Be the product the assistant recommends
Shoppers increasingly ask an assistant what to buy before they ever reach a store. That conversation happens entirely outside your site, it ends in a shortlist of a few products, and if your catalogue is not legible to the model you are not on it, regardless of how good your product pages are for human visitors.
For a retail or D2C team this is a catalogue problem before it is a content problem. Product structured data, feed accuracy, crawler access to product pages and presence on the review sources assistants trust decide whether a product is eligible to be recommended at all. This page covers what each assistant needs, what to measure, and the order to fix things in.
Why retail teams treat this as its own channel
The shortlist forms before the visit
When a shopper asks an assistant for the best option in a category with a budget and a use case, the answer is three to five products with a sentence of reasoning each. By the time anyone lands on your product page the comparison has already happened, and being absent from it leaves no trace in your analytics. There is no impression to lose and no bounce to see. That is why recommendation share on category prompts has to be measured directly rather than inferred from traffic.
Feed and schema completeness decide eligibility
Assistants that recommend products need structured facts to reason with: what the product is, who makes it, what it costs, whether it is in stock, and what people think of it. Missing GTIN, brand, price, currency or availability in Product structured data, or a merchant feed with stale stock, frequently disqualifies a product before quality is even considered. This is mechanical work, it can be automated across a catalogue, and it is where most retail teams find their fastest win.
Third-party sources dominate recommendations
Review platforms, editorial roundups, community threads and marketplace listings carry more weight in shopping answers than a brand's own site does. An assistant asked for the best option leans on sources that compare several products, which your product page by definition does not. Your own site is one input among many, and the citation data for your category will name the others. That list is the outreach brief.
Every market is a separate answer
Sources differ by country, currency and language, so the products an assistant recommends in one market say little about another. A brand that is well covered by UK review sites can be absent from the same question asked in Germany or India. If you sell in several markets, the prompt set, the structured data and the source work all need to run per market rather than be assumed to generalise. GPT Shopping at /features/gpt-shopping tracks recommendation share per assistant and per market.
How AI shopping visibility works
Making a catalogue legible to assistants that recommend products, then measuring whether it worked, in the order that finds problems cheapest first.
- 1
Complete your Product structured data
Audit every product page for Product JSON-LD with name, brand, GTIN or MPN, image, description, offers with price, priceCurrency and availability, and aggregateRating where reviews exist. Do it across the whole catalogue with a crawl, not a sample, because gaps cluster in older or imported products. Keep the markup identical to what the page displays: a schema price that differs from the visible price is worse than no schema. Generate and check templates with /tools/json-ld-generator and /tools/json-ld-validator, then have the platform emit it for every variant.
- 2
Get the feed right
Where an assistant supports a merchant or product feed, feed accuracy and freshness govern eligibility. Reconcile the feed against the site for price, stock and product identifiers, and fix the sync cadence so a sell-out is reflected within hours, not days. Stale data is worse than absent data because it produces a wrong answer with your name on it. Use the same identifiers in the feed, the page schema and the marketplace listings so an assistant can tell they are one product.
- 3
Open access to shopping assistants
Confirm product and category pages are fetchable by OAI-SearchBot and ChatGPT-User, PerplexityBot and Perplexity-User, and ClaudeBot, and that bot-management rules do not challenge them. Check the served robots.txt with /tools/robots-txt-validator, then fetch a product page with a non-JavaScript client and confirm the price, availability and JSON-LD are in the raw HTML. Headless storefronts often render these client-side, and those are the exact fields the assistant is looking for.
- 4
Fix the catalogue content assistants read
Product descriptions written for a human skim often lack the facts an assistant needs to match a use case: materials, dimensions, compatibility, who it is for and who it is not for. Add a short specification block and a plain-language "best for" line to each product, and make sure category pages have a real introduction that explains how to choose, not just a grid. This is the on-site equivalent of the roundup an assistant would otherwise cite.
- 5
Measure recommendation share
Build a prompt set of the category, budget, use-case and comparison questions your shoppers ask, per market, and run it on a schedule against each assistant. Record which products are recommended, in what order, with what reasoning, and which sources are cited. Recommendation share on category prompts is the headline metric; the cited-source list is the work plan. Track new products from launch so you can see when they first become eligible. The Visibility Tracker at /features/brand-visibility-tracker runs the category prompts.
