Which sources do AI shopping assistants draw on when recommending products?
AI shopping assistants build product answers from several kinds of source at once: product data supplied by merchants, pages crawled from the open web, third-party reviews and ratings, and marketplace or directory listings. Which of these a given assistant uses, and how it combines them, varies by product and by feature, and no vendor publishes a full weighting.
Merchant feeds are the most structured source. Where a retailer or brand submits a product data feed to a shopping platform, the feed supplies identifiers, price, stock status, images and attributes in a machine-readable form. Feeds are the cleanest way to tell an assistant exactly what a product is.
Web pages are the broadest source. Assistants that browse or retrieve content read product pages, spec sheets, comparison articles and help documents. Here, clarity matters more than volume: a page that states the product name, model, key specifications and use cases in plain text is easier to quote accurately than one that hides specs in images or tabs.
Reviews and marketplace listings act as corroboration. An assistant that finds the same specification on your site, in a retailer listing and in an independent review has more reason to repeat it. Conflicting figures across those sources are a common reason an assistant hedges or omits a product.
- Merchant product feeds (where the platform accepts them)
- Brand product pages and spec sheets crawled from the open web
- Third-party reviews, test results and ratings
- Retailer and marketplace listings
- Industry directories and comparison content
Which product feed attributes most affect whether an assistant can surface a product?
The attributes that matter most are the ones that let a system identify the exact product, confirm it can be bought now and state its key facts without guessing. In practice that means a unique identifier, a current price and availability, and a complete specification set. Optional fields such as lifestyle images or marketing descriptions help less than these.
Identifiers come first. A GTIN (the barcode family that includes UPC and EAN) lets a system match your product to the same item sold elsewhere. Where products have no GTIN, a manufacturer part number (MPN) plus the brand name is the usual fallback. A B2B product with neither should still publish a stable SKU and a model number on every page that describes it.
Price and availability must match the live page. A feed that shows a price of 499 while the page shows 549, or lists an item as in stock when it is discontinued, teaches a system to distrust your data. Update frequency matters as much as the initial value.
Specifications decide whether a comparison answer can include you. For a B2B item, the fields that buyers compare on, such as dimensions, materials, voltage, certifications, capacity, compatibility and warranty terms, should be filled in with units. Vague values like high quality or industrial grade give an assistant nothing to compare.
| Attribute | Why it matters | Common failure |
|---|---|---|
| GTIN, MPN or SKU | Matches the exact product across sources | Missing, or different IDs across feed and site |
| Price and currency | Lets the assistant quote a current price | Feed price differs from page price |
| Availability | Shows whether the item can be bought now | Stale stock status |
| Product title | Names the item the way buyers search | Internal codes instead of descriptive names |
| Specifications with units | Supports comparisons | Blank fields or marketing adjectives |
| Brand and category | Places the product in the right set | Generic or inconsistent category labels |
How much do third-party reviews and ratings influence AI product recommendations compared with a brand's own site?
Third-party reviews influence AI recommendations mainly as evidence about performance and reliability, not as a replacement for product facts. A brand's own site is usually the source for what a product is; reviews are the source for how it behaves in use. Assistants that answer questions such as which option is most durable tend to lean on independent sources, while questions about specifications tend to lean on the manufacturer.
No public study quantifies that split, so treat any single percentage you read as unverified. What can be observed is that a review which names a product model, states a use case and describes a measurable result is easier for an assistant to quote than a generic star rating. Reviews that are dated, unattributed or about a different model number add noise.
For B2B sellers, the practical steps are to collect reviews from verified customers on your own pages, to ask distributors and resellers to show the same model name and specification, and to reference independent test reports where they exist. Do not write reviews yourself or pay for placements that present as independent. That risks removal and undermines the trust signal you are trying to build.
Do marketplace listings and industry directories help or hurt visibility in AI-generated comparisons?
