An ad can bring in more orders while giving you less reason to change your campaign. If it also received more impressions, more of the cheaper placements, or more time for orders to appear, the total alone tells you very little about the creative. Start by finding out what each version was actually shown against. That gives you a better question than which row finished first.

Start with what was eligible to run

Save the configured split and pull the delivered impressions, clicks, spend and selected conversion event for the same dates. Keep the timezone and attribution settings with the export. If the test platform reports assignment separately from delivery, keep both columns. A planned split is a record of the setup. Impressions and spend are records of what the auction actually delivered.

Google Ads explains that its configured traffic split governs auction eligibility and that impressions and spend can still differ. The auction rules on another platform may work differently, so check what its experiment control actually allocates before interpreting the numbers. See the Google Ads custom experiment documentation for a concrete example. https://support.google.com/google-ads/answer/10683687?hl=en

Put the counts next to the rates

Suppose a store is testing two ads for the same product page. Version A receives 20,000 impressions, 400 clicks and 20 completed orders. Version B receives 10,000 impressions, 250 clicks and 15 orders. A has more orders in total. B has a 2.5 percent click rate against A's 2 percent and a 6 percent order rate per click against A's 5 percent. These are illustrative counts. They tell you what happened in this export and give you a starting point for checking the delivery mix.

Keep the denominator beside every rate. A row with two orders from ten clicks can look exceptional in a sorted report. Put it beside a row with twenty orders from four hundred clicks and the different amount of evidence is obvious. Keep spend beside order count too, especially when the versions entered auctions with different prices. A version with no orders has no observed cost per order for that window. Leave that field blank and retain the spend and exposure so the reader can see what the row represents.

Look for overlap before attributing a difference

Split the report by the most plausible source of uneven delivery. For an inventory change, start with placement or format. If one ad started late, start with dates when both were eligible. Within a useful slice, place the two versions side by side and keep their counts visible. A placement where only one version appeared can explain part of the aggregate while offering no side-by-side comparison.

FroggyAds describes per-creative reporting across its push, native and display formats. If a report combines those formats, keep them separate while you check whether both versions had meaningful exposure in each. That makes it easier to tell whether a promising difference came from the creative or from where it ran. https://froggyads.com/creative-testing/

Keep the decision as narrow as the evidence

Check whether the newest dates had the same chance to collect orders. If order counts are still moving, keep the export as a dated snapshot and repeat the same report after the outcome window has matured. If the delivery mix changed during the test, record when it changed and check the overlapping period separately. This is also the moment to check that both creatives were approved and pointed to the intended destination. A rejected ad or a landing page change can make a tidy comparison misleading.

If the next test is about the headline, leave the rest of the ad and the destination fixed where your setup permits it. Google Ads describes ad variations as a way to compare a single text change in its responsive search workflow. The practical point for this audit is to keep the follow-up narrow enough that the next report can answer the question you wrote down. https://support.google.com/google-ads/answer/7438541?hl=en

Write the decision beside the export. For example, keep A running and retest B on the placements both versions reached, using the same order event and a fixed comparison window. Include the reason for the retest and the date when someone should look again. The next person should be able to open the report and see which figures support that decision.