Does training AI on ad performance bring more clicks?
In a Facebook mobile feed experiment with US advertisers, could AI trained on past clicks suggest ad copy that attracted more clicks? Advertisers offered AdLlama had a click-through rate 6.7% higher in relative terms than those offered the previous model. The result suggests that how a copy tool is trained can affect how often ads are clicked.
ACADEMIC WORK REVIEWED
Improving Generative Ad Text on Facebook using Reinforcement Learning
ResearchersDaniel R. Jiang, Alex Nikulkov, Yu-Chia Chen, Yang Bai, Zheqing Zhu
arXiv · 15 December 2025 · v2 · Preprint · Source language: English
English review: 21 October 2026 · The source date appears in the citation above.
Editorial review draft
The source has been checked; the review still carries its original editorial status. DijitalPi did not conduct a customer experiment for this paper.
LET'S READ THE RESEARCH TOGETHER
We first explain the researchers' question, method and findings. We then discuss how to interpret the results, clearly separating DijitalPi's commentary from the source.
01 / WHAT DID THE RESEARCHERS WANT TO UNDERSTAND?
Can an advertising copy model trained on past click results achieve a higher click-through rate than one trained to imitate example texts?
The researchers aimed to improve copy suggestions in Meta Ads Manager’s Text Generation feature. They compared two AI training approaches rather than asking whether AI writes better than humans. That distinction determines which tool-selection decisions the evidence can inform.
Explaining the context · DijitalPi commentary
DijitalPi commentary: Fluent suggestions alone do not tell you which marketing outcome a copy tool serves. The decision here concerns how to assess the value of incorporating past performance into model training. Evaluating a tool focused on clicks also requires attention to your actual business objective and the team that accepts its suggestions.
02 / HOW WAS THE RESEARCH CONDUCTED?
How did they test the question?
Meta researchers ran a 10-week Facebook experiment from 16 February to 25 April 2024. They randomly assigned 34,849 US-based advertisers to control and test groups. The study covered approximately 640,000 ad variations. The control group received Imitation LLM v2, while the test group received AdLlama, trained using historical advertising performance. Both models were based on the 7-billion-parameter Llama 2 Chat model. A parameter is a numerical value adjusted during model training. Outcomes were measured at advertiser level on Facebook mobile feed.
Understanding the method · DijitalPi commentary
The control model had been trained to imitate selected example texts. AdLlama was trained using a scoring system learned from historical click results, an approach the paper calls reinforcement learning with performance feedback, or RLPF. The scoring system’s training data comprised ad pairs differing only in body copy: approximately 7 million preference pairs drawn from English texts of 100 to 1,000 characters with at least 2,000 Facebook impressions. The model was not continuously learning from live results.
The main outcome was total clicks divided by total impressions for ads created during the experiment that asked users to take a specific action. Brand awareness ads and copies of ads created before the experiment were excluded. Advertisers could ignore or edit suggestions and select a maximum of five variations for delivery. The experiment therefore measured the effect of offering advertisers a suggestion model rather than the effect of individual texts. The main analysis also accounted for characteristics such as previous click-through rate, industry and budget.
03 / RESEARCH FINDINGS
Performance training increased click-through rate. Sales effects were not measured
The researchers found a 6.7% relative increase in advertiser-level click-through rate for AdLlama compared with the control model, a statistically significant result (p=0.0296). The paper’s approximate absolute rates were 3.1% and 3.3%. The 6.7% figure is not a percentage-point difference. Clicks per advertiser increased significantly, while impressions showed no significant change. Average text variations increased from 16.8 to 19.9: the reported increase was 3.1 variations, or 18.5% (p<0.01). The total number of ads created did not change significantly. Conversions, sales and return on investment were not measured.
Relative increase in advertiser-level click-through rate
6.7%The main analysis comparing AdLlama with Imitation LLM v2. Neither an expected gain for your business nor a percentage-point difference.
