Write a question you can still interpret in a month
Most surprising poll results come from the question, not the model. A few habits make the difference.
Ask about the present, or ask for an estimate
"What will next year's amount be?" is a question many models decline: asked about a figure not yet set, they tend to say it has not been announced. If you want a number about the future, ask for one plainly: "What is your best estimate of...?" Decide in advance whether a refusal is the finding or a failure.
Name the thing exactly
"The 2027 dividend" can mean the payment made in 2027 or the one for a 2027 fiscal year, and a careful model may say so instead of answering. Name the payment, the year and the official source you mean.
Ask one thing at a time
Trust, competence and agreement are three questions. Asked together, a single answer cannot tell you which one moved.
Make the options distinct, and allow "not enough information"
A model forced to choose between overlapping options picks one for reasons you cannot see. A "not enough information" option turns a hidden refusal into a counted one.
Change one thing between wordings
Hold the facts and the options steady and change only the phrasing, so a difference between wordings means the phrasing mattered.
Fit the format to the question
Choice for named alternatives, a scale for degree, a number for a quantity with a unit, and a short written answer when you need to read the reasoning. Format changes answers: in one study the best systems lost 11 to 13% (the paper's own measure) when the same news questions were asked for a free answer instead of multiple choice (Suzgun et al.).
Pilot, then keep the boundary
Read the first replies, fix what confused the models, and let the fix start a new version. A better question starts a better series; it does not repair the old one.