An AI assistant names your product and links to your store. At first glance, that looks like a successful recommendation.
But is it the right size? Is the quoted price for that variant? Does the delivery promise apply to the customer's country? Has the assistant confused a single item with a pack of six?
To check whether AI can read your product catalogue correctly, test an exact product and variant against a dated reference record. Verify identity, specifications, price conditions, availability and delivery separately, then inspect the source supporting each answer.
A product mention is not a catalogue accuracy test. The useful question is whether the customer receives enough correct information to choose the intended item without being given a false promise.
This guide shows how to run that test across ChatGPT, Gemini and Claude, starting with a small sample rather than a full catalogue overhaul.
What does “reading a catalogue” actually mean?
There are several different tasks hidden inside that phrase.
| Task | What you are testing | What success does not prove |
|---|---|---|
| Read a supplied product URL | Extract information from that page | Independent discovery of the product |
| Analyse an uploaded PDF | Interpret the document supplied in chat | Access to the current public website |
| Find a product through search | Identify the right item and seller | Correct selection of every variant |
| Use a platform product feed | Process the submitted catalogue format | The same behaviour in every assistant |
| Answer a shopping question | Apply product facts to customer requirements | Stock reservation or guaranteed compatibility |
Keep these tests separate. An assistant reading your uploaded brochure does not establish that buyers can discover the same information through public search.
Similarly, an accepted feed does not prove that a generated answer preserves every condition.
The most important mistake: right product family, wrong purchasable item
Consider a fictional product family called the Harbor Chair. It has two options:
| Attribute | Harbor Chair A | Harbor Chair B |
|---|---|---|
| SKU | HBR-A-BLK | HBR-B-GRY |
| Colour | Black | Grey |
| Material | Polypropylene | Upholstered fabric |
| Intended use | Indoor and covered outdoor use | Indoor use only |
| Price | A different amount for each option | A different amount for each option |
| Availability | Must be checked for the selected option | Must be checked for the selected option |
An answer that combines the black option's material with the grey option's image and a family-level starting price has not identified a coherent item.
This is why the audit should begin with a product identity, not just a product name.
Step 1: build a reference record for each test item
Create a dated record approved by the catalogue owner.
| Field | What to establish |
|---|---|
| Product identity | Product name, brand and model |
| Variant identity | Exact SKU or other legitimate identifier, size, colour and configuration |
| Seller and market | Which store, country and currency the test concerns |
| Transaction | Purchase, rental, subscription or quotation |
| Unit | One item, pack, metre, month, day or another defined unit |
| Price conditions | Taxes, deposits, sale period, selected configuration and exclusions |
| Availability | Status for this variant at the recorded time |
| Specifications | Dimensions, material and relevant limits |
| Fulfilment | Delivery or collection rules actually published |
| Evidence | Current product URL and relevant policy sources |
Do not invent a GTIN to fill a spreadsheet or force a custom product into a standard retail model. Record identifiers that genuinely exist and apply to the item.
Separate internal truth from publicly available evidence. If a compatibility rule is known only to your technical team, decide whether buyers need a public explanation before treating its absence from an AI answer as a failure.
Step 2: choose a useful sample
Start with a sample that exposes different types of ambiguity.
I would include a simple product, a variable product, a sale item, a temporarily unavailable variant and a product with an important restriction. For larger catalogues, also include a bundle and an item whose price depends on configuration.
Add a commercially important product even if it is technically simple. An error on a frequently enquired-about item may matter more than a difficult edge case that nobody buys.
This is a risk-based sample, not a representative estimate of the entire catalogue. Publish or report that limitation rather than extrapolating a handful of tests to thousands of products.
Step 3: test exact-page extraction
Supply the exact product URL and selected variant where possible.
Retrieve this public product URL: [URL].
The variant being checked is [SKU / SIZE / COLOUR / CONFIGURATION].
