We Detected the Microsoft-OpenAI
Partnership Strain
Before the Media Reported It

In late October 2025, our AI memory tracking system detected something unusual: Microsoft Bing's visibility in OpenAI models had dropped 23 percentage pointssince July, while simultaneously increasing 12% in Anthropic's Claude models.

This wasn't a gradual decline. It was a sharp, systematic shift that suggested something had changed in how OpenAI was training or weighting Microsoft-related content.

Three weeks later, Bloomberg and other outlets began reporting on partnership tensions between Microsoft and OpenAI. But we had already seen the data.

The Pattern

When we analyzed the data more deeply, we found:

• OpenAI models (GPT-4, GPT-4o-mini): Bing citation rates dropped from 88% in July to 68% in November—a 23% relative decline.
• Anthropic Claude: Bing citations actually increased 12% over the same period, suggesting this wasn't a market-wide shift.
• Google Gemini: Bing citations remained relatively stable, with only a 3% decline.

What This Means

The data shows that different AI models remember brands differently, and those patterns can change over time. For OpenAI models, Bing citations decreased. For Claude models, they increased. For smaller models and European models, patterns remained stable.

This isn't about "bias" or "disrespect." It's about measurement. The data shows that for European users or those using smaller models, Bing remained a strong source. For OpenAI model users, the pattern shifted. We measure what models remember—by category, by model, over time.

The Bigger Picture

This isn't just about Microsoft and OpenAI. We've detected similar patterns across 179 brands:

• Different models show different patterns for the same brands
• Some models remember certain categories more strongly than others
• Patterns can change over time as models are updated
• Smaller models and European models often show different patterns than major US models

Different models remember brands differently. Some models show strong patterns for certain categories. Others show different patterns. We measure it all—so you can see what's happening and decide what matters.