Why 37% of AI Systems Disagree on Brand Recommendations: Key Insights for Optimizing LLM Visibility

Why 37% of AI Systems Disagree on Brand Recommendations: Key Insights for Optimising LLM Visibility Here’s a stat that stopped me in my tracks: 37% of the time, ChatGPT and Google AI Overview give conflicting recommendations for the same brand . Let that sink in. If your brand is visible on one platform, there’s more than a one-in-three chance it’ll be invisible on the other. For businesses relying on AI-driven discovery, that’s not just a curiosity , it’s a potential disaster. Across our dataset of 3221 AI-generated prompts, we tested how different generative AI systems , including ChatGPT, Google AI Overview, Perplexity, and Claude , respond when asked about the same brands. The results were eye-opening. Some brands dominated consistently, but others practically disappeared depending on the system. And that divergence isn’t random; it reflects how differently these models are trained, updated, and tuned. In this article, we’ll get into the nuts and bolts of why AI systems disagree , why that disagreement matters for your business, and most importantly, how to tackle this problem head-on. If you’re serious about LLM visibility optimisation, you’ll want to read to the end. Trust me. Illustration of ChatGPT and Google AI Overview side by side, split by "37% divergence" What’s Causing the 37% Divergence? Let’s start with the obvious question: why do AI systems give different answers about the same brands? Isn’t AI supposed to be consistent and data-driven? Well, in theory, yes. In practice, it’s a bit of a minefield. Here’s what’s really going on: Training data: Every AI system is trained on distinct datasets. Google AI Overview, for example, may favour sources from its own search ecosystem, while ChatGPT’s training mix includes a broader scrape of the internet. What gets included , and what doesn’t , massively shapes brand visibility. Model updates: ChatGPT, for instance, openly operates with knowledge cut-offs (e.g., September 2021 for some versions), whereas Google