The Operations Bridge

There’s a controlled study worth knowing about if you use AI to analyze your own data: researchers fed an LLM 1,000 records in 11 different formats — CSV, JSON, XML, YAML, Markdown, and others — then tested how accurately it could answer questions about that exact same data.

Same data. Same questions. Same model. The only thing that changed was structure.

CSV — the default almost everyone reaches for — came in at 44.3% accuracy. A markdown key-value format hit 60.7%. That’s a 16-point swing in accuracy from formatting alone, with nothing about the underlying data any different.

This is the part of “clean your data” that usually gets skipped. Most people hear “garbage in, garbage out” and stop at removing errors and duplicates. But you can have flawless, verified data and still get a materially worse answer, just because you handed it to the model in the wrong shape. Integrity isn’t only about accuracy — it’s about structure the model can actually read well.

Worth noting: the tradeoff runs both ways. The best-performing format also used nearly 3x the tokens of the leanest one — more cost, more latency, for the accuracy gain. That’s not a reason to ignore format. It’s a reason to be deliberate about the choice instead of defaulting to whatever’s easiest to export.

So what — the export button gives you CSV by default. That doesn’t mean it’s what your AI understands best. Now what — before your next AI-assisted analysis, ask what format you’re actually feeding it, and whether anyone chose that on purpose.

Source: “Which Table Format Do LLMs Understand Best? (Results for 11 Formats)”