Read a result the way an editor would
A percentage on a chart is the end of a chain: a question, some models, some calls, some replies, some of them answers. Check the chain before you quote the number.
1. What is it out of?
"14 of 20 readable replies" and "14 of 26 calls" are different findings. Look at the refused, unreadable and cut-off counts beside every share: a model that declines half the time is telling you something the share hides.
2. How wide is the range?
70% from 20 answers has a range of 48% to 85%. The same 70% from 200 answers is 63% to 76%. Quote the range, or the sample size, and never a bare percentage.
3. Is it one wording or all of them?
A combined figure can hide two wordings pulling in opposite directions. Check each wording before you trust the total.
4. Did the model change, or the poll?
Compare runs of the same version. A version boundary on the chart means the question or the panel changed, and the two sides are not the same measurement.
5. Did the same model answer?
Check the model that answered on both runs. A change of model name, or of the host that served it, is a reason the answer moved that has nothing to do with the world.
6. A move is a question, not an answer
Ranges that do not overlap are worth a look; they are not a full test, especially across many comparisons. Read the replies on both sides of the change. The chart tells you something moved; the replies may tell you why.
7. Consistent is not correct
Twenty identical answers can be twenty identical mistakes. Check the facts against a record.