Scaling Creative With AI
This AI system personalizes Google's Holiday Shopping campaign across hundreds of NYC locations
If you’ve been walking around New York City lately, you might have noticed the Google Shopping ads on the many LinkNYC kiosks dotting the city’s streets and subway stations. A clever, neighborhood-specific headline is followed by a relevant product or two, and closes with a reminder that Google Shopping is where to find all the best holiday deals in one place.
What you may not have realized is that the ads are entirely synthetic, created by an AI system that Addition designed and developed using Google’s foundation models, Gemini and Imagen. Powered by local insights and data from maps and search, each ad is 100% unique and contextually generated to the location it’s running in: right down to the city block.
How It Works
To create AI content that doesn’t feel like it was generated by AI, we designed a system with a creative idea embedded in an ensemble of highly opinionated AI functions. Our creative approach was to capture hyper-local insights in short headlines and trending products that connect to the idea that Google Shopping has all the best deals in once place. Each AI function in our system is aligned with creative direction and examples to scale this concept ad infinitum.
Data Inputs
The AI system is designed to take in data to inform it’s outputs:
Media Data: A CSV file containing the address and latitude/longitude of each Link to use as initial input.
Google Maps: The system pings the Google maps API to get a detailed map image of the location.
Search Data: Search data highlighting trending products that we might want to feature in ads.
Local Insights and Generative Copy
All this data is now fed into a series of Gemini functions that produce creative insights about the location where the ad will be running. In this case, the insight needs to be something unique about the location: perhaps a nearby landmark, the types of people that frequent it, or simply the vibe. LLMs like Gemini possess a deep repository of this kind of information.
Next up, the system generates headlines informed by the insights. Using human-crafted examples and a chain-of-thought technique, each insight gets transformed into unique headline options.
Image Generation
The system uses Imagen3 to generate product images for the ads. To ensure consistent quality, we use AI vision functions and a human review step to create a library of hundreds of product images that are visually consistent and brand-safe. A subsequent step in the system identifies the most relevant products to include in any ad based on the location and creative.
Content Evaluation
Each of these preceding steps focus on scale. We want lots of options, which are then fed back into an evaluation system to review each ad for brand safety, location relevance, and creative quality.
Video Editing
The system takes the winning ads and runs them through an AI-enabled video pipeline that adjusts for copy length and seamlessly integrates the generated images.
Final Review and Trafficking
Brand managers at Google are given an interface where they can review the best ads on a map and approve them for deployment. Once approved, the system sends the ads directly to the LinkNYC displays through a custom API interface.
Realtime Content
While the goal of this system was to enable hyper-local relevance, it also gives us the power to turn ads on and off quickly, thereby introducing the opportunity to push out real-time messages. We experimented with this during the Macy’s Thanksgiving Day parade with a creative push that incorporated pre- and post-parade messaging.
From Efficiencies to Relevence at Scale
One of the biggest problems marketers have faced in the past decade is scale. The sheer scale of content that brands need to produce to maintain table stakes in a fragmented digital ecosystem is overwhelmingly time- and resource-intensive. As a result, most brands fail to lean into the opportunities that digital technology enables in terms of localization, personalization, and real-time strategies.
AI has the power to change this dynamic. Smart brands can leverage AI, not only to achieve table stakes when it comes to content production, but to lean into strategies that make their content more relevant. Achieving this without compromising creative quality or creating an ‘uncanny valley’ vibe requires careful planning. Most of all, though, it requires human creativity.
The new human creativity in this equation is more systematic than artisan. It’s not about a writer sitting down and crafting one headline after the next. It’s about providing enough instruction, examples, and high-level direction – informed both by creative art and copy, and a creative technology standpoint – to enable AI to scale your idea into a myriad of outputs that would not have been possible without machine intelligence.
Addition is an applied AI studio for modern brands.
Visit Addition.ml to learn more.