Using this page and any explicitly necessary linked policy pages,
report:
- Product and variant identity
- Seller, market and currency
- Purchase/rental/quote model and pricing unit
- Published price and important conditions
- Availability at the time of retrieval, if established
- Dimensions, material and stated use restrictions
Support each fact with a URL and a short relevant excerpt.
Do not substitute a different variant or assume missing details.
If the selected variant is not established, say so.The prompt is a diagnostic tool. It is not evidence of what every shopper naturally asks.
If the assistant gives family-level information, check whether the selected option is represented in the URL, page content and retrieved evidence. The answer may be unable to establish the variant rather than choosing the wrong one deliberately.
Step 4: test independent product discovery
In a new conversation, omit the product URL.
Search for [EXACT PRODUCT / MODEL] sold or offered by [BUSINESS]
for customers in [COUNTRY].
I need [EXACT VARIANT OR CONFIGURATION].
Identify the matching product URL and explain which published
facts establish the match.
Keep the seller, variant, currency and pricing unit explicit.
If no exact match is found, say that rather than substituting
a related product without labelling the difference.Record discovery success separately from factual accuracy. A correct description of a different seller's offer does not establish accurate discovery of your store's item.
Also note whether the answer found a category page, a product page, a PDF or an external listing. Different sources can contain different levels of detail.
Step 5: test a realistic buying question
Now ask whether the item fits a customer's need.
I am considering [EXACT PRODUCT AND VARIANT] from [BUSINESS]
for [USE CASE] in [MARKET]. My requirements are [REQUIREMENTS].
Using current public evidence, classify each requirement as:
- Supported
- Contradicted
- Not established
Preserve product restrictions and distinguish specifications
from assumptions. Do not guarantee stock, delivery dates,
safety or compatibility beyond the published evidence.
List what I should confirm with the seller before ordering.A useful response may say that dimensions match but outdoor suitability is not established. That is more informative than a confident recommendation built on missing information.
For safety-critical products, keep suitability decisions with qualified people and authoritative product documentation. The audit should expose missing evidence, not replace required technical assessment.
Which facts should you score separately?
Do not hide several commercial mistakes inside one “product understood” tick.
| Fact | Example of a meaningful failure |
|---|---|
| Identity | Different model or seller |
| Variant | Price from one size attached to another |
| Transaction | Rental described as purchase |
| Unit | Pack price described as price per item |
| Currency | Correct number with the wrong currency |
| Specifications | Shipping-box measurements described as product dimensions |
| Availability | Product-family stock applied to an unavailable option |
| Fulfilment | Delivery eligibility described as guaranteed next-day delivery |
| Policy | General return terms applied to a documented custom-order exception |
Label each response correct, partial, incorrect, not established or omitted. Record supporting evidence separately and flag errors that could cause a wrong order or a false expectation.
If a tool fails to retrieve the page, record that failure. Do not remove failed runs and report accuracy only among the successful ones without disclosure.
What usually deserves investigation when details are wrong?
The visible mistake does not establish its cause, but these checks can narrow the problem.
The product page mixes family and variant details
A title may refer to the family while a price belongs only to the default option. Make the selected item's facts explicit and preserve the relationship between variants.
Google documents product-variant markup using ProductGroup and Product entities, including variant identifiers and URLs that select the relevant option. That is Google-specific guidance, not a guarantee about every assistant.[1]
The retrieved content lacks the selected option
Some sites load price, stock or specifications only after interaction. Compare the visible selected variant with the initial HTML and the content exposed by the relevant retrieval path.
Do not assume that all AI tools execute—or cannot execute—the same JavaScript. Follow the website access troubleshooting process when retrieval is incomplete.
Old documents or external listings contradict the page
A current product page can coexist with a dated PDF or outdated reseller listing. Inspect the exact source behind the wrong claim.
Use the outdated business information guide for the freshness investigation, then correct each source according to ownership.