- 6
Work the review and roundup sources
Where the citations for a category question come from review sites, editorial roundups and community threads, that is where the effort goes. Make sure the product exists on the review platforms that get cited, that the listing is complete and current, and that recent reviews exist. Pitch editorial roundups with the facts they need. Reddit monitoring at /reddit-monitoring shows which community threads are being cited and where the brand is discussed. See /solutions/off-site-authority for the full playbook.
Retail use cases
Category recommendation share
How often your products appear when an assistant answers "best X for Y" in your category, per assistant and per market. This is the retail equivalent of category rankings and the number the merchandising team will want on the monthly report.
Competitor benchmarking
Which brands the assistants recommend instead of you, in what order, and which sources they cite for it. The source list matters more than the ranking, because it tells you where the recommendation is actually decided and who you need to be present alongside.
Launch visibility
Whether a new product is discoverable to assistants at all in its first weeks. New SKUs often lack reviews, feed entries or complete schema, so they are invisible long after they are live on the site. Track from launch to see the moment eligibility is reached.
Catalogue health
Structured data and feed completeness across the full catalogue, checked continuously rather than once. Gaps recur every time products are imported, variants are added or a theme is updated, and each gap is a product that cannot be recommended.
Price and availability accuracy
What assistants say your product costs and whether it is in stock, compared with the truth. Wrong answers here cost sales and trust, and they usually trace back to stale feeds or old third-party listings the model still trusts.
What differs per assistant for retail
Eligibility rules and source preferences differ by surface. Measure each one; the fixes overlap but the results do not.
ChatGPT
Shopping answers combine live page fetches through OAI-SearchBot and ChatGPT-User with merchant product data where a feed is supported. Product schema, accurate price and availability, and a fetchable, server-rendered product page are the eligibility checks. For "best" questions it leans on review and roundup sources, so your presence there decides whether you are named at all.
Perplexity
Perplexity shows its sources prominently and leans on recently updated pages, so it is the fastest place to see whether schema and content changes registered. PerplexityBot and Perplexity-User both need access. Shopping-style answers cite review sites, retailer listings and community threads heavily; a complete, current listing on the platforms it cites matters as much as your own page.
Google AI Overviews and AI Mode
Both draw on the Google index and Shopping Graph, so Merchant Center feed quality and Product structured data feed directly into product-level answers. Crawling is via Googlebot, so existing coverage carries over. Price, availability and identifier consistency between feed and page are the most common issues, and when the two disagree one of them is ignored.
Gemini
Gemini grounds on Google Search and Shopping data, so the same feed and schema work applies. Google-Extended governs whether content is used for Gemini training and grounding beyond search, so blocking it can thin coverage here specifically. Entity clarity matters: a consistent Organization record with sameAs helps Gemini tie products to the right brand.
Amazon Rufus
Rufus answers from Amazon's own catalogue, listing content, reviews and Q&A, not from your website. If you sell on Amazon, listing completeness, attribute fill, A+ content and review recency are the levers, and your site is irrelevant to it. If you do not sell there, you will not appear, and shoppers asking Rufus are being recommended competitors who do.
The numbers retail teams report
Recommendation share is the headline; the rest explain why it moved or did not.
- Recommendation share
- The share of tracked category prompts where at least one of your products is recommended, per assistant and per market. Read it as eligibility plus preference: a product that is never recommended is usually ineligible for a mechanical reason before it is losing on merit.
- Catalogue completeness
- The share of products with complete Product schema, a matching feed entry and a server-rendered page. It is a leading indicator: recommendation share cannot rise for products that fail it, and it drops silently after imports and theme changes.
- Cited-source list
- The domains cited when assistants answer your category questions, ranked by frequency. Each entry is either a review platform to be present on, a roundup to pitch, or a competitor page to outcompete. It is the work plan rather than a KPI.
- Price and availability accuracy
- The share of answers where the stated price and stock status match reality. A wrong price or an out-of-stock recommendation is a lost sale and a trust hit, and it traces to feed cadence or stale third-party listings.
Mistakes retail teams make on the first pass
Mistake · Checking schema on a sample of hero products and assuming the catalogue matches.
Fix · Crawl the full catalogue for Product JSON-LD completeness on a schedule. Gaps cluster in older, imported and variant products, and each gap is a product that cannot be recommended.
Mistake · Letting the feed and the page disagree on price or stock.
Fix · Reconcile feed against site automatically and fix the sync cadence. When the two disagree one of them is ignored, and stale data produces confident wrong answers with your name attached.
