AI Shopping SEO: Getting Products Recommended

AI shopping assistants recommend products the way a knowledgeable salesperson does: they match a stated need to a product they believe fits it, and they need facts to justify the match. That makes AI shopping SEO a data problem before it is a content problem. The assistant can only recommend what it can identify, price, verify as in stock and describe accurately, and most of those facts come from your product feed, your product page markup and the reviews attached to your items.

This is a different job from getting an article cited in an AI answer. Citation rewards a clear passage that answers a question. Recommendation rewards a complete, consistent, machine-readable product record. This post covers which inputs these systems draw on, how eligibility works in practice, and what a store can change, without pretending anyone outside the platforms knows the exact ranking formula.

Recommendation Is Not Citation

When an assistant answers a question like "why does my sourdough collapse," it needs a passage it can paraphrase and attribute. When it answers "best waterproof hiking boots for wide feet under 150 dollars," it needs a short list of specific products, each with a price, availability, a link and a reason. The second task leans on structured product records rather than prose, and prose pages are the wrong tool for it.

Shopping-oriented features in the major assistants and search engines generally sit on top of commerce data infrastructure that predates the chat interface. Google has a Shopping Graph fed largely by merchant feeds and page data. Other assistants either license or crawl product data, or accept feeds from merchants directly. The details differ by platform and change often, so treat any claim about one system's exact inputs with caution. What stays stable is the kind of evidence all of them need: identity, attributes, price, stock, policies and independent opinion.

The practical consequence is that your blog can be excellent and your catalog still invisible. Teams that invested in informational content for answer engines often discover that their product pages are thin, their feed is stale and their variants share one title. Fixing that is the work this post is about.

The Inputs Assistants Can Actually Use

It helps to sort the evidence into four layers, because each one is fixed in a different place and by a different person on your team.

• Feed data. A structured file or API submission, most commonly through Google Merchant Center, listing each product with an ID, title, description, link, image, price, availability, brand and identifiers. This is the most direct channel, and for some surfaces it is the only one that counts.
• On-page data. Product structured data in schema.org vocabulary (Product, Offer, AggregateRating, Review), plus the visible page itself. Crawlers use it to confirm the feed and to fill gaps, and a mismatch between the two is a red flag.
• Review and reputation data. Ratings and review text on your site, in review programs you participate in and on third-party sites where your products are discussed. Assistants lean on this to decide which of several similar products to surface.
• Merchant and policy data. Shipping cost and speed, return policy, payment options and basic trust signals about the store. A product can be perfect and still lose to one from a merchant whose terms are clearly stated.

Feed Attributes That Decide Eligibility

Most eligibility failures are boring. The product is disapproved, or approved with warnings, or approved but missing the attribute a shopper's query depends on. Work through the feed the way a reviewer would.

Identifiers. A GTIN (the barcode number), an MPN and a brand let a platform recognize that your product is the same physical item sold by others. Without them your listing is an orphan that cannot be matched to reviews, specifications or price comparisons. Google's feed specification requires GTINs for many branded products, and supplying made-up values is worse than leaving the field out, because it triggers disapproval. If you make the product yourself, a consistent brand and MPN still matter.

Titles. A feed title is not a page title. It should lead with what the thing is, then the attributes a shopper filters on: brand, model, size, color, material, quantity. "Trailhead Waterproof Hiking Boot, Wide Fit, Men's Size 10, Brown" gives an assistant something to match. "Best Boots Ever!!!" gives it nothing and breaks feed policies.

Descriptions. Write them for extraction. State what the product is for, what it is made of, dimensions, compatibility, care instructions and what is in the box. Conversational queries often hinge on one of these details, such as "fits a 15-inch laptop" or "safe for sensitive skin," and if the detail is not in your data, the product cannot be selected for that query.

Price and availability. These must match the landing page, and they must be current. Assistants that surface prices want to avoid sending a person to a page that contradicts the answer. A feed that updates once a day while your stock changes hourly will produce exactly that contradiction. Most platforms let you submit incremental updates or use an API for price and stock, and this is worth setting up.

Category and product type. Use the platform's taxonomy for the category and your own detailed product type for the path a human would recognize. Both are used, and a product placed in the wrong category is effectively filed in the wrong aisle.

Product-specific attributes. Apparel needs gender, age group, size and color. Electronics need model numbers and compatibility. Food and supplements need ingredients and dietary flags. These optional-looking fields are the ones that long, specific conversational queries map onto.

Structured Data on the Product Page

Markup is the second line of evidence, and it is also the one that works when you are not submitting a feed at all. A complete Product block names the item, its image, brand, SKU and GTIN, and nests an Offer with price, currency, availability and condition. Add shipping details and return policy markup where the vocabulary supports it, since those answer the questions assistants increasingly get asked.

