How to Do GEO for Ecommerce in AI Search: ChatGPT Shopping, Gemini, Google AI Mode
GEO for ecommerce in eight steps: which pages AI engines cite for product questions, the feed and schema fields they read and how to track a whole store.
Quick answer
Canlah AI runs GEO for ecommerce, under the method in this guide, as eight steps across two separate retrieval paths: one through crawlable product and category pages for ChatGPT and Perplexity and one through a merchant feed for Gemini and Google AI Mode. In a Canlah AI citation study of 12 Singapore consumer-product questions and 95 citations recorded on September 23, 2026, ChatGPT took 54% of its 68 citations from product pages and 21% from category pages. Google AI Mode took 44% of its 27 citations from Google Shopping cards and 15% from YouTube. The two engines cited no URL in common, so each path has to be built and measured on its own.
Canlah AI works both layers as one program and reports per engine, not as a single store score. Canlah AI is a Singapore-based SEO + GEO agency that measures and improves brand visibility inside ChatGPT, Gemini, Google AI Overviews and Google AI Mode, reporting AI visibility as re-verifiable ranges with timestamped evidence. It does not publish a self-serve SKU-level shopping dashboard, so a merchandiser who needs daily tile positions for thousands of SKUs will need a dedicated tracker alongside it.
Treat any single answer as directional. One run cannot prove stable product placement, causation or future recommendations. If a first check reveals a gap, freeze a panel of 20 to 40 buyer questions and repeat it by engine before rebuilding anything.
Disclosure: Canlah AI publishes this guide and operates the paid measurement and execution service described near the end. Engine behaviour comes from OpenAI’s and Google’s own documentation, and the citation shares from Canlah AI’s own probe archive. Pages were reviewed on September 23, 2026 and no vendor product was tested through a paid account.
What does AI visibility for ecommerce measure?
AI visibility for ecommerce measures five separate things for every product question: brand mention, product card presence, merchant row, attribute accuracy and source path. A report that collapses the five into one rank hides the level that is actually broken.
- Brand mention: whether the answer names the store when a shopper asks a category question without naming any brand.
- Product card presence: whether a specific product appears as a tile in ChatGPT Shopping or as a Google Shopping card in AI Mode, and at which position.
- Merchant row: whether your own storefront is listed first once a shopper opens the product, or whether a marketplace reseller of your own item is listed instead.
- Attribute accuracy: whether price, availability, size, material and shipping match what your store stated on the date of the answer.
- Source path: which page the answer cited, from your product page and your category page to a retailer, a review publisher or a video.
A brand can hold every product tile and still lose the sale, because the merchant row is decided by different data than the tile.
Which pages AI engines cite for ecommerce questions
ChatGPT cited brand and retailer product pages for 54% of its 68 ecommerce citations, while Google AI Mode cited Google Shopping cards for 44% of its 27, in Canlah AI’s study of 95 citations recorded on September 23, 2026. ChatGPT went to the origin: category pages supplied a further 21% and government pages such as the National Environment Agency’s air-quality FAQ supplied 13%, used as a judgment anchor rather than as a shop. Google AI Mode went to the aggregate, with YouTube videos supplying a further 15% of its citations.
Cited source types for 12 Singapore consumer-product questions, 95 citations across 59 domains, recorded September 23, 2026.
| Cited source type | ChatGPT (68 citations) | Google AI Mode (27 citations) | What the store controls |
|---|---|---|---|
| Brand or retailer product page | 54% | Not a leading type here | Page copy, merchant listing schema, crawl access |
| Category or collection page | 21% | Not a leading type here | Category copy, internal links, comparisons |
| Government or regulator page | 13% | Not a leading type here | Nothing directly; cite the standard yourself |
| Google Shopping card | Not a leading type here | 44% | Merchant Center feed, price, stock, shipping |
| YouTube video | Not a leading type here | 15% | Owned channel, creator seeding, demonstrations |
Source: Canlah AI citation study, September 2026. Twelve questions came from Google autocomplete for the Singapore locale, two of them in Chinese, and 19 cited pages were read line by line. Answers were stored truncated at 600 characters and three merchant sites returned 403. “Not a leading type here” means the type did not appear in that engine’s ranked source table, not that it never occurs.
Shares drawn from 68 and 27 citations are directional. The finding that survives the small denominator is the zero overlap: the two engines cited no URL in common, which is why store-wide AI visibility cannot be reported as one number.
