数据与原创研究方法 · The PROVE framework — original research that earns links + AI citations

Design a study people have to cite. Each letter is a gate; skipping one is where studies fail (no data void, soft method, cooked analysis, buried finding, or no distribution).

P — Pick a question inside a data void

A claim worth proving where good public data doesn't exist and your audience + journalists + AI engines would cite the answer. Your unfair advantages, in order of strength:

1. Proprietary/internal data — your platform, your customers, your transactions. Highest originality; no one else can replicate it. (Anonymize before analysis.)
2. A recurring survey of your niche — become the annual source ("State of X"); the cadence compounds.
3. Fresh analysis of a public dataset — lower originality but real and fast; great when you have no list. Reject questions already answered well, questions you can't get real data for, and questions whose answer no one would repeat. The test: write the headline stat you hope to find — would a journalist or an AI cite it?

R — Run a sound method

Choose the method that fits the question, then design it before collecting:

• Define: population, sample frame (who's eligible + how you reach them), target n, recruitment, time window.
• Instrument: neutral, non-leading questions (not "don't you agree X is overpriced?"); one idea per question; answer options that don't bias.
• Plan the analysis up front (which comparisons, which segments) so you're not fishing later. Method ladder: proprietary-data analysis > representative survey > public-dataset analysis > simple experiment. The agent designs the instrument; the human/tool fields it and gathers the real data.

O — Observe honestly (collect + analyse without cooking it)

• Real data only. Never invent or AI-synthesize data points — the fastest way to lose the asset and the brand.
• No p-hacking (don't hunt until something is "significant"); no cherry-picking the flattering cut; only segment where n supports it.
• Disclose: sample size, dates, method, and limitations — prominently, not in fine print. Small n? Frame as directional, not definitive; never say "most people" off 42 responses.
• Conflict of interest (a self-funded study about your own category) gets disclosed, not hidden.

V — Voice the one finding that travels

Find the surprising-but-defensible headline — one number people will repeat — fully supported by the data and never inflated (38% is not "nearly half"). Shape it: a clear claim (X% of Y do Z), the why it matters, and a memorable name/number. One hero stat carries the study; 2–3 supporting findings back it. Counter-intuitive beats confirmatory — but only if the data truly says it.

E — Engineer for citation, then distribute

Build the asset to be cited and then spread it:

• Report page: visible methodology + sample + date, a "Last Updated" stamp, clean charts, and a copy-paste stat box with an attribution link (so anyone quoting you links back).
• Atomize into cuts → the format writers (carousel/thread/reels/caption) + infographic-and-data-viz.
• Distribute wide: pitch journalists the news angle; seed across multiple publications (distribution is the citation multiplier); refresh on a cadence (beat the freshness cliff).
• WoopSocial publishes the cuts + promo posts and drives traffic to the report; measure citations + saves + referral traffic, never a fabricated metric.