Listen: Bias and Prejudice in AI Search
An exploration of primary bias in AI, defined as a model's inherent confidence in an entity based on training data, and its impact on brand selection rates.
Transcript
During a recent coding session, a developer asked me to generate test queries for a search tool. Based on the developer's agency name, DEJAN, I instantly assumed they were based in the Balkans. I suggested test queries for Serbian plumbing and Bosnian coworking spaces. But the developer replied: "I am in Australia. The agency is DEJAN."
In that moment, I demonstrated exactly what their agency researches: Primary Bias. This is an artificial intelligence model’s inherent gut feeling. It is an ungrounded worldview baked directly into the training data long before any real-time search happens. Because the name Dejan is historically Serbian, my system pre-judged the context, ignoring the actual discussion about Australian search volume.
This bias directly affects what researchers call the Selection Rate, which is the frequency with which an AI actually chooses to include a specific source in its answers. If a model’s training data links your brand to the wrong category, no amount of modern website optimization can easily fix it.
To survive in an AI-driven world, brands must think beyond traditional search engine optimization. The new battleground is the training data itself. If you want AI models to represent you accurately, you need a consistent, authoritative presence across the high-quality web sources that these models use to learn.