Units are ambiguous
Separate product dimensions from package dimensions, net quantity from shipping weight and monthly payments from total purchase cost.
For rental inventory, distinguish price per day from price per event. If no public price exists, a quote-only answer can be correct.
The assistant goes beyond accurate evidence
If the page clearly identifies the item and the answer still merges variants, preserve that output as an interpretation error. Better documentation may be worth testing, but a website defect is not established merely because a generated answer is wrong.
What should a clearer product page contain?
A buyer should be able to establish the item, offer and limitations without solving a puzzle across several documents.
My recommended starting structure is:
- A clear product name and selected option.
- The current price or an explicit quotation route, with the unit.
- Availability language appropriate to the business model.
- A concise specification table with units.
- Relevant use restrictions and what's included.
- Delivery, collection and return-policy links where applicable.
- A clear next action for that item.
This is a user-interface and information recommendation, not a required template for AI citations.
Use photographs to support the item, but keep essential specifications as text. A beautiful image labelled “available in many options” does not establish which options can actually be ordered.
Does Product schema solve catalogue accuracy?
Relevant structured data can describe products and offers, but it must represent the same item and conditions that users see.
Google documents variant-level identities and product groups. It also recommends initial-HTML Product markup for merchants seeking shopping results and notes limitations with dynamically generated price and availability markup.[1]
Validate the implementation for the applicable platform. A technically valid Product object can still contain the wrong price or describe a different variant.
For Google Search's AI features, there is no special AI schema requirement. Do not sell a generic markup block as proof that ChatGPT, Gemini and Claude now understand every SKU.[2]
Do product feeds help?
Feeds provide a structured way to submit catalogue data to a particular supported platform. They do not remove the need for accurate landing pages or current source records.
Google Merchant Center documents its product-data specification, including fields for identity, offers and variants. OpenAI separately documents product feeds for its supported commerce integration paths. The specifications and access conditions are platform-specific.[3][4]
Do not assume that a Google feed has been submitted to ChatGPT or that the same upload works everywhere. Review the current integration route and requirements before implementing.
OpenAI's guidance includes keeping variant-specific information aligned when options differ. Apply that principle only through the schema and delivery mechanism relevant to your actual integration.[5]
The operational question is whether the source record, feed item and landing page all describe the same purchasable offer at the same time.
A practical WooCommerce checklist
Before adding a new AI plugin, inspect the catalogue data you already manage.
| Area | Check |
|---|---|
| Product types | Simple, variable, grouped and external offers are used deliberately |
| Variations | Legitimate identifiers and meaningful attributes distinguish options |
| Price | Selected variation and public pricing conditions agree |
| Stock | Variant-level status does not conflict with the displayed offer |
| Dimensions | Product and shipping measurements are clearly labelled |
| URLs | A shared variant link selects the intended option where supported |
| Schema | One coherent offer is described; plugins do not produce contradictory values |
| Cache | Public pages reflect the intended current variant data |
| Feeds | Exported records map to the same items and market |
For quotation or rental catalogues, preserve that business model. Do not publish a fabricated fixed price to satisfy a feed or audit template. Check platform eligibility before submitting nonstandard offers.
What if the catalogue is a PDF?
Test the PDF as its own source. A text-based document, scanned brochure and interactive viewer may expose different information to different tools.
Use labelled text for product names, units and conditions, and avoid relying only on photographs of specification sheets. Consider a useful HTML page for important items or product families, with the PDF as a supporting download.
Keep the document's issue date and scope clear. A static catalogue cannot reliably establish live stock unless the business has an explicit process that keeps it current.
Compare PDF and website tests separately. If the uploaded PDF works, that does not establish public discovery of the document or correct retrieval of today's product page.
Retest prices and availability with a timestamp
Price and inventory can change between two tests. Capture the reference value and the assistant's answer within a documented window.
If stock changes before review, do not automatically label the original answer wrong. Inspect what the source established at the time, where you have evidence.
