts."
This is the work Segwise, an AI creative analytics and generation platform, is built to remove. Its Creative Tagging Agent watches, listens to and plays every creative you run and tags each element with multimodal AI across video, audio, image, text and playables, so the tagging is consistent by default instead of by discipline. Every tag is mapped to metrics like ROAS, CPA, IPM and CTR, which is step two and step three done automatically. And because Segwise joins creative data across 15+ ad networks and your MMP, whether AppsFlyer, Adjust, Branch, Singular or Kochava, the install and revenue side is already sitting next to the creative, no cross-referencing required.
That join is what changed things for WGT. With MMP and network data in one view, WGT reports its early concept-to-test cycle runs about 5x faster than it did when the data lived in separate dashboards. Asset clustering groups ads that share the same footage or audio so you can isolate which single change, a hook swap or a new CTA, actually moved ROAS, which is step five without the manual setup. You can read more on that approach in the guide to asset clustering and ROAS.
Ask the data a plain question Segwise's Creative Strategy Agent lets you ask things like "what is different about my top five creatives versus my bottom five" and answers from your own tagged data, so finding the winning element is a question, not a spreadsheet project.
The shift is simple to state and hard to do by hand: stop asking which ad won and start asking which element won, because the element is the thing you can repeat. Do it manually to understand it. Automate it so you can actually keep up. If you want the element-level view without building it yourself, that is where tagging plus the MMP join earns its keep.
For the wider picture on measuring and improving creative, see the full guide to creative analytics, and if you would rather compare dedicated tools, the roundup of creative tagging solutions is a good next read.