Being the product AI recommends

A small brand makes an insulated bottle that's lighter than anything the big names sell. Asked for the best bottle for a long hike, an assistant names two big brands and a third brand the founder has never heard of, which sells mostly through marketplaces. Nothing about the bottle is the problem. Getting recommended is three jobs done in a fixed order: a product page that states the facts, the same facts everywhere else the product appears, and other people saying so.

First, a product page that states the facts

An assistant reads your product page the way a hurried expert would: as text and data, looking for the answer to a specific sub-question. The page earns a place in the answer when the facts are there in a form a machine can lift out.

1. One product per page, with its canonical name in the title and the main heading, spelled exactly as it's spelled everywhere else. Settle the variant question first, so there's one URL for the product rather than one per colour (Variants and duplicate content).
2. A first paragraph that says what it is, who it's for and what makes it different, in plain sentences. "An insulated steel bottle for day hikes, built to weigh less than others of its capacity" is quotable. "Hydration, reimagined" is not.
3. Specs as text in a table: capacity, weight with and without the lid, materials, dimensions, what's in the box. Specs inside an image or a PDF don't exist to the model.
4. Product structured data stating the price, the currency, the availability, the brand and the GTIN, matching the visible page.
5. The comparison questions answered on the page. Who it's not for. How it differs from your other bottle. Whether it fits a standard car cup holder. These are the sub-questions from AI shopping answers, answered at source.

Then make sure the page can be fetched. Google documents that its AI features in Search use the same crawl as Search itself, so a page Google indexes is available to them. OpenAI, Anthropic and Perplexity each document the name of their own crawler, and your robots.txt decides whether those crawlers can read the page. llms.txt and AI crawlers lists them and what to allow. Writing for AI citations covers the passage structure that makes a page easy to quote.

Second, the same facts everywhere the product appears

Assistants cross-check. When a product appears in several places with the same name and the same key specs, the model can be confident it's one product and describe it. When the store says "Summit 750", a marketplace says "Summit Bottle 0.75L", the feed says "SUMMIT-750-BLK" and the weights disagree, the model has three half-products and names none of them. That preference for agreeing sources is observed in how assistants answer rather than documented by any of them, and it's consistent enough to build on.

The anchor is the GTIN. Google's Merchant Center help documents how GTINs let it match one item across sellers into a single product in its catalogue, and the same number is what product databases elsewhere key on. Every listing of your product, everywhere, should carry the same GTIN, the same brand, the same model name and the same core specs.

The practical fix is a product fact sheet: one document per product with the canonical name, the GTIN, the specs as you want them stated, and the approved short description. Your product page, your feed, your marketplace listings and any stockist's listing are copied from it. When a stockist writes their own version, send them the sheet. Consistency here is boring, and it's the step most brands skip while wondering why a marketplace-only competitor gets named.

Third, other people saying so

Reviews on your own page establish that customers exist. Mentions elsewhere establish that the product is known, and that's what separates a product that gets named from one that's merely accurate. Roundups that already rank for your "best" searches, comparison articles, forum threads where someone recommends the bottle by name, video reviews, and reviews on retailer sites all count, and they're read directly by the assistants that search before answering.

Earning them is ordinary work. Find the roundups that rank for the "best [product type] for [use]" searches you identified in Buying intent for stores and offer the writers a product to test. Answer questions in the communities your buyers use, as yourself, and let your product come up when it's the honest answer. Publish your own comparison pages that name competitors fairly (Comparison and FAQ content), because an honest comparison is exactly the passage an assistant quotes.

Don't manufacture any of it. Seeded forum posts and bought reviews breach Google's spam and review policies and, in the US, the Federal Trade Commission's rule on fake reviews, and a model summarising sentiment reads an astroturfed thread the same way a person does.

Check it, then keep checking

Work in the order above. A mention in a roundup does nothing for a product whose page can't be read, and a perfect page does little for a product the rest of the web has never heard of.

After the fixes, ask the question yourself. Once a month, put the same "best [product] for [use]" question to the assistants your customers use, in the same words, and record which products are named and which pages are linked. Answers shift between sessions, so a product that appears in most of them is the signal and a single absence isn't. Measuring your AI visibility covers how to keep that record honest.

What to take away

• Getting recommended is three jobs in order: a product page that states the facts as text and data, the same facts on every listing of the product, and independent mentions.
• A product page earns a place in an answer when its name, purpose, specs, price and availability can be lifted straight out of the text and the structured data.
• The GTIN and a single product fact sheet keep your store, your feed, marketplaces and stockists describing one product, which is what assistants look for.
• Mentions in roundups, forums and reviews are what get a product named rather than merely described, and they're earned, never manufactured.

Next

The chapter on measurement starts with tying search and AI visits to orders, in Tracking revenue from search and AI.