E-commerce
July 1, 2026
A chatbot may recommend an unsuitable product: wrong size, poor compatibility, incorrect use, ignored budget, or too generic advice. The customer may then buy the wrong item or lose trust in the brand.
The response must not be limited to "sorry". The bot must acknowledge the discrepancy, understand the real need, propose a correction, and escalate if the recommendation caused a post-purchase issue.
This guide explains how to handle poor AI recommendations with transparency, method, and a customer-centric approach.
Summary
Why is a bad recommendation sensitive?
A recommendation is perceived as advice. If it is bad, the customer does not just think the chatbot was wrong; they may think the brand does not understand their need.
The risk is even higher for technical products, sizes, allergies, compatibility, or expensive purchases. An error can lead to a return, a complaint, or a negative review.
The chatbot must know how to correct its advice with as much care as it gives it.
Convert over 2,000 customers on average per month with Qstomy.
The world’s 1st Shopify AI dedicated to customer conversion
Empowering 200+ e-commerce merchants
What types of errors can be recognized?
Frequent errors include incompatible recommendations, over-budget products, unsuitable sizes, advice that ignores a client constraint, unavailable products, or overly general responses.
The bot must also recognize when the client reports an error: "it's not compatible," "you recommended this to me," "that doesn't match," or "I told you that...".
How to respond when the customer reports the error?
The response must acknowledge the problem without arguing. The bot can say: "You are right to point that out. The recommendation does not seem to take into account your constraint. I will start over from your exact need."
Next, it must ask the right missing questions and propose a correction. This attitude gives the customer the feeling that the brand takes the error seriously.
How do I correct the recommendation?
The bot must start over from the essential criteria: usage, size, compatibility, budget, preference, availability, and personal constraint. It must not simply propose a second product at random.
The correction must explain why the new choice is more suitable. For example: "This model is compatible with your device, unlike the previous one, and stays within your budget."
What if the customer has already purchased?
If the customer purchased because of a bad recommendation, the topic becomes support. The bot must check the order, the return policy, and the exchange possibilities.
It must not promise an exceptional refund if it does not have the authority to do so. On the other hand, it must pass on the context: recommendation given, product purchased, problem observed, and customer request.
Which flow to follow?
The flow must correct the situation without holding the customer responsible.
Acknowledge the report and rephrase the need.
Identify the overlooked constraint: size, usage, compatibility, budget, or availability.
Verify which products are actually suitable.
Explain the new recommendation with a clear reason.
Transfer the case if the error has already resulted in a purchase or a dispute.
Which messages should be used?
For compatibility: "Thank you for bringing this to our attention. This product does not seem to match your device. I will look for a compatible option for your model."
For a size: "The previous recommendation does not take your measurements sufficiently into account. Let's start over with your usual size and fit preferences."
For an already made purchase: "I will forward the context to our team so they can check the best solution based on your order and return policy."
How to learn from these mistakes?
Each incorrect recommendation must be classified: missing product data, missing compatibility rule, question not asked, product unavailable, or misunderstanding of the need.
This classification transforms a disappointing conversation into a system improvement. It allows for the correction of product sheets, recommendation rules, or chatbot questions.
It also helps the product and support teams to identify weak areas in the catalog. If several customers report the same error, the problem probably does not stem from an isolated conversation.
Which KPIs should be monitored?
Track disputed recommendations, returns linked to poor advice, the most affected product categories, accepted corrections, and post-purchase escalations.
If a category concentrates errors, the problem often stems from insufficient product information or weak qualification questions.
Which mistakes should be avoided?
Avoid defending the bot's initial response, suggesting the same product again, denying the customer's constraint, or covering up the mistake.
A customer more easily forgives an acknowledged mistake than a response that seems to try to be right.
_
How can Qstomy help?
Qstomy can detect bad recommendation signals, resume need qualification, and hand off post-purchase cases with full context.
The bot can also flag categories where recommendations need improvement.
Explore the AI sales agent, AI support or request a demo.
IRECO-BOT Checklist (12 steps)
Validate IRECO-SUP #439 + IRECO-MAP /pages/reco-feedback
Export IRECO-MAP + exclude purchased rules JSON
Configure 12 intents bot_ireco_* section 3
Implement flow IB-1 to IB-8 + correction engine IB-5
Activate guardrails no defend + catalog only + max 2 SKU
Route ireco vs #439 product_q
Placements thumbs down T1 + widget footer T3 + feedback page T4
Structured feedback menu reason codes
Staging tests 8 scenarios: bad bot, widget, owned, correction accept, opt-out, allergy handoff, defend zero, product q route
IB-8 learning export weekly merch review calendar
Monthly ireco_bot KPI dashboard + defend audit
A/B ack copy + correction format 4 weeks
In brief
#440 = bot feedback reco tier 1, #439 agents P1 gesture
Acknowledge never defend: trust rule #1
IRECO-BOT: feedback → verify → correct grounded → log
Max 2 alternatives: no carousel post complaint
KPI ireco_correction_accept_rate: target > 55%
FAQ
Difference #439?
#439 customer service agents merch flag gesture opt-out. #440 bot ack feedback correction learning self-service.
Bot defends its suggestion?
No. Rule 1 acknowledge never defend. Always IB-6 ack first.
How many alternatives?
Max 2 SKU grounded IB-5 exclude owned OOS complained.
Allergy conflict?
bot_ireco_handoff_439 immediate P1. No alternative without metafield verify.
How does it learn?
IB-8 ireco_feedback_log weekly export merch + bot corpus review.
Going further
This week: activate thumbs down T1 on bot suggestions, configure feedback reason codes menu, test IB-5 correction exclude owned, schedule weekly IB-8 learning log review with merch.
Share this guide #440 with product and support: a bot that says "you are right, here are two suitable alternatives" is worth ten algorithm defenses, a product hallucination is worth a lost customer and a Twitter screenshot.
Product question without reco complaint?
bot_ireco_route_product_q to product questions bot, not IRECO-BOT flow.
Enzo
July 1, 2026
