%20-%20kopie.png)
Cart
0
your cart
No items found.
Product is not available in this quantity.
By AI Marketing Entrepreneurship · Published October 13, 2026
Prepared with AI assistance using linked primary sources. Facts checked October 13, 2026.
An AI shopping audit tests two different things: whether a product appears for a relevant need, and whether the information shown is accurate enough to support a purchase. Start with real customer requirements and a small product sample. Being included in an answer is useful only when the product is suitable and the buying information is dependable.
OpenAI describes shopping research as an experience that can clarify preferences and consult product information. That makes incomplete specifications and unclear suitability worth investigating. It does not mean every merchant or product will appear for every relevant request. [S9]
Select around ten products representing your catalogue: a best seller, a new item, a product with variants, one with compatibility requirements, and an item that is frequently returned because expectations were wrong. Include a product with excellent documentation so the audit has a useful reference point.
Create a dated fact sheet for each item. Record the exact name and variant, price and currency, stock state, dimensions, materials, compatibility, delivery conditions, returns policy, and canonical product URL. These facts are your verification reference, not a promotional script.
For example, a fictional compact desk lamp may be 38 cm tall, need mains power, and offer adjustable brightness. If it is not rechargeable, that limitation belongs in the fact sheet and on the product page. An attractive recommendation that calls it cordless would be a failed result.
Use three prompt types. A need-led prompt describes the situation without naming your brand: a reading lamp for a small desk, within a specified budget and delivery market. A constraint-led prompt adds requirements such as maximum height, power source, or compatibility. A named-product prompt asks about a particular item and is primarily an accuracy check.
Do not combine those results into one discovery score. Naming your product gives the assistant information it did not have in the unbranded test. Also avoid loading the prompt with unusually specific wording copied from your page; that can turn a customer test into a page-finding exercise.
Use a practical matrix of five shopping needs, each expressed as a broad question and a constrained question. Repeat the resulting ten prompts in each assistant on two dates. Test named products separately. Record the user market, language, visible product or mode, and whether a dedicated shopping experience was available. Also record memory and personalisation settings: OpenAI says shopping research may use saved memory to tailor recommendations. Keep these settings consistent and do not pool a dedicated shopping mode with a general chat mode without labelling the difference. [S9]
For each response, capture the recommended product, merchant, displayed price, linked destination, and explanation of fit. Open the destination at the time of review. Check whether it is the correct variant and whether the proposed use actually matches the specification.
| Check | A useful pass condition |
|---|---|
| Product identity | The link and description refer to the same item and variant |
| Suitability | The product meets the shopper's stated constraints |
| Price and market | Currency, offer conditions, and market match the request |
| Availability | The destination reflects the current purchasable state |
| Buying information | Important delivery, compatibility, or return conditions can be verified |
Mark missing facts separately from false ones. A recommendation can be accurate but incomplete. Record contradictions, such as a product summary claiming an included accessory that the destination lists as optional.
Inspect inconsistent product names, missing variant specifications, stale offers, and policy pages that are difficult to find. If your feed, product page, and support answers disagree, rewriting the introduction will not resolve the core issue.
Write factual descriptions around customer decisions. For the fictional lamp, show height, footprint, power source, brightness controls, and supplied accessories in visible text. Explain who it suits and the situations where it does not. Keep essential details accessible without requiring a visitor to interpret a lifestyle image.
Google documents Product structured data and Merchant Center feeds as ways to provide product information for supported Google experiences. Keep any markup consistent with the visible page and the actual offer. These mechanisms are not a universal submission system for every assistant, and they do not guarantee inclusion. [S10]
OpenAI separately says shopping research may use merchant product data supplied through the Agentic Commerce Protocol, alongside public product information and other retail sources. Check its current merchant participation guidance if that route applies to your business; supplying product information is not a guarantee of recommendation. [S9]
For ordinary Google Search AI features, Google says there is no separate special markup requirement. Do not confuse a search eligibility check with a promise of a Gemini app recommendation. [S1]
Choose the most consequential issue, such as a missing compatibility restriction. Update the relevant catalogue fields, page text, and maintained feeds together. Record the exact changes and their dates. Avoid simultaneous large rewrites if you want to understand which information was confusing.
Repeat the original prompts after allowing time for systems to encounter the new information, and continue checking the source page itself. There is no fixed recrawl or recommendation timetable that applies across all products. Changes in assistant behaviour can also reflect changing availability, competing offers, or system updates.
Report suitable recommendations, accurate recommendations, valid product links, and unresolved errors separately. For a hypothetical set of 20 valid responses, appearing in five is 25% observed inclusion. If two of those five recommend an incompatible product, raw inclusion overstates useful visibility. Those numbers illustrate the calculation; they are not measured performance.
Where consented analytics and order records allow it, inspect what identifiable visitors do after arriving. A product page that answers the right question may reduce confusion even when referral volume is small. Avoid attributing every later sale to an AI recommendation without evidence of the journey.
Usually the first task is to improve the existing product page. Creating parallel pages with conflicting prices or specifications introduces another maintenance problem. Add a guide only when it answers a distinct comparison or use-case question.
Audit ten products, identify the three most consequential information gaps, and correct them consistently. Keep the original prompt set and evidence so the next review measures change against a documented baseline.
%20-%20kopie.png)
