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

=== Assistant: Short answer: it works reasonably well for broad opinion levels, and poorly for fine-grained inference. ===
Short answer: it works reasonably well for broad opinion levels, and poorly for fine-grained inference.

Tell it like it is:
* Good at: - National-level percentages (± a few points). - Tracking trends over time. - Salient, well-understood issues with stable opinions.
* Weak at: - Small subgroups (youth, minorities, niche political segments). - Low-salience or cognitively demanding questions. - Anything requiring behavioral accuracy rather than stated attitudes.
* Core problem: Weighting fixes known demographics, not unknown biases. If politically interested, media-savvy, or survey-prone people answer differently—and they do—weighting does not fully correct that.
* Empirical bottom line: YouGov-style panel surveys usually land within ~2–4 percentage points of high-quality probability polls on headline questions, but can miss badly when: - turnout assumptions matter, - opinions are socially sensitive, or - the issue is new or complex.

Conclusion:
Weighted online panels are good enough for population mood, not good enough for precise social diagnosis. Treat them as informed estimates, not measurements in the natural-science sense.