Marketplace and directory listings help when they confirm your data and hurt when they contradict it. They add surfaces where an assistant can find your product, price and reviews, but each listing is also a copy that can drift out of date. The net effect depends on consistency more than on the number of listings.
A useful way to think about it is as a set of mirrors. If the marketplace title, model number, specification and price match your feed and site, the assistant gets repeated confirmation. If one listing uses an old model name, a wrong voltage or a discontinued price, the assistant may discount all of them or cite the wrong one.
Industry directories work similarly for B2B. A complete, current entry in a recognised trade directory can confirm your company name, product category and certifications. An outdated entry with a former address or a withdrawn certificate is a liability. Audit each listing once a quarter against the same source of truth you use for your feed.
Which structured data should you implement to support AI product discovery?
Implement schema.org Product markup on every product page, with an Offer nested inside it for price, currency and availability. Structured data does not guarantee that an assistant will recommend a product, but it removes ambiguity about what the page describes and it matches the fields feeds already carry, so the two sources reinforce each other.
The minimum useful set is the product name, description, brand, identifier (gtin, mpn or sku), image, and an offer with price, priceCurrency and availability. Add aggregateRating only when it reflects real, published reviews, because markup that does not match visible content can be treated as misleading. Use the values displayed on the page, exactly as shown.
Validate the markup with a structured data testing tool after every template change, and check that the JSON-LD values match the feed. Mismatches are the most common error and the easiest to fix.
- Choose one source of truth for each product (PIM or ERP) and generate both the feed and the page markup from it.
- Add a Product block with name, description, brand, sku, mpn and gtin where they exist.
- Nest an Offer with price, priceCurrency, availability and a canonical URL.
- Add aggregateRating only when it matches visible, published reviews.
- Validate each template, then compare a sample of markup values against the live feed.
How can a B2B company test whether assistants recommend its products, and what should it fix first?
Test by running a fixed set of buyer prompts in each assistant you care about, recording what it says about your product, whether it names your brand, which sources it cites and whether the facts are correct. Repeat the same prompts on a regular schedule, because answers change between runs and across sessions. A single screenshot tells you little.
Build the prompt set from how buyers really ask. Use three types: category questions (best industrial sensor for a cold-storage site), specification questions (which valve handles this pressure and temperature) and comparison questions (brand A versus brand B for a given use). Log the answer, the citations and any errors in a simple sheet. Tools such as Rixe, an AI visibility platform, can automate this daily tracking across ChatGPT, Gemini, Perplexity and Google AI Overviews, which is useful when manual checks become too slow.
Fix in this order: correct factual errors in feeds and pages first, because a wrong price or spec damages trust in everything else. Then fill missing identifiers and specifications. Then tidy marketplace and directory listings. Only after that should you write new comparison or guide content, targeting the prompts where you are absent or misdescribed.
- Write 20 to 30 buyer prompts across category, specification and comparison types.
- Run each prompt in every assistant you track, once a week, from a clean session.
- Record whether your brand appears, how it is described, which sources are cited and any factual errors.
- Trace each error to its source: feed, product page, marketplace listing or review.
- Fix data errors first, then missing attributes, then listing consistency, then new content.
- Re-run the same prompts four weeks later and compare the logs.
What should a B2B team do first, and what remains unproven?
Start with data accuracy and identifiers, because those are the parts you control completely and the parts an assistant most needs to match your product. Content and reviews matter, but they build on a correct foundation. Doing the basics well is low-risk even if the exact weighting of sources turns out to differ from what you expect.
Be clear about what is verified and what is not. It is documented that structured data vocabularies such as schema.org exist and that feeds carry identifiers and availability. It is not publicly documented how any assistant ranks products against each other, what share of an answer comes from reviews, or whether a given markup change will change a recommendation. Plan on measurement, not on guaranteed outcomes.
A reasonable first quarter is to fix the highest-error products, align feed and page data, add markup to the main templates, collect verified reviews, and start the prompt log. Review the log after each cycle and let the observed gaps, not assumptions, set the next priorities.