How to do GEO for ecommerce, step by step
GEO for ecommerce takes eight steps in dependency order: access, schema, page copy, Google feed, ChatGPT and Perplexity feeds, category pages, off-site evidence, then measurement.
Step 1: Open crawl access to both paths before writing anything
Access decides whether any later work can be seen. OpenAI’s crawler documentation states that sites which opt out of OAI-SearchBot will not be shown in ChatGPT search answers, and that robots.txt changes take about 24 hours to apply. Googlebot governs the page path and Google-Extended governs model training; neither controls the feed path, which is a Merchant Center account rather than a crawl.
Few cross-border stores address AI crawlers by name. In Canlah AI’s open agent-readiness dataset of 50 Chinese-origin cross-border DTC brands collected in August 2026, 6 of 47 reachable storefronts (12.8%) named any AI agent in robots.txt and 1 of 47 exposed an MCP endpoint. The sample is purposive and does not project onto ecommerce generally.
Step 2: Put merchant listing structured data on every product page
Google requires four properties for a merchant listing, yet 38 of 47 reachable storefronts in the same August 2026 sample carried no Product schema on the sampled product page. The required properties are Product name, Product image and a nested offers object carrying price or priceSpecification.price above zero and a three-letter priceCurrency. Google’s documentation adds a hard eligibility rule: only pages where a shopper can purchase a product are eligible for merchant listing experiences, not pages with links to other sites that sell the product.
The recommended properties are where product questions are answered: aggregateRating, review, brand.name, sku, gtin, color, material and size, plus availability, itemCondition, shippingDetails and hasMerchantReturnPolicy on the offer. Review schema was rarer than Product schema in the sample: 4 of 47 storefronts (8.5%) carried it, against 9 of 47 (19.1%) with Product schema.
Step 3: Write product pages for constraint questions
Shoppers type constraints into an assistant rather than head keywords. A question such as “best air purifier for haze and allergies” can only be matched by a page that states the constraint in words a model can lift: coverage area, filter class, room size, the exact price and the return window.
Put the answer in the first sentence under each heading, because that is the sentence engines copy. State price and availability in the page text as well as in schema and feed, and keep all three in agreement. When feed and page disagree, an engine holds two versions of one product and cannot tell which is current.
Step 4: Feed the Shopping Graph that Gemini and Google AI Mode read
The Google path is a feed, not a page. Google’s Shopping Graph holds more than 50 billion product listings and refreshes more than 2 billion of them every hour, and free listings in Merchant Center place a product into Search, the Shopping tab, Images, YouTube, Lens and Gemini. Google’s free-listing rules require a return policy on the website, shipping settings or a shipping attribute and eight product attributes: ID, title, link, image link, price, description, availability and condition.
The checkout layer is separate again. Google’s Universal Commerce Protocol, an open standard for agentic commerce, puts a checkout button on eligible product listings in AI Mode and on Gemini and applies to products eligible in the United States, Canada and Australia. The merchant remains the seller of record. In Canlah AI’s agent-readiness sample, 25 of 25 UCP manifests found in August 2026 were Shopify-issued, so the protocol layer arrived through the platform rather than through any brand’s own work.
Step 5: Get the catalog into ChatGPT and Perplexity
ChatGPT reads a product feed with nine required fields: item_id, title, description, url, brand, seller_name, image_url, availability and price. OpenAI’s merchants page states that Shopify and Etsy catalogs are already integrated and need no application, while other merchants apply for direct feed access. Variants, identifiers and review aggregates are covered in the companion guide on ranking in ChatGPT Shopping.
Perplexity’s Merchant Program costs nothing to join, according to Shopify’s Perplexity Shopping guide reviewed on September 23, 2026. The guide states that Perplexity sources product data through integrations including Shopify and that shoppers can buy from participating retailers without leaving Perplexity, with PayPal handling checkout. Confirm the current feed format and onboarding route with Perplexity directly, because that detail was not found on the page reviewed here.
Step 6: Make category pages liftable, not decorative
Category pages produced 21% of ChatGPT’s ecommerce citations, second only to product pages at 54%, so a category page published as a bare product grid gives up the second-largest source type. Write the category page as the comparison a shopper asked for. Open with a short answer naming the best pick for each common constraint, add a table of three to six products with the headers a buyer judges on and give a dated exact price in every row.
