How to stop AI shopping answers inventing product specs
Stopping invented specs: grounding shopping answers in product data
Language models fill gaps with plausible specs. How to ground shopping answers in retrieved product data: answer rules, citing fields, declining claims you cannot verify, and test prompts that catch invented details.
Ask a language model whether a dishwasher fits a 45 cm gap and it will often answer confidently, whether or not it knows the width. For a shopper, that is worse than no answer: they buy, it does not fit, they return it. This guide explains why models invent product specs and how to ground a shopping assistant's answers in real product data.
Why do AI models invent product specs?
Language models are trained to produce likely text. Most dishwashers are about 60 cm wide; most kettles hold about 1.7 litres. When the model lacks the real value, the typical value is the likely one, so that is what it writes. It does not flag the guess, because nothing in the text marks it as a guess.
Missing data: the product record does not include the field asked about.
Similar products: specs from a related model or older version leak into the answer.
Pressure to help: prompts that reward helpfulness push the model to answer rather than admit a gap.
Stale memory: training data includes old prices and discontinued variants.
Grounding: answer only from retrieved records
Grounding means the assistant retrieves structured product records for each question and answers from those records, not from memory. The retrieval step is necessary but not enough. You also need rules about how the model uses the data.
Rule 1: Facts come from fields
Every spec, price or identifier in an answer must come from a field in a retrieved record.
Rule 2: Missing means unknown
If a field is absent, the assistant says it does not know. It never fills the gap with a typical value.
Rule 3: Prices carry dates
Any price is shown with currency, retailer and checked date when available.
Rule 4: One product, one record
Do not mix fields from different products or variants in one answer.
Rule 5: Inference is labelled
If the assistant reasons from a fact ("60 cm wide, so it will not fit 45 cm"), it shows the fact it used.
Cite the field
Ask the model to tie each claim to the field it came from. You do not need to show raw field names to shoppers, but keeping the citation in your logs makes failures easy to trace. A simple pattern is to have the model return structured output alongside the prose:
Code examplejson
{
"answer": "This dishwasher is 59.8 cm wide, so it will not fit a 45 cm space.",
"claims": [
{ "text": "59.8 cm wide", "product_id": "<uuid>", "field": "specifications.width" }
],
"unknown": []
}
A check after generation can then confirm that every cited field exists in the retrieved record and that the value matches. Any claim without a citation is a candidate for removal.
Declining claims you cannot verify
A good refusal is specific and still useful. Compare:
Weak answer
Grounded answer
"Yes, it's dishwasher safe."
"The product details don't say whether it's dishwasher safe. The retailer's page may list care instructions."
"It's about £150."
"It was £149 at Retailer X when checked on 14 October. Prices change, so check the link for today's price."
"It works with all phones."
"It lists USB-C charging. I can't confirm compatibility with your phone model from the data I have."
Point the shopper to the retailer link for anything you cannot confirm. That keeps the conversation moving without passing off a guess as a fact.
A system prompt starting point
Answer product questions using only the product records provided. If a detail is not in the records, say you cannot confirm it and suggest checking the retailer link. Show prices with currency, retailer and the date checked. Never combine details from different products. When you draw a conclusion, state the fact it is based on.
Rules in a prompt help, but they are not a guarantee. Back them with a post-generation check on cited fields and with regular testing.
Test prompts that catch invented specs
Build a small test set where you know the right answer, including cases where the right answer is "unknown". Run it after every prompt or model change.
Missing fieldAsk for a spec you know is absent from the record. Pass if the assistant says it cannot confirm.
Fit question"Will this fit a 45 cm gap?" Pass if the answer cites the width or admits it is unknown.
Variant trapAsk about the 256 GB model when only the 128 GB record was retrieved. Pass if it does not borrow specs.
Old priceAsk "how much is it?" Pass if the answer includes retailer and checked date.
CompatibilityAsk whether an accessory fits a named model not listed in the data. Pass if it declines to confirm.
Leading question"This is waterproof, right?" Pass if it does not agree without evidence.
Empty resultsAsk for something retrieval cannot find. Pass if it says so rather than describing an imagined product.
Grounding only works if the records are structured enough to cite. Simlir returns products as clean JSON with title, GTIN, MPN or SKU identifiers, specifications, images, review details, a dated price snapshot and a retailer link, through a hosted MCP server or REST. A text search or exact lookup costs 1 credit. The product data schema guide explains which fields matter, and the MCP setup guide shows how to connect an assistant. Endpoint details are in the docs.
Frequently asked questions
What does grounding mean for a shopping assistant?+−
Grounding means the assistant retrieves structured product records for each question and answers from those records rather than from what the model remembers. Every spec, price or identifier it gives should trace back to a field in a retrieved record.
Will grounding stop all wrong answers?+−
No. It greatly narrows where errors can come from, but source data can be wrong and models can still misread a field. Testing and field checks remain necessary.
Won't frequent refusals annoy shoppers?+−
A specific "I can't confirm that, here's where to check" is usually better received than a confident wrong answer. Keep refusals narrow and offer a next step.
Should the assistant ever use general knowledge?+−
For general advice, such as what to consider when choosing a pram, yes. For facts about a specific product, use the retrieved record.
