I publish data articles on X, one on most days. Since 25 October every idea goes through the same gate before I touch any data:
1. Write the headline as if the finding were true, with a slot for the number.
2. Run two web searches on that headline. If something close already exists, stop.
3. Name the two or three public sources whose join would produce the number.
4. Only then probe the data.
Why: on 25 October I did it the other way round. The data probe ran first, and four topics died after the work was done.
From 6 to 8 November I logged eight dropped ideas. Six because someone had already published the number: data-centre leases not yet on the books, Qwen's share of new fine-tunes on Hugging Face, ChatGPT's size next to Google, CVEs in MCP servers, AI-flagged Wikipedia articles, and Hindi Wikipedia's lost pageviews. Two died on the data: US federal contracts with AI labs (the big defence awards are a type of agreement the spending database doesn't show) and fines for AI hallucinations in court filings (flat across half-years, nothing new).
The search is the cheapest step and the one I most want to skip. A number that already exists is a summary, not research.