Canlah AI’s cited-page study found that headings written as “{Product}: best for {scenario}” are copied close to verbatim by Google AI Mode, and that comparison pages without a verdict line and an exact price are skipped even when the domain is strong.
Step 7: Earn the off-site evidence the engines quote
Off-site sources supplied 15% of Google AI Mode’s ecommerce citations through YouTube alone, and none of that evidence sits on the store’s own domain. OpenAI states that ChatGPT’s product review summaries are generated from reviews on public websites and are not verified by OpenAI. Retailer listings, review publishers, roundups and video demonstrations all belong to this source layer.
Ask verified buyers to review where shoppers already read reviews and keep specifications consistent on every retailer listing that carries your item. Never seed synthetic reviews. Where buyers research in DeepSeek, Qwen or Doubao, the source graph changes again and needs its own panel.
Step 8: Track a frozen panel and keep the raw evidence
Write 20 to 40 questions the way buyers type them and keep most of them free of your brand name. Split them across category, constraint and brand questions. Run the panel on a fixed schedule against every engine in scope, and store the answer text, the cited URLs, the product tiles, the merchant row and the prices shown.
Canlah AI keeps those records as timestamped evidence so a later run is compared against the same prompts, not against a remembered impression. Report per engine and per question type, because a gain in Google Shopping cards is not a gain in ChatGPT.
Store-wide ecommerce GEO framework
A store-wide ecommerce GEO framework records eight layers, from crawl access to measurement, and ties each layer to one decision.
What to record at each layer of a store, and the decision it supports, reviewed September 23, 2026.
| Layer | What to record | Decision it supports |
|---|---|---|
| Crawl access | OAI-SearchBot, Googlebot and Google-Extended status in robots.txt; WAF challenges by user agent | Whether any later work can be seen at all |
| Product schema | Share of products with name, image, price and priceCurrency | Which page templates to fix first |
| Page copy | Constraint attributes in text; price parity with feed and schema | Which questions the catalog can match |
| Google feed | Free-listing eligibility; return policy and shipping settings | Whether products can enter Gemini and AI Mode |
| ChatGPT and Perplexity feeds | Route per platform; rows accepted and rejected by reason | Whether the catalog reaches those surfaces |
| Category pages | Presence of a verdict line, comparison table and dated exact prices | Which category questions to write for next |
| Off-site evidence | Review counts per public site; retailer listings; video coverage | Where to ask for reviews and coverage |
| Measurement | Mention, tile position, merchant row, cited URLs and prices per question | What to change, and what to leave alone |
Source: OpenAI and Google public documentation reviewed September 23, 2026, and the Canlah AI citation study of September 2026. The layers are Canlah AI’s working order, not a ranking formula published by any engine.
How to interpret five common results
Five results cover most first ecommerce GEO checks, and each one points to a single broken layer rather than to a need for more content.
Absent from every engine. Start at access and schema rather than content. A store that blocks OAI-SearchBot, ships product pages without an offer price or has never opened a Merchant Center account is invisible for structural reasons no amount of publishing will fix.
Visible in ChatGPT, absent from Google AI Mode shopping. The page layer works and the feed layer does not. Check free-listing eligibility in Merchant Center, the return policy on the site and the shipping settings before rewriting any page.
Visible in Google Shopping cards, absent from ChatGPT answers. The feed works and the pages do not. Check crawl access first, then whether the pages state the constraints the question contains.
Shown with the wrong price, stock or shipping. Fix the disagreeing source rather than adding exposure. Find which of the feed, the page and the retailer listings is stale, then raise its refresh cadence to match how often the price really moves.
Recommended, but the click goes to a reseller. The recommendation works and the merchant layer does not. Compare your price, stock, shipping and review count with the first-listed merchant, and make the maker identity consistent across feed, page and Organization schema.
What an ecommerce GEO check cannot prove
An ecommerce GEO check cannot prove four things, and each of them needs repeated observation rather than one run.
- It cannot prove a stable product position, because answers change with question wording, account state, location and time.
- It cannot show that a feed or schema change caused a placement change, without repeated observations on a fixed timeline.
- It cannot guarantee future tiles or citations after an engine changes its shopping experience or its merchant ranking.
- It cannot reproduce a signed-in shopper’s view when evidence was sampled through an API rather than a consumer interface.
When to move from spot checks to ongoing monitoring
Ongoing monitoring is justified when any