AI-Driven Ad Campaign Automation: LLMs Enabling Fully Automated Product Ad Placement

Introduction

Nowadays, AI is widely used in AdTech for dynamic ad personalization and traffic filtering. But recent advances in AI offer the promise of doing more. We believe that in the coming year, AI will also be used for ad campaign configuration. Currently, ad campaigns are primarily configured manually by individuals working in AdOps but AI and ML may greatly reduce the manual effort associated with campaign setup. Theoretically, AI can handle tasks like the following:

• adjusting bidding strategies based on real-time performance data

• allocating and optimizing multichannel budgets

• performing automated A/B testing of ad or campaign configurations to find optimal settings

• segmenting audiences dynamically or defining IAB content or audience segments in real time with LLM development
  
  

These are just some examples of potential AI automation in ad campaigns. Now, let’s discuss the real product placement campaign configuration problem that AI is helping to automate on one of our projects.

The Current state: Manual Campaign Configuration

One of our clients across multiple AdTech products and solutions has a product targeted to e-commerce and content creators. It allows content creators to monetize their content by embedding relevant product ads into blog posts, articles, and other types of content. Currently, the relevance is defined by humans—the person who configures the campaign needs to pick products from a giant catalog. These products will then be embedded into the text as ads. You can imagine the scale of the manual work, because every blog post or article may have a different topic and require its own set of products per ad placement.

How Can AI Help?

Previously, we talked about the LLM capability to understand text and implemented a proof of concept (PoC). The PoC showcases an ability to understand the topic of the content, turn it into standard IAB categories, and use them for targeting in real-time.

To automate this process, we also decided to leverage vector search solutions and got great results on test pages, and now, we are about to proceed with real-world production testing with live publishers. The solution was able to completely and automatically configure products for page-specific product ad placements without any human intervention. Of course, in the future, we are planning to add a user interface(UI) for humans to be able to tune certain LLM behavior, introduce a blacklist of products that the publisher doesn’t want to see on the page, and in general provide more control over the solution.

Solution components

The solution consists of multiple components. The first problem to solve was to grab the content of the page and prepare it for analysis.

Then, the LLM analyzed the page content to extract key meaning and potential keywords that are relevant to products and that best represent the content.

The third step was to search for relevant products and keywords. Product search is not an obvious topic, because semantic search works better with generic product descriptions, but if the content mentions product or brand names or exact models, then lexical search works better. So, a hybrid search that combines lexical and semantic (vector) search gives the best results in terms of relevancy.

And finally, the given list of relevant products was used as a campaign configuration for specific ad placements.

Of course, the system is wrapped with multiple metrics, so we can compare the efficiency of human-based and AI-configured campaigns. Also, to improve the user experience and optimize costs, we have added multiple caching layers. The system is capable of working end-to-end in real-time; however, the content of a specific blog post often doesn’t change after it is posted, so caching may save costs on the usage of LLMs and other infrastructure and also greatly reduces product embedding latencies.

Conclusion

This was just an example where AI was able to completely configure a product ad campaign without human intervention, and we believe that similar solutions will be the trend in the next year. There are many use cases where AI can reduce the reaction time or manual work of ad campaign configuration. Of course, some knobs and handles may be needed on the user interface side to keep the AI within certain boundaries, but this kind of solution opens great opportunities for ad campaigns to automatically configure themselves and update their configurations based on performance or typical conversion path.

Author: Alexey Rosolovsky