The most common implementation mistake is generating markup from a template that ignores variants. A shirt sold in six sizes and four colors should not present as one product with one price and one availability value if those differ. Either mark up the selected variant accurately or expose variants through separate URLs or a proper variant structure, and keep the values in agreement with what a visitor sees.

Validate with the Rich Results Test and watch the enhancement reports in Search Console. Then compare the markup against the feed for the same product. If the price is 49.00 in one and 54.00 in the other, a platform has to choose which to trust, and sometimes it chooses neither.

Also keep the page honest in plain HTML. A product whose specifications exist only inside an image, or whose price is injected by script that a crawler never runs, is incompletely described no matter how good the markup is.

Reviews: What They Do and What They Cannot Fix

Reviews serve two jobs in a recommendation. Quantitatively, a rating and a count help rank similar products. Qualitatively, the text supplies the reasons an assistant repeats: "reviewers say the sole grips well on wet rock but runs narrow." That sentence is only possible if someone wrote the review it came from.

Specific reviews are worth more than enthusiastic ones. A review that mentions the use case, the size ordered, how long the item has been used and one flaw is informative to both people and machines. You can encourage this without manipulating anything by asking a neutral follow-up question after purchase, such as what you used it for and whether the sizing was accurate, rather than asking for five stars.

Do not fabricate, incentivize selectively or gate reviews to positive customers. Review platforms and consumer protection regulators treat that as deceptive, and the penalty is worse than a mediocre average. Conversely, a handful of candid negative reviews that mention a fixable issue, answered by the merchant, often read as more credible than a spotless record.

Reviews also live outside your store. Independent roundups, forum threads, comparison sites and video reviews all shape what a model has absorbed about a product category. You cannot control them, but you can make sure products reach reviewers honestly, that your brand and model names are used consistently and that your own pages contain the specifications those outside sources will quote.

A Worked Example: One Listing, Three Fixes

Take a small shop selling insulated water bottles. Its listing for the flagship model has the title "Summit Bottle," a two-sentence description about "adventure," no GTIN, one image and a feed refreshed weekly. A shopper asks an assistant for a leak-proof insulated bottle that fits a car cup holder and keeps drinks cold through a long day.

The assistant cannot confirm any of those attributes from this listing. There is no capacity, no diameter, no statement about the lid seal and no insulation detail. It will recommend a competitor whose record states a 0.75 liter capacity, a base diameter that fits standard cup holders and a gasketed lid.

The fixes are unglamorous. First, rewrite the title to the pattern of brand, product, capacity, material and color, and add the GTIN from the barcode the shop already prints on its packaging. Second, add the specifications the query turned on, such as base diameter and lid type, to both the description and the on-page table, then mirror them in the structured data. Third, move stock and price to more frequent updates and confirm the landing page matches.

None of this guarantees a recommendation, and nobody outside the platforms can promise one. What it does is remove the reasons the product could not be recommended, which is the only part the shop controls.

Merchant Signals and Product Page Content

Assistants that recommend products have to worry about sending people somewhere unsafe or disappointing. Store-level information therefore carries weight. Publish clear shipping costs and delivery windows, a readable return policy, a real contact method and accurate business details. Where a platform offers a merchant profile or policy submission, complete it fully. Incomplete merchant information is a common cause of products being approved but rarely shown.

On the product page itself, answer the questions a buyer asks before purchasing. A compact questions-and-answers block, a sizing guide, a compatibility list and a comparison against your own sibling products all give an assistant extractable sentences. They also reduce returns, which is a business reason to do them that has nothing to do with AI.

Keep pages crawlable. If you block the crawlers of assistants you want recommendations from, the catalog may still appear through feeds on some platforms, but page-level evidence is unavailable to them. That decision is a real trade-off and depends on the platform and on your goals, and it is worth making deliberately rather than by an inherited robots file.

Measuring Whether It Worked

Measurement here is imperfect, so be modest about it. Start with platform diagnostics: Merchant Center reports on disapprovals, warnings and missing attributes, and these are deterministic. Clearing them is progress you can verify even if you cannot see the recommendation layer.

Then test by prompting. Keep a list of the shopper questions your best-selling products should answer, run them in the assistants your customers use, and log whether your products appear, how they are described and which competitors appear beside them. Answers vary from run to run, so repeat each prompt several times and look for patterns rather than single results.

Finally, watch referral traffic. Some assistants pass identifiable referrers or tagged links and some do not, so segment what you can and accept a blind spot for the rest. Compare the same products before and after feed fixes, and change one thing at a time so you can tell which fix did the work.