E-commerce
June 28, 2026
A chatbot claims "100% organic cotton" while the product sheet indicates a blend. It promises delivery within 24 hours while the policy announces 5 business days. It invents a promo code. This is an AI hallucination: a plausible but false response, delivered with confidence.
In e-commerce, this is not just a technical anecdote. Returns, chargebacks, regulatory complaints, reviews saying "the chatbot lies". RAG alone is not enough: runtime validation, confidence thresholds, and monitoring are required (Swept AI, preventing hallucinations 2026).
This guide #123 covers types, causes, technical and operational guardrails, monitoring, and Shopify improvement loops. Distinct from response quality (#116) and cleaning hub data (#122): here we focus on AI hallucination prevention. Complements chatbot limits (#124) and Shopify fallback (#279) when the API stops responding.
Summary
What is a chatbot hallucination in e-commerce?
An e-commerce chatbot hallucination is a factually incorrect response, not anchored in your data, presented with confidence.
Hallucination vs agent error
The agent forgets or uses an outdated macro. The bot generates content never written in the corpus. At scale, 1,000 conversations = a hundredfold risk. Perfect syntax = more subtle detection.
Why e-commerce is vulnerable
Changing catalog: daily stock, prices, promos
Precise policies: return, delivery, warranty
Critical specs: dimensions, ingredients, compatibility
Brand voice: the bot = official in the eyes of the customer
Concrete example
"Is this serum suitable for oily skin?" Bot: "100% oil-free formulation dermatologically tested" without a PDP source. Product contains argan oil. Return + 1-star review.
E-commerce bot = retrieval-first, not creative writer. See response quality (#116).
Why are hallucinations dangerous?
Bot hallucination risks impact revenue, reputation, and compliance.
Business consequences
Returns: product "not as promised by the bot"
Chargebacks: customer quotes bot delivery promise
Negative reviews: "the chatbot lies"
Agent rework: fixing a bot error costs more
Legal: unauthorized health claim
Reverse trust effect
Well-spoken and polite bot: the customer is more likely to believe a hallucination than a hesitant human response. An invented return policy can cost 10× in undue returns + CSAT + agent time.
What types of hallucinations are most commonly encountered?
Six types of hallucinations dominate on Shopify DTC.
1. Invented product specs
Size, material, compatibility, missing PDP ingredient. "Compatible with iPhone 15" without checking the catalog.
2. Invented policies
"Free returns for 60 days" vs 30 actual days. "Free shipping from €30" vs €50.
3. False inventory and promotions
"Yes, size M in stock" when the variant is OOS. Invented coupon code or expired promo announced as active.
4. Incorrect delivery time
Domestic vs International confusion. Carrier delays ignored.
5. Health or usage advice
Therapeutic claim generated for cosmetics or supplements. Critical regulatory area.
6. Product recommendation
Incompatible cross-sell or discontinued SKU. Inferred size guide instead of chunk size chart.
What technical and operational causes provoke them?
Understanding the causes of hallucinations guides the right fixes.
Technical causes
Free generation without RAG: LLM fills the gap by inventing
Poor RAG chunk: size issue, delivered chunk retrieved
Weak prompt: no instructions to "cite sources" or "say I don't know"
Missing confidence threshold: 40% response sent as 95%
High temperature: LLM creativity > 0.2 on facts
Operational causes
Stale corpus (2023 hub vs 2026 PDP), contradictions between hub and policy, too many day-1 intents without validation, multilingual gap in Markets, fine-tuning on obsolete data. RAG grounding typically reduces hallucinations by 60 to 75% (IrisAgent, reducing hallucinations 2026).
See cleaning terms hub (#122), train Shopify bot, prioritize automation (#120).
Which technical safeguards should be deployed as a priority?
Four stacked technical guardrails, not a single lever.
1. Strict RAG + runtime validation
The bot responds from approved chunks. The validation layer checks price, policy, and stock against an authoritative database before sending it to the customer. Unsourced claim = block or escalation.
2. Structured data first
Stock, price, and allergens via Shopify metafields API. No text generation for structured data.
3. Confidence scoring
85% standard threshold, 95% regulated. Below threshold: handoff or "I don't know". Mandatory source citation for factual claims.
4. Blocklist and human-in-the-loop
Forbidden words: cures, 100% guaranteed, always in stock. Launch phase: agent approves uncertain responses. LLM temperature 0 to 0.2 for factual support.
eGrow reminds us: real-time catalog grounding + guardrails = 2026 DTC strategic necessity (eGrow, e-commerce hallucinations 2026).
How to build an anti-hallucination corpus?
The corpus is the foundation: stale corpus = hallucination factory.
