The Automated Copywriter: Algorithmic Rephrasing of Health-Related Advertisements to Improve their Performance

Brit Youngmann, Elad Yom-Tov, Ran Gilad-Bachrach, Danny Karmon

Introduction

Public health authorities increasingly rely on search engine advertising to "nudge" populations toward healthier behaviors, such as cancer screenings or smoking cessation. However, the efficacy of these campaigns is often limited by the quality of the ad copy. Writing high-performing advertisements is a specialized skill requiring marketing expertise and extensive A/B testing—resources that are frequently scarce in public health sectors. Researchers at Microsoft Research have addressed this gap by developing an "Automated Copywriter," a deep learning system designed to rephrase health-related search advertisements to maximize user engagement.

Problem Statement

An advertisement's performance, typically measured by its Click-Through Rate (CTR), is highly sensitive to linguistic nuances, emotional resonance, and the clarity of its "call-to-action." In the context of public health, poorly phrased ads represent more than just a marketing failure; they represent missed opportunities for medical intervention. This research explores whether the copywriting process can be automated by training algorithms to identify and replicate the linguistic patterns that drive engagement in the health domain.

Methodology

The researchers framed ad optimization as a sequence-to-sequence (seq2seq) translation task. In this paradigm, a low-performing advertisement serves as the "source language," while a high-performing version of the same ad serves as the "target language."

The Pipeline

  1. Normalization & Preprocessing: To ensure the model learned generalizable structures rather than specific medical facts, the team used Named Entity Recognition (NER) via the SpaCy library. Specific terms (e.g., "Shingles vaccine") were replaced with generic tokens like <CONDITION/TREATMENT>. Dates, names, and monetary values were similarly abstracted.
  2. Candidate Generation: A 2-layer Long Short-Term Memory (LSTM) network was trained on pairs of ads triggered by the same search query, where one ad significantly outperformed the other. The model learned to "translate" the style of the low-CTR text into that of the high-CTR text.
  3. Selection & Ranking: To choose the best variation from multiple generated candidates, the system employed a LambdaMART learning-to-rank model. This model evaluated candidates based on features such as Flesch-Kincaid readability scores, sentiment analysis, and lexical diversity.
Examples of the ad transformation pipeline, showing the original text, the normalized version using generic tokens, and the final generated ad.

Datasets and Domains

The study utilized a dataset of over 114,000 English-language search advertisements from Microsoft Advertising, collected in early 2019. The research focused on two primary health domains:

  • Medical Symptoms (MS): Queries related to acute conditions like "vertigo," "cough," or "earache."
  • Preventative Healthcare (PH): Queries related to long-term health behaviors, such as "vaccine," "weight loss," or "stop smoking."

Ground truth for the training and evaluation was derived from real-world CTR data for ads that had been displayed at least 100 times.

Key Insights

  • Implicit Psychological Learning: Analysis of the generated ads revealed that the model independently learned to increase both Arousal (excitement/agitation) by 87% and Valence (positivity) by 36%. This suggests the algorithm "discovered" that emotionally charged, positive language is a primary driver of clicks in health contexts.
  • Action-Oriented Shifts: The model consistently injected more "Calls-to-Action" (CTAs). In the Medical Symptoms domain, it favored verbs like "learn" and "treat," whereas in Preventative Healthcare, it gravitated toward urgent directives like "quit" and "stop."
  • Structural Generalization: By using tokenization (e.g., replacing specific diseases with placeholders), the model successfully learned the underlying rhetorical structure of a persuasive health ad, allowing it to remain effective across different medical conditions.

Results

The system demonstrated significant performance gains in both offline simulations and live environments. In offline testing, the ranking model predicted that generated ads would outperform originals in 81% of Medical Symptom cases and 76% of Preventative Healthcare cases.

Comparison of Click-Through Rates (CTR) between original and generated ads across Medical Symptoms (MS) and Preventative Healthcare (PH) domains.

In a live validation campaign on Microsoft Advertising, the rephrased ads achieved a staggering 68% average increase in CTR compared to the original versions. The generated copy was found to be more concise and frequently included persuasive keywords related to convenience (e.g., "fast delivery") and advantage (e.g., "discounts").

Limitations and Future Work

While the results are promising, the researchers noted that CTR is a proxy for interest rather than a guarantee of behavioral change (conversion). Furthermore, the live validation was conducted on a relatively small sample of 20 ad pairs. Future research will focus on utilizing Transformer-based architectures (such as BERT or GPT) to improve linguistic fluency and automating the final "realization" step to eliminate the need for manual grammatical corrections.

Summary updated

Abstract

Search advertising is one of the most commonly-used methods of advertising. Past work has shown that search advertising can be employed to improve health by eliciting positive behavioral change. However, writing effective advertisements requires expertise and (possible expensive) experimentation, both of which may not be available to public health authorities wishing to elicit such behavioral changes, especially when dealing with a public health crises such as epidemic outbreaks. Here we develop an algorithm which builds on past advertising data to train a sequence-to-sequence Deep Neural Network which "translates" advertisements into optimized ads that are more likely to be clicked. The network is trained using more than 114 thousands ads shown on Microsoft Advertising. We apply this translator to two health related domains: Medical Symptoms (MS) and Preventative Healthcare (PH) and measure the improvements in click-through rates (CTR). Our experiments show that the generated ads are predicted to have higher CTR in 81% of MS ads and 76% of PH ads. To understand the differences between the generated ads and the original ones we develop estimators for the affective attributes of the ads. We show that the generated ads contain more calls-to-action and that they reflect higher valence (36% increase) and higher arousal (87%) on a sample of 1000 ads. Finally, we run an advertising campaign where 10 random ads and their rephrased versions from each of the domains are run in parallel. We show an average improvement in CTR of 68% for the generated ads compared to the original ads. Our results demonstrate the ability to automatically optimize advertisement for the health domain. We believe that our work offers health authorities an improved ability to help nudge people towards healthier behaviors while saving the time and cost needed to optimize advertising campaigns.
BibTeX
@inproceedings{youngmann2020automated,
  author = {Brit Youngmann and Elad Yom-Tov and Ran Gilad-Bachrach and Danny Karmon},
  title = {{The Automated Copywriter: Algorithmic Rephrasing of Health-Related Advertisements to Improve their Performance}},
  booktitle = {The Web Conference},
  year = {2020},
  doi = {10.1145/3366423.3380211},
}
Figure