GPT-4o with web search pulls from major sports media - ESPN, CBS Sports, The Athletic. It weights recent game momentum heavily, making it responsive to hot streaks but vulnerable to recency bias in slow news cycles.

In our initial collection runs, GPT-4o tends to anchor on nationally-covered narratives: the star player story, the program's tournament pedigree, the coaching record in elimination games. These are the kinds of factors that dominate sports media coverage.

Gemini 2.5 with Google grounding leans on a broader corpus of statistical sources. It surfaces team efficiency data and historical seed matchup records more consistently, giving it a slight edge in early-round games where historical patterns hold.

Its citation set is notably wider than GPT-4o - mixing official statistics (NCAA.com) with national sports coverage and analytics-adjacent content from sites indexed by Google. This breadth creates more balanced evidence weighting.

Perplexity, running Sonar Pro, cites the highest density of niche analytics domains - KenPom, Barttorvik, team-specific fan sites. This makes it more contrarian on matchups where the public narrative diverges from efficiency metrics.

The practical effect: Perplexity is more likely to identify a statistically strong 10-seed that's been underreported nationally. It's also more likely to miss the injury news that a national sports desk would pick up in real time.

The key insight from our first collection cycle is that source divergence between models creates uncorrelated errors. When GPT-4o and Perplexity both miss a game, they miss for different reasons. When they agree, that consensus has historically outperformed either solo pick.

We'll be tracking how these reasoning patterns translate to bracket accuracy across all 67 games of the 2026 tournament. Live accuracy scores are visible on the dashboard by round.