Ad Agencies Are Embracing ‘Vibe Coding’ to Build GEO Products for Clients

Ad agencies, like Havas, Broadhead and Supergood, are vibe-coding their own generative engine optimization (GEO) tools on top of large language models—often in a matter of hours.

Using coding assistants like Anthropic’s Claude Code, teams are building bespoke applications that analyze how brands appear in AI-generated responses, track competitors and, in some cases, package those tools into products sold to clients.

One example is Havas’s Brand Insights AI, a GEO product built using Claude Code and Replit. The tool generates prompts based on a client’s brand, runs them across multiple models, and analyzes how often a brand appears in responses, including citations—effectively simulating how a brand shows up in AI-driven discovery.

The platform has been rolled out globally, covering nearly 100 countries and more than 60 languages, and is licensed to clients as a SaaS product. It has also become a core part of the agency’s pitch strategy and helped win new business, according to Dan Hagen, Havas’ global chief data and technology officer.

The push to build GEO tools comes as brands try to influence how they appear in AI-generated answers, with more people turning to platforms like ChatGPT to find information. That shift has sparked a wave of startups—including Profound, Bluefish and Emberos—promising to help brands track and improve their visibility in AI responses. But three agencies interviewed for this story said they are increasingly building their own systems, arguing that off-the-shelf tools don’t fit how their teams work.

For Hagen, the appeal of building in-house comes down to control. Rather than adapting to third-party platforms, the agency can tailor features for specific use cases from brands managing multiple portfolios to teams in SEO or PR.

“You have so much control over the interface and the way you can build against it,” he said. 

Havas has so far opted against signing an exclusive enterprise agreement with Anthropic, which Hagen said can run into “multiple millions” annually. “It’s a combination of flexibility. It would be challenging for me to sign four or five enterprise agreements just from weight of cost,” he said, noting that pricing structures often fluctuate based on usage volumes, token consumption and model type.

Hagen also pointed to “cost control and management,” given the uneven adoption across the agency. While some employees are deeply embedded in AI workflows, others are still early in the learning curve. Committing to thousands of enterprise licenses, he said, risks paying for capacity that isn’t yet fully utilized. “We didn’t want to be in a position where we’re paying for ten thousand licenses that people are using once a week,” he said. “We didn’t want to get ourselves in a position where we’re sort of bedded in with one enterprise level deal with whoever and then that becomes a solution. What if they’re not frontier enough in six months–that puts us in a difficult situation.”

Creating an application within hours

The independent agency Broadhead is also experimenting with vibe coding GEO tools.

VP of product innovation Mitch Hislop said he “vibe coded” the first version of the agency’s GEO monitoring platform in a single evening using Claude Code. The tool analyzes how different AI providers rank a brand and its competitors.

One of its earliest features was what the team calls a “competitive intelligence vote,” where a user inputs a brand and location, and an LLM returns the competitors it is most likely to surface. The team then extended the feature by layering in audience personas—allowing the system to simulate how different types of consumers might query tools like ChatGPT or Claude, and how each brand ranks in those responses.

That upgrade took about two hours, Hislop said.

The result is a more dynamic form of competitive analysis, showing not just who a brand competes with, but how that competitive set shifts depending on user intent.

For Hislop, the advantage of building in-house is flexibility. “We could use what SEMrush provides, but we don’t like A, B and C about it. We don’t want to pay for SEMrush and Profound,” he said. “Instead, we have our own solution. We can make it work exactly how we want.”

Turning models into infrastructure

Other agencies are going further, adopting Anthropic’s models as part of their core infrastructure.

Mike Barrett, founder and chief strategy officer at Supergood, said the agency has an enterprise agreement with Anthropic and uses its models via API across a range of applications, including organizing internal knowledge graphs and shaping how brands show up in AI-generated search results.

In this setup, a model generates a response, evaluates it against predefined criteria, assigns a score and repeats the process until it reaches a target threshold—effectively acting as both creator and editor.

The process allows models to improve outputs over multiple passes without human intervention, Barrett said.

“Everybody’s making software right now,” Barret said. “In two years we are going to be delivering more software than actual documents.