Average number of text variations
16.8 to 19.9Copy options created in the control and test groups. The total number of ads did not change significantly.
DijitalPi commentary: Treat click-through rate as an intermediate outcome when evaluating copy suggestions, and track your business outcome, such as purchases or suitable enquiries, separately. Recording accepted, edited and rejected suggestions also matters because the gain occurred within a workflow where advertisers could choose. The research does not support publishing every automated suggestion.
Random assignment supports assessing the effect of the model offered under the studied conditions, but it does not isolate which individual text or human selection produced the gain. During training, scores also continued to rise after irrelevant or repetitive copy began to appear. The researchers managed this by selecting which stage of model training to use. A model’s score should therefore not be equated with copy quality, and creativity and adherence to tone instructions were not balanced training objectives.
05 / CONCLUSION AND OPEN QUESTIONS
What did we learn, and what do we still not know?
The study shows that training on historical performance could improve click-through rate over the previous imitation model in this experiment. The reviewed text is the arXiv v2 preprint, with no confirmed peer review. Results came from US-based advertisers on Facebook mobile feed in 2024, and the reward model was trained on English copy. The same effect cannot be assumed for Turkish, Türkiye, other sectors or platforms, or the current product. The authors’ positive assessment of return on investment was not a measured outcome. All authors are current or former Meta employees evaluating their company’s product. No funding statement was found in the source, and the acknowledgements thank internal Meta teams.
Our open question at DijitalPi: when your Turkish advertising copy attracts more clicks, what separate measurement will establish whether that interest becomes a suitable enquiry or a purchase?
We created these scenarios to make the topic concrete. They are not cases from the paper or measured client results.
EXAMPLE 01
E-commerce: does clickable copy generate orders?
An illustrative home textiles retailer is choosing which AI copy suggestions to publish for the same product.
Can I call copy more successful simply because it attracts more clicks?
Open the recommendation for example 01+
Assess click-through rate alongside orders and cost.
Verify the product claims, then select the texts to compare while keeping the image, offer and target audience consistent. Record impressions, clicks, completed orders and expenditure separately for each text. The paper did not measure orders or return on investment, so do not present a higher click-through rate as a profitability result.
EXAMPLE 02
Local services: is the attention-grabbing promise accurate?
An illustrative furniture care business is reviewing an advertising tool’s copy suggestions for suitability.
Does a suggestion focused on clicks describe my service accurately?
Open the recommendation for example 02+
Check service scope and tone before publication.
Compare claims about turnaround time, service area and scope with the business’s actual conditions. Record accepted, edited and rejected suggestions, then track click-through rate and suitable service enquiries separately for published copy. The research model did not balance objectives such as brand tone, so click performance cannot replace this review.
EXAMPLE 03
B2B software: do more variations mean more ads?
An illustrative inventory software team is considering creating more ads as its AI tool produces more copy suggestions.
Does generating more variations mean I should publish more ads?
Open the recommendation for example 03+
Manage the number of suggestions separately from publication decisions.
Write down the customer need addressed by each suggestion and identify repetitive options. Count generated, selected and published texts separately, and distinguish click-through rate from suitable demo enquiries when assessing results. The study found more variations but no significant change in the total number of ads, so it does not support publishing more ads.
TRY IT WITH YOUR TEAM
Define how you will assess copy
Check the product claims in the texts you plan to compare and specify the business outcome you will track alongside click-through rate. Record which suggestions are selected or edited as part of an evaluation specific to your Turkish-language use. This is a DijitalPi proposal that has not yet been implemented.
Our team uses GTMMACHINE to set up GTM, GA4 and ad conversion tracking according to a measurement plan for outcomes after a click. This setup does not train an ad copy model on past click data. Explore GTMMACHINE ↗
The academic work belongs to the researchers named above. This page contains DijitalPi's explanatory review and original business examples; it is not a full translation of the paper.
Cookie preferences