LO
DeploymentCognitive automation2024-04-25
Meta trained its ad-text generator on the results of advertisers' own copy tests; in a 10-week trial with 34,849 advertisers, click-through rose 6.7% and advertisers made 18.5% more variations
Copywriteroccupation page →
Event date / reported
2024-04-25 · reported 2025-07-29
Evidence stage
DeploymentAn employer has put it into production. Can move the baseline — weighted by scale and how similar the setting is.
Tasks this bears on
Reading results and iterating
Understanding why one variant converted and another did not, and turning that into the next brief.
Being augmented✓ Evidence-backed
Where this applies
A paper by Meta researchers about the ad-text generator in Meta's own ads product. The model was trained on pairs of ads in which advertisers changed only the text, treating the variant with the higher click-through rate as preferred, so the results of copy tests became the training signal; in a randomised test from February to April 2024 across 34,849 US advertisers and about 640,000 ad variations, it raised click-through by 6.7% against the earlier model, and advertisers using it created 18.5% more variations. Advertisers must still select each variation before it runs. Meta sells both the ads and the tool.
What this means
Reading which copy won and feeding that into the next round is being built into the generator itself at one of the largest ad platforms: past test results train the model. The advertiser still chooses what runs; the loop from result to next draft is increasingly automatic.
What it does not yet show
One platform's own tool and trial; it does not show fewer copywriters or how agencies read results.
What you can check
Open arXiv:2507.21983 and find "AdLlama improves click-through rates by 6.7%".
Does it change the assessment?
No. The impact index is never moved by a single event. What this record did: the 1 linked task judgement above now rest on evidence instead of inference.
Source
Meta — "Improving Generative Ad Text on Facebook using Reinforcement Learning", arXiv:2507.21983 (submitted 29 July 2025; revised 15 December 2025) · verified 2026-09-27 · Claude (VOLO agent) · interpreted 2026-09-27 · Claude (VOLO agent)
Primary source — published by the party that did this, or the authority of record. No co-signature needed.