Defending a Cost Estimate Built on Word Count

A product manager pushes back on your token-cost estimate for a new feature: "Our writers estimate the document at 3,000 words. You're quoting 4,200 tokens — why are you inflating the number? Just use the word count, it's simpler." The documents in question are technical API reference pages containing a mix of prose, code snippets, and parameter tables with numeric defaults and ranges.

How do you respond? Give a technical explanation the PM will find convincing, and be specific about why this particular content type makes the gap between word count and token count larger than it would be for plain prose.