A/B testing

Instagram’s visible experiments have included changes to profiles, navigation and viewing experiences.7910 Instagram’s profile announcement specified that rearranging icons, buttons and tabs would leave the photos and videos in the profile grid unchanged.7 In September 2020, Mosseri announced tests of three versions of Instagram’s home screen, placing Reels and Shop tabs in alternative layouts in response to changes in video viewing, shopping and the role of creators.10 His June 2021 recommendation experiment included choices about topics people wanted to see more or less of; those choices belonged to that historical test.3 Mosseri invited participants in the May 2022 Feed experiment to comment, and his July explanation said the full-screen experience was not yet good enough to release to everyone.59 Meta’s Newsroom described a limited Reels-first mobile-app test in India in September 2025 that opened directly in Reels and included a Following tab offering recommended content from followed accounts, content from mutual follows, or posts and Reels in newest-first order.11

Meta’s engineering team, writing in Instagram Engineering, described how preparing alternatives came before comparing groups in its video-ad processing test.2 The team generated alternative video encodings—processed versions of the videos—only after ads were successfully published, so preparing the test did not affect ad creation.2 Meta’s engineering team then used ads with both old and new versions available to give the new versions to half the users and the old versions to the other half.2 The team reported that the new versions performed much better, providing evidence for the rollout.2

Meta’s research uses online A/B comparisons to measure the results of recommendation and advertising alternatives.41213 For Instagram short-video recommendations, Meta’s OneShot research reported gains in sessions, time spent, post impressions and likes, including a 0.136% increase in time spent and a 0.251% increase in likes.4 For advertising recommendations serving Facebook and Instagram users, Meta’s research compared several configurations against a smaller model called MoNN Small, with the strongest reported configuration improving its online ads measure by 2.57%.12 Meta’s recommendation research also tested an arrangement in which a shared large model summarized a person’s interest in each possible item, and smaller models for different recommendation areas used those summaries alongside their own information.13 That Meta research reported an online A/B comparison against a baseline without this shared-model-then-smaller-model arrangement, as well as cumulative gains from later versions, so those two sets of results describe different comparisons.13

Meta’s SlimPer research extends the evaluation of Instagram Feed and Reels changes beyond engagement by checking whether other measures get worse.6 Meta’s research reported no deterioration in checks covering integrity, diversity and how recent the content was, alongside significant engagement gains.6 Meta’s research also reported approximately neutral training and inference GPU capacity.6 The Meta paper described pre-testing and backtesting before release to full traffic on Instagram Reels and Feed.6