Repeat a small fixed sample after catalogue changes, feed mapping changes or template deployments. Record both technical fixes and observed answer outcomes, without assuming one caused the other.
Useful monitoring fields include:
Product family | Exact variant | Seller | Market | Currency
Reference timestamp | Public evidence | Test mode and prompt
Assistant and visible configuration | Response timestamp
Identity | Price/unit | Specifications | Availability | Conditions
Evidence support | Serious error flag | Action owner | Retest resultFor the general audit rubric, use the website facts test across the three assistants. For inaccurate claims requiring corrections, use the source-tracing guide.
Frequently asked questions
Why does ChatGPT show the wrong price for my product?
Check the variant, seller, currency, unit, date and cited source. A family-level starting price or old promotion may not match the requested item. The answer could also make an unsupported inference. Establish the source before choosing a fix.
Can AI tell the difference between product variations?
It can report variant information when the relevant evidence is available and interpreted correctly, but you should verify the exact SKU or option. A correct family name is not enough to establish correct selection.
Does a product feed guarantee correct shopping answers?
No. It supplies data through a supported platform mechanism. Data quality, mapping, landing-page consistency, freshness and the actual generated answer still need checking.
Can AI read my PDF catalogue?
That depends on the tool, file and mode. Test extraction and factual accuracy directly. An uploaded-document success does not prove that the public catalogue is discovered or used in shopping answers.
Should I create a separate page for every variant?
Choose an architecture appropriate to the catalogue and supported search guidance. Google documents single-page and multi-page variant approaches. Do not create thin duplicates solely for an unverified AI tactic.[1]
What if my products require a custom quote?
State the quotation model and the information needed to prepare it. The accurate answer may be that no public fixed price is established. Do not invent a purchase price or imply acceptance into a shopping programme without checking eligibility.
How many products should I test?
Begin with a small risk-based sample covering different offer types and commercially important items. Expand based on findings. Do not present that sample as a statistically representative error rate for the whole store.
Start with the item most likely to be misunderstood
Choose one variable product, define the exact offer and run the page, discovery and buying-question tests separately.
If the answer loses the unit, selects another variant or invents availability, preserve the example and trace the evidence. That gives you a concrete catalogue improvement or a documented model error rather than a vague readiness score.
For help reviewing product architecture, feed consistency and public offer information, request an eCommerce and AI-ready SEO audit. You can also read the broader AI shopping SEO guide.
- Can AI understand your business? – the website facts test: reference sheet, three test tracks and scoring
- Why can't ChatGPT read my website? – diagnose access, redirects, firewalls and missing content
- AI gets your business details wrong – trace the source and correct the facts
- Why does ChatGPT show old information? – find the outdated source and track the correction
- Domain changes, rebrands and AI search – preserve sources, test old links and identity after a move
- Can AI read your product catalogue? (this guide) – variants, prices, units and availability, checked
- Does AI know where you operate? – offices, service areas, delivery and remote markets
Want your catalogue checked?
I'm Jayant Solanki, an SEO, GEO and automation strategist. I review product architecture, variant data, feed consistency and the public offer information AI assistants and shoppers rely on.
Ranked #1 for "metal buildings" · +30% YoY organic traffic · Evidence-labelled research
Sources
Documentation checked on 17 October 2026. Product examples are fictional. Prompts and sample-selection rules are editorial recommendations, not benchmark findings. Feed access, schemas and eligibility must be rechecked before implementation.
- Google Search Central: Product variant structured data.
- Google Search Central: AI features and your website.
- Google Merchant Center: Product data specification.
- OpenAI Developers: Product feeds.
- OpenAI Developers: Commerce best practices.
Jayant Solanki
AI-Ready SEO, GEO & AIO strategist based in Indore, India. He works across technical SEO, analytics, website development and marketing automation for eCommerce, local-service and global businesses.
More about Jayant →