Sources of truth (priority order)
Live Shopify product pages
Up-to-date site policy pages
Versioned terms and conditions hub
Approved Gorgias macros
Centralized marketing promo calendar
Weekly maintenance
Sync stock and price catalog. Deduplicate to one answer per question. Version date on each chunk. Remove expired promos. Scan for contradictions hub vs. product page. Conflict rule: product page beats generic hub.
Chunking
One chunk = one atomic piece of information. "30-day return policy" separated from "free returns status." Designated corpus owner (bot admin), not "installed once by IT."
How to detect and monitor hallucinations?
Without monitoring, hallucinations scale before being corrected.
Automated signals
Unmatched rate: question without relevant chunk
Low confidence log: responses below threshold
CSAT bot < 4: post-response dissatisfaction
"That's wrong" escalation: customer contradicts bot
Bot reason feedback: Shopify tag if applicable
Monthly QA audit
50 sample conversations: each factual claim vs source. Target accuracy 97 %+. "Report incorrect response" button post-chat. Compare bot vs agent responses on identical questions.
Regression test suite
100 gold standard questions. Run after each corpus update. Fail if accuracy < 95 %. Add each real incident to the suite.
How to manage fallback, handoff, and correction?
Better to say "I don't know" than to make a confident invention.
Phrase fallback type
"I am not sure of the exact answer. I'm putting you in touch with an advisor who will check [subject] with the most up-to-date information."
Internal culture
Fallback = prevention success, not bot failure. Deflection KPI must not punish justified handoff. Precision > deflection.
Enriched handoff
Agent receives question, chunks retrieved, confidence score, product context. Resolves without asking the customer again.
24-hour correction loop
Incident detected
Correct corpus or metafield
Add regression test
Root cause: stale, bad chunk, not RAG?
Shopify API down: "check product page" don't guess stock. See reduce AI tickets.
Which areas require maximum caution?
Three high-caution zones: maximum cost if hallucination occurs.
Regulated products
Corpus locked, 95% threshold, never generative health advice. Handoff for medical keywords.
Prices, promos, stock
Live price API. Whitelist promo codes marketing sync. Real-time Inventory API. "Variable stock, check PDP" message if API lag occurs.
Large catalog and subscriptions
5,000+ SKUs: mandatory product context. Billing Recharge via API, no generative next billing date. Sustainability certification claims: approved legal chunks only.
How do you organize continuous improvement?
Continuous improvement transforms incidents into permanent barriers.
Weekly ritual
Export top 10 unmatched + low confidence
Review CSAT bot < 4
Update corpus or intents
Re-run 100-question test suite
Incident post-mortem
Template: question, bot answer, correct answer, root cause, fix, test added, date. "Incorrect information" feedback = priority review.
Seasonality
Pre-BFCM: 100% audit of corpus promos on D-7. Post-event: remove promo chunks within 48 hours. New SKU week 1: temporary corpus overlay until PDP is stable.
How does Qstomy prevent hallucinations?
Qstomy stacks anti-hallucination guardrails for Shopify stores.
Features
Locked corpus RAG: syncs PDP + policies
Source citation: chunk link per response
Confidence gate: threshold per intent
Structured API: stock, price, metafields
Blocklist claims: health, promo, warranty
Regression export: gold questions suite
Weekly report: top hallucination risks
DTC Encrypted Scenario
Outdoor brand with 500 conv/month. Month 1 without guardrails: 8 incidents with invented specs. Month 2 Qstomy RAG + 90% confidence + citation: 0 incidents in Month 3, 98% audit accuracy, 4.4 bot CSAT, 48% deflection maintained.
14-Day Setup
Corpus contradictions audit
Import structured PDP + policies
Confidence thresholds per intent
50 test gold questions
Active citation + blocklist
Weekly calendar review ritual
Explore AI support, Shopify, request a demo.
Which operational playbooks should be launched this week?
Playbook 1: audit 20 bot conversations
Verify each factual claim vs PDP source or policy. Accuracy score. List discrepancies. Delay 3 hours.
Playbook 2: scan corpus contradictions
Compare terms hub vs policy pages vs macros. Flag return/delivery/promo inconsistencies. Correct before go-live. Delay 2 hours.
Playbook 3: 50 gold test questions
Draft validated responses for WISMO, return, stock, promo, specs of top 10 SKUs. Run after each corpus update. Delay 4 hours.
Playbook 4: tightening confidence + fallback
Threshold 85% global, 95% regulated. Fallback phrase configured. Active claims blocklist. Delay 1 ops day.
Playbook 5: weekly bot quality ritual
Support + bot admin 30 min: unmatched, low confidence, customer reports. 1 corpus fix minimum. Recurring.
Useful links
A bot that says "I don't know" protects your brand. A bot that invents destroys it. Test 10 tricky questions on return policy and stock: a single invention = block the go-live.
Enzo
June 28, 2026
