A Shopper Asked Google About Your Product. The Answer Was Wrong.

In this article: the part of Interpretability Engine Optimization (IEO) that takes the guesswork out of what you hand a machine about your product. It's the validation step, and it works on one rule: verified, never invented. Every statement about a product has to trace back to something the business published. Where the page is silent, nothing gets filled in.

Here is what happens without it.

Imagine you sell a rain jacket. A good one, for what it is: light, packable, made for showers on a trail day. Your product page says so in plain words: water-resistant shell for light showers. Seams are not taped.

A shopper asks Google's AI Mode whether the jacket is waterproof. The answer comes back: "Yes, it's waterproof."

She buys it. Two weeks later she's standing in real rain, soaked, and the jacket comes back to you with a one-star review saying it doesn't do what it promised. You never promised that. A machine did, in your name.

Nobody called you to check, and it won't show up in your reports. Your page ranks, your feed is approved, your ad is running. The only trace is a return and a bad review.

Why a machine gets a clear page wrong

Your page was clear to a person, and the machine can read it too. Reading isn't the problem. "Water-resistant" sits in a sentence next to "shell" and "showers", and nothing in the product's data spells out the difference between water-resistant and waterproof.

So when the question arrives, the system leans on what rain jackets usually are. Rain jackets are usually waterproof, so yours must be too.

This isn't a glitch nobody has noticed. OpenAI says of its own shopping research that it can get product details wrong, price and availability included, and tells shoppers to check the merchant's site. And in September 2026 Google started asking retailers to send answers to product questions through Merchant Center, because its AI needs more than a name and a price to describe what you sell.

Product feeds are twenty years old. The machines reading them aren't.

The technical name: unsupported inference

In IEO, this failure has a name: unsupported inference. A statement shows up in the answer that the business never published. It was guessed from what similar products usually are.

The Methodology calls it the most expensive kind of failure, because it looks like success. The answer is fluent and confident, and nothing flags it until a customer finds out.

The validation step exists for this. On the jacket, it means three moves:

1. Resolve what the page actually claims. "Water-resistant shell for light showers. Seams are not taped" becomes explicit facts: water-resistant, yes. Seams, not taped. Waterproof rating, none stated.
2. Answer the question before anyone asks it, in the store's own words.
3. Leave out what the page doesn't say. No waterproof rating on the page means none gets written.

Where it lands: the answers Google now asks for

Google calls them conversational attributes: questions and answers about each product, how products relate to each other, and variant options. The questions and answers go in an attribute called question_and_answer, which has two parts: the question and the answer. In a test with lululemon, details supplied by the brand showed up in AI Mode's recommendations.

Here is the jacket, left to the machine and prepared from its own page:

Nothing in the last column is new. It's the store's own page, sorted into answers a machine can use.

What to do this week

1. Pull the five questions your customers ask most. Your support inbox and your reviews already know them.
2. Check that each answer is on the product page, in plain words. If it isn't, fix the page first. No feed can carry an answer the store never wrote.
3. Write each one as a question and an answer, in the page's own wording, and add it to the product in Merchant Center under question_and_answer. Google takes up to 30 pairs per product.
4. Start with the answer that would hurt most if a machine got it wrong. Waterproofing, allergens, compatibility, sizing.
5. Don't copy your other fields into it. Google asks you not to duplicate what you already send in the title, description, product details or highlights. Use the questions for what those fields leave open.

Google's AI will answer questions about your products whether you prepare them or not. The answers can come from your page or from a guess.

Practical example

A sunscreen page says: "Mineral formula, SPF 30."

A shopper asks an AI assistant: "Is it reef-safe?"

Before you read on, decide what can be stated and what can't.

Verified: mineral formula, SPF 30. Both are on the page.

Not on the page: reef-safe. Mineral sunscreens are often sold as reef-safe, and that's exactly the guess a machine makes. So that field stays empty until the brand confirms it.

That empty field has a name too: an honest empty. It's the rule from sections 4.2 (Verified, never invented) and 5.4 (Unsupported inference) of the IEO Methodology, applied to one product.

Check what's explicit, what's missing and what contradicts itself on one of your products → Run the free audit

To go deeper → The IEO Methodology, sections 4.2 and 5.4 (doi.org/10.5281/zenodo.22416784) · nsolvia.com/ieo

Sources: Google, "Retailers: Get ready for AI-driven holiday shopping," Sept 16, 2026 · Google Merchant Center Help, "Question and answer [question_and_answer]" · OpenAI, "Introducing shopping research in ChatGPT," Nov 2025.

← The Platforms Now Describe Your Products For You · ⌂ IEO · Next: Your About Page Says Who You Are. No Machine Can Use It. →

Series: Interpretability Engine Optimization · Knowledge Domain: Interpretability (IEO)

Interpretability Engine Optimization (IEO)
The UNDERSTAND Layer
nsolvia.com/ieo