Why 39% of AI Systems Disagree on Brand Recommendations: Insights from 2,881 Prompts

Why 39% of AI Systems Disagree on Brand Recommendations: Insights from 2,881 Prompts Here’s the awkward truth most businesses haven’t wrapped their heads around yet: AI doesn’t offer universal answers. What ChatGPT recommends isn’t guaranteed to match Google’s AI results. And don’t even get me started on Perplexity or Claude. Across our dataset of 2,881 prompts, we found that 39% of the time, AI systems give conflicting brand recommendations for identical queries. Let that sink in. Nearly four out of ten searches generate divergent results depending on the AI you’re asking. So, why’s this happening? And more importantly, what are the implications for businesses trying to be visible in this fragmented AI world? It’s a question I’ve been obsessed with since day one at Contxt, and frankly, the deeper we dig into the data, the more complicated it gets. But that’s a good thing, because understanding these discrepancies is how brands can actually strategise for better AI visibility. Let’s break it down. Why AI systems disagree First things first: AI brand recommendations aren’t governed by some universal algorithm. Each AI system, whether it’s ChatGPT, Google AI, or Claude, has a unique architecture, training approach, and way of processing prompts. They’re not speaking the same language, even when answering the same question. Here’s what I keep seeing across our experiments: Training data variation: Some models are trained on vast datasets (ChatGPT, looking at you), while others, like Google AI Overview, lean more heavily on real-time web crawling. It massively impacts what they “know” about your brand. Query interpretation: AI systems don’t interpret prompts identically. A query like “What’s the best industrial pump manufacturer in the UK?” may trigger brand visibility on ChatGPT but fail to on Google AI if the phrasing doesn’t align with their indexed content. Biases and filters: Each AI provider applies its own filters, some explicit (like anti-spam mechanisms) and so