Marketing

L'Oréal built CreAItech, an in-house system that pipes together AI models from multiple providers — Seedance, Google's Gemini, Veo3 and Imagen, Adobe, Stable Diffusion and, as of this year, OpenAI — into one production workflow, so a marketer can switch models depending on what a given asset needs rather than being locked into one vendor's output. It runs on Nvidia infrastructure L'Oréal built out specifically to handle the volume once usage outgrew a single system. On the company's full-year earnings call, CEO Nicolas Hieronimus told analysts staff had used the CreAItech suite to produce 50,000 marketing assets in the roughly year since it rolled out, at 40% lower production cost than before — framing it modestly for a headline number: 'We have to learn to master the tools, but it's really a new era.' The new OpenAI deal folded into the system also gives L'Oréal access to a life-sciences reasoning model for its research side, and lets the company feed its own current product data into ChatGPT's underlying models, so AI assistants answering questions about L'Oréal products draw on the company's own information rather than whatever is floating around the open web.

Working with the studio Silverside, Svedka rebuilt its long-running robot mascot, Fembot, and introduced a new companion character, Brobot, dancing together at a party for what the brand calls the first primarily AI-generated national Super Bowl commercial. 'Primarily' is doing real work in that description: humans still wrote the storyline and creative direction, and reconstructing Fembot meant training the underlying AI systems to replicate her specific facial expressions and body movements convincingly enough to hold up on the biggest screen of the year, not generating a generic robot from scratch. Getting a consistent, ownable character to move and emote correctly took about four months of iteration, not a single prompt.

Tool used Silverside

Chief communications officer Leslie Cafferty said the fix cuts against a decade of social-media-trained instinct: AI assistants need real substance to quote, not a punchy one-liner. Booking is adding meaningfully more detail to its own web pages so there's something worth citing when an AI assistant answers a travel question, moving past the 'short and pithy' copy that wins on Instagram but gives a language model nothing to work with. Other marketers making the same shift described it in similar terms: Getty Images' Susan Nomecos summed up the underlying risk as 'garbage in, garbage out' for any brand feeding thin content into these systems, and JPMorgan Chase took a parallel approach on the listening side, training AI to read 'nuanced emoji, sarcasm, and slang' in social conversation well enough to catch sentiment that traditional keyword monitoring misses entirely.

The savings show up in repeat work, like resizing and localising. One-off flagship work still takes months.

Also worth a look. Volume content production, One-off hero campaigns

• Discussion 1
• Discussion 2

Google updated this guidance in December 2025. The penalty targets thin pages made in bulk, not the tool.

Also worth a look. Publishing AI drafts with real editing, Publishing AI pages at volume

• Discussion 1

Coca-Cola took the same criticism and ran the ad anyway, saying it tested better than almost anything before it.

Also worth a look. McDonald's Netherlands, Coca-Cola

• Discussion 1
• Discussion 2

Track mentions and citations, not clicks. Digiday reported the median company appears in 16% of relevant AI answers.

Also worth a look. Rewriting site content for depth, Tracking mentions and citations

• Discussion 1
• Discussion 2

Neither company published a number. Both built the capability in-house rather than buying a finished product.

Also worth a look. Consumer insight and listening, Product imagery for e-commerce

• Discussion 1
• Discussion 2