Does Context Format Matter? What 10,000 Experiments Revealed

Practitioners argue endlessly about whether to format context as JSON, YAML, Markdown or XML. A large-scale 2026 study put the question to the test. The answer is more liberating than most expect.

Spend time in any context engineering community and you will find passionate debates about format. Some swear by JSON for its precision. Others prefer YAML for its readability. Still others insist XML tags help models parse boundaries, or that Markdown is most token-efficient. Everyone has an opinion, and everyone assumes the choice matters significantly. A 2026 study with nearly 10,000 experiments tested that assumption directly.

The Study

The research systematically varied context format, YAML, Markdown, JSON, and others, across a large number of controlled experiments, measuring the effect on agent accuracy. If format mattered as much as the debates suggest, the data would show clear winners and losers. The scale of the study, nearly 10,000 experiments, was designed to detect even modest effects with statistical confidence.

The Surprising Finding

Format choice had no statistically significant effect on aggregate accuracy. Whether the context used one format or another, the models performed roughly the same. The passionate debates, it turns out, were largely about something that does not move the needle. This is one of those research findings that initially disappoints (no simple format trick to boost performance) and then liberates (you can stop worrying about it).

Why This Is Liberating

If format barely matters, you should not agonise over picking the theoretically optimal one. The mental energy teams spend debating serialisation is better spent elsewhere. Pick a format that is clear and consistent for the content at hand, structured data in JSON or YAML, instructions in Markdown, content needing clear boundaries in XML tags, and move on. The format is a detail; the substance lies elsewhere.

What Actually Matters

While format choice does not move accuracy, two related things do. First, content and selection: what information you include and how relevant it is matters enormously, far more than how you serialise it. Second, schema design, the actual structure of the information, distinct from its format. A well-designed schema that makes relevant fields explicit and relationships clear helps the model, regardless of whether you express it as JSON or Markdown. The insurance example that reached 95% accuracy succeeded because of a well-designed targeted schema, not because of a format choice.

The Takeaway

Do not over-think format. Pick something clear and consistent and spend your effort where the research says it counts: on selecting the right content, designing good schemas, and applying the four strategies. The format debate is a distraction from the work that actually determines whether your context engineering succeeds. Let it go, and focus on substance.