Editing Openai/69566be9-e9c0-8009-992a-722fd2764e7a (section)

=== Assistant: That rule-of-thumb is not marketing fluff, but it is also conditional. ===
That rule-of-thumb is not marketing fluff, but it is also conditional.

Where the ~2–4 pp comes from (roughly):
* Method comparisons by survey methodologists (e.g., work summarized by Pew Research Center) repeatedly show that well-weighted online panels (YouGov, Ipsos KnowledgePanel, etc.) often differ from probability-based telephone/ABS surveys by only a few percentage points on headline items (vote intention, broad policy support).
* YouGov’s own election validations (UK and US) show average absolute errors in that range when turnout models are calibrated correctly.
* Academic reviews (AAPOR, ESOMAR discussions) converge on the same finding: online ≈ probability, for toplines, after weighting.

But the fine print matters:
* The 2–4 pp figure applies to: - national aggregates - high-salience questions - experienced polling organizations
* It breaks down when: - turnout is mis-modeled (US midterms, referenda) - questions are cognitively demanding or low-salience - subgroups are small or politically disengaged

Bottom line (no sugar-coating):
The statement is defensible as an average, not a guarantee. When these polls fail, they don’t fail by 1–2 points—they can miss by 10+ points. The method is robust but brittle: solid under the right conditions, unreliable when stretched beyond them.