When AI Chatbots Give Sizing and Fit Advice, Accuracy Gets Tested
Store chatbots now answer fit questions with total confidence. What happens to returns and trust when the confident answer is wrong, and how stores can audit it.
Ask a store chatbot whether a jacket runs small and it will answer instantly, fluently, and sometimes wrongly. Sizing advice is one of the most consequential things a chatbot can give, because a wrong answer becomes a return, and the shopper remembers who advised them.
Fit advice is a useful test case for AI accuracy in commerce: it is specific, checkable against the size chart, and directly tied to money. This guide covers where chatbot fit advice goes wrong and how a store can audit it before shoppers do.
Confident answers built on thin data
A chatbot answering fit questions is usually working from product descriptions, reviews, and general knowledge, blended fluently. Where the size chart is not in its data, it improvises from category norms, which may be the opposite of how this brand actually cuts its garments.
The failure is invisible in the transcript. The answer reads like expertise, cites no source, and gives the shopper no way to check it. Fluency is doing the persuading, and fluency is not evidence.
The return is the receipt
Bad fit advice converts directly into returns with a reason attached: item did not fit as described. Tracking return reasons for orders that touched the chatbot gives a store the first honest measure of its advice quality, and it is a measure most stores never take.
Compare advised orders against the size charts verdict. Where the bot recommended a size the chart would not have, each incident is a training, data, or guardrail fix with a price tag already attached.
Give the bot the chart, and a way to say unsure
The fixes are practical. Put the actual size chart and fit notes into the data the assistant can reach, instruct it to answer from them, and let it say when a product has no reliable fit information instead of filling the gap with a guess.
Route ambiguity to the chart itself: a good answer quotes the measurement, names the source, and links the guide, so the shopper can verify in one step. Verifiable answers survive being wrong occasionally. Unverifiable confidence does not.
Audit on a schedule, with mystery shoppers
Sample real fit questions monthly, run them against the assistant, and score the answers against the charts. Include edge cases: tall and petite ranges, between sizes, and products whose fit notes say they run unusual. Edge cases are where improvisation lives.
Retest after any model, data, or catalogue change. Chatbot advice drifts the same way every automated system drifts, silently, and the audit is the only instrument that notices before the returns data announces it at scale.
Related reading
If this topic is useful, these earlier guides go deeper on connected questions: When AI Search Answers Send Shoppers to Out of Stock Products, AI Written Product Descriptions, Accuracy, and the Returns They Cause, and When AI Recommendations Get Prices and Stock Wrong.
Frequently asked questions
Why is sizing advice a high risk chatbot task?
Because it is specific, acted on immediately, and expensive when wrong: the shopper buys the advised size, it does not fit, and the store pays for the return while the shopper loses trust in the advice channel itself.
How can a store measure chatbot fit accuracy?
By joining chatbot touched orders to return reasons, and by sampling fit questions and scoring answers against the actual size charts. Return data shows the cost, sampling shows the causes.
Should chatbots refuse fit questions they cannot answer?
They should answer from the size chart when it is available and say plainly when it is not, pointing to the chart or a human. A stated uncertainty preserves trust. An invented measurement spends it.
How often should chatbot advice be audited?
Monthly sampling as a rhythm, plus a retest after any model update, data change, or major catalogue addition. Advice quality drifts silently, so the audit has to run whether or not a problem is suspected.