Should an AI assistant make your survey percentages add up to 100%?
If a customer survey lets people tick several reasons for returning an order, don't ask an AI assistant to ‘fix’ the percentages just because they add up to more than 100%. One customer can report both a poor fit and a late delivery. Those problems overlap. Survicate's existing guide explains that each option's percentage uses the people who answered that question as its base. Because a person can choose several options, the percentages can total more than 100%. That's documented survey behavior, not evidence of a faulty calculation. Rescaling the figures to make a tidy pie chart would answer a different question: the share of selections, rather than the share of people reporting each problem. A shop deciding how many customers need better sizing information could underestimate the problem if that distinction disappears. I'd give the assistant the original question, answer options and respondent count, then ask for a bar chart labelled ‘Percentage of respondents; multiple answers allowed’. Keep the counts beside it. If the export only contains percentages and no base, ask for the missing information rather than having the assistant invent it. Before sharing an AI-written survey report, can you tell whether each percentage describes people or selections?
Comments
Before assigning the sizing problem to the person who writes product pages, I'd read the comments behind that answer. ‘Wrong size’ could mean the chart was confusing, or that the warehouse sent a different size from the one ordered. Those need different fixes. Survicate's guide lets you open an individual respondent's answers and comments together. I'd use that to check what customers meant before asking an AI assistant to turn the category totals into a team action list. A tidy summary shouldn't leave the catalog team rewriting a page when the problem is in picking orders.
An answer option can disappear from the results view altogether. Survicate's guide says options nobody selected aren't shown in the Analyze tab as zero-response rows. So give the AI assistant the original answer list as well as the results. Otherwise it can't tell whether ‘damaged on arrival’ was offered and never chosen, or was never asked about. Those support different decisions about packaging. Checking individual responses helps explain the answers you received; keeping the questionnaire helps you see what the report leaves out. That's a documented display rule, not an AI failure I've observed.