Mistake · Rendering price, availability and schema client-side on a headless storefront.
Fix · Server-render or pre-render product and category templates so those fields are in the raw HTML, then verify with a non-JavaScript fetch. Retrieval agents do not wait for the bundle.
Mistake · Treating the product page as the main source for "best X" answers.
Fix · Assume the recommendation is decided on review platforms, roundups and community threads, and work those sources. A product description cannot compare products; the sources assistants cite can.
A worked example: a D2C skincare brand on Shopify
A hypothetical D2C skincare brand on Shopify sells well through paid social but never appears when shoppers ask an assistant for a moisturiser for sensitive skin under a given price. A catalogue crawl shows Product schema is present in the theme but missing GTIN on most variants and missing aggregateRating everywhere, because reviews are loaded by a third-party widget after the page renders.
The team fixes the schema template to emit GTIN and brand on every variant, moves the review summary into the server-rendered HTML with AggregateRating markup, and adds a short specification block to each product: skin type, key ingredients, fragrance-free or not, and a plain "best for" line. A non-JavaScript fetch of a product page now returns price, availability and rating in the first response.
A prompt set of forty questions across the UK and US, covering skin concerns, budgets and comparisons against named competitors, is tracked weekly across ChatGPT, Perplexity, Google AI Mode and Gemini. The cited-source list shows two beauty editorial sites and one large community thread deciding most of the answers. The brand is on neither editorial site and is mentioned in the thread once, unfavourably, over an old formulation.
The marketing team pitches both editorial sites with the specification facts, and the community thread is addressed with a factual reply about the reformulation. Over the following months the product begins to appear in Perplexity answers first, then ChatGPT, with the editorial roundup cited as the reason. Recommendation share is reported monthly next to paid social, and the catalogue completeness check runs after every product import.
Frequently asked questions
Do I need a merchant feed to appear in AI shopping answers?+
Not universally. Some assistants read a merchant feed, others read on-page Product structured data, and many recommendations come from third-party review content that mentions you. Complete on-page schema is the common denominator and the right starting point. Add a feed where an assistant supports one, and keep the two consistent.
Why is my best-selling product not recommended?+
Sales rank is not an input the assistant sees. The usual causes are incomplete Product schema, a blocked or JavaScript-dependent product page, stale feed data, or absence from the review and roundup sources the assistant cites for that category. Check them in that order; the first three are quick to verify.
Does this apply outside the US?+
Yes, and answers vary considerably by market because the sources differ. Review sites, retailers and community threads are national, and pricing and availability are per market. If you sell in several countries, build a prompt set per market and measure each rather than assuming the US result generalises.
How does this relate to agentic checkout?+
Visibility gets you recommended; agentic commerce protocols let an agent transact on the shopper's behalf. They are sequential. Being buyable by an agent is worth little if you are never on the shortlist, and being on the shortlist without a buyable path hands the sale to a retailer who has one. Agentic Readiness at /features/agentic-readiness checks both.
Which crawlers do I need to allow for shopping assistants?+
The retrieval agents: OAI-SearchBot and ChatGPT-User for ChatGPT, PerplexityBot and Perplexity-User for Perplexity, and ClaudeBot for Claude. Google AI Overviews and AI Mode use Googlebot and Merchant Center data. Training crawlers such as GPTBot and CCBot are a separate policy decision and should be handled with separate rules.
Do reviews on my own site count?+
They help when they are marked up with Review and AggregateRating schema and visible in the HTML, because they give the assistant a rating to reason with. They carry less weight than independent review platforms, which assistants treat as more trustworthy for comparisons. Do both; the on-site markup is the cheaper half.
My storefront is headless. Is that a problem?+
It can be. Headless builds often render price, availability and structured data on the client, which means a retrieval agent fetching the raw HTML sees a product page with no product on it. Server-render or pre-render product and category templates so those fields are in the initial response, then verify with a non-JavaScript fetch.
How quickly do changes show up in answers?+
Schema and rendering fixes are picked up when the page is next fetched, which for retrieval agents happens at query time, so those can register within days. Feed changes follow the feed sync cadence. Off-site work is slower because it depends on third parties publishing. Track weekly and report monthly.
Explore the features behind this solution
Go deeper
See whether assistants recommend your products
Check catalogue legibility and recommendation share across the assistants your shoppers use.