Start free trial
Glossary

Hallucination detection: definition and examples

Catching AI engines when they fabricate product details that the PDP never claimed — price, features, variants, availability.

Last updated Q1 2026

Hallucination detection identifies when an AI engine fabricates product-level details — price, features, variants, or availability — that the source PDP does not support.

In detail

AI engines summarize, condense, and sometimes invent. Hallucination detection is the practice of reading each AI answer against the canonical product data and flagging mismatches. The check runs at two layers: structured fields that can be diffed mechanically — price, SKU, variant labels, stock — and free-text claims that need careful pattern matching against the PDP description, metafields, and supporting marketing copy.

Common hallucination patterns as of Q1 2026 include stale pricing cached from earlier indexing, feature invention (claims the PDP never made), variant confusion (mixing specifications across sizes), and phantom availability. The patterns vary by engine; ChatGPT Shopping tends toward feature invention while Perplexity tends toward stale pricing in eCommerce Insights's observation to date.

Why it matters

A hallucinated price or feature is a conversion problem before it is a brand problem. A buyer who arrives at the PDP expecting the AI-quoted price and finds a different one often leaves. A buyer expecting a feature the product does not have will refund and post a negative review.

Hallucinations also deplete AI visibility over time. Engines quietly down-weight sources that produce correction-worthy answers, so letting hallucinations persist makes the underlying SKU less likely to be cited later. Fixing hallucinations is part of AI visibility maintenance, not a separate compliance chore.

Example

For example: a ceramic-mug brand sees ChatGPT describe the 12oz hand-thrown mug as "dishwasher and microwave safe" when the PDP explicitly says hand-wash only. The claim is a free-text hallucination — the engine blended the mug into a generic ceramic-mug answer. eCommerce Insights flags the mismatch, the team updates the metafield that expresses care instructions in structured form, and the next two weekly tracking runs show the hallucinated claim gone from ChatGPT answers.

Related terms

Ask AI about hallucination detection

Have your favorite AI engine summarize this for your specific use case.

Frequently asked questions

What kinds of product hallucinations are most common?
As of Q1 2026 the common patterns are stale price quotes (the engine cites a last-known price that no longer matches the PDP), feature invention (a claim not on the page), variant confusion (mixing specs between a small and a large), and phantom availability (saying in stock when the Shopify admin shows out of stock). Detection is easier for structured fields, harder for free-text feature claims.
How does eCommerce Insights detect hallucinations?
eCommerce Insights compares each AI answer against the canonical PDP pulled from the Shopify admin. Structured fields — price, variants, stock — are diffed directly. Free-text claims are checked against the description, metafields, and approved marketing copy. Mismatches are flagged with a severity label so teams triage price hallucinations before wording ones.
What can a Shopify brand do about AI hallucinations?
Three fixes help most, in order: keep Product JSON-LD fresh so engines have the current source of truth; expand feature bullets and metafields so the right information is cheaper for retrieval to reach than the wrong information; and request corrections through engine feedback channels where available. eCommerce Insights automates the first two and tracks whether hallucinations recur.

Related guides

Catch AI hallucinations before your shoppers do. Start a free trial or read Wikipedia on AI hallucinations.