A/B Testing for Product Feeds

A/B testing for product feeds is the process of running two or more versions of your product data against each other to see which performs better on a channel. With Feedoptimise you can test any attribute, such as titles or images, measure each version on real channel data, for example from Google Shopping, and apply the winning version to your live feed.

You can test any attribute in your feed, including titles, descriptions, images, product types, prices and custom labels. Any modifier on the platform can generate the test version, so AI titles, edited images, formulas and overrides can all be tested.

Yes. A single experiment can modify one attribute or several at once, for example a new title and a new image together, and you choose exactly which fields each test changes.

The duplicated method clones the original product under a new ID and runs the test version alongside it. The rotated method keeps the same ID and cycles versions through it over time. The hybrid method rotates fresh copies while still comparing them against the original.

Yes. Each test has a start date and an optional end date, so it goes live and wraps up automatically. You can also replace the fixed end date with stopping rules, so the test concludes on its own once your chosen metric and confidence thresholds are met, or when its maximum runtime is reached. If you set both, the test ends at whichever comes first. Conditions let you scope it to certain categories or product ranges.

You set the stopping rules when you create the test: a minimum amount of a metric you choose per version, such as clicks, impressions or conversions, a statistical confidence level and a maximum runtime. For example, rotate the original and a new title until each has 100 clicks and the result reaches 95% confidence, with a 30 day cap. The experiment monitors its own data and concludes as soon as those conditions are met.

Confidence measures how unlikely it is that the difference between your versions is down to random variation rather than the change you made. At 95% confidence, the gap is large enough that chance becomes an unlikely explanation, which is what most people mean when they call a result statistically significant. Until that point the experiment keeps running instead of calling a winner on a few noisy days of data.

They describe the same idea from two directions. A result is statistically significant when it is unlikely to be down to chance, and the confidence level is the threshold you set for that. An experiment set to 95% confidence will only declare a winner on a result that is statistically significant at that level. Feedoptimise uses confidence in the interface because you enter it as a number, for example 95%, and you can watch it build while the test is still running.

The maximum runtime acts as a safety net. When the timeout is reached without a statistically significant result, the experiment folds automatically to whichever version you nominated in advance, the original data or the tested version. Nothing is left running indefinitely and no inconclusive test blocks your feed.

Every version is measured on real channel performance such as clicks, conversions, CTR, ROAS and more, all at the product level. When the test ends, you can set Feedoptimise to implement the winning version directly into your live feed.

That is up to you. Each experiment can be evaluated at group level, one winner across the whole test, or at item level, where every product gets its own verdict on its own data and version B can be live on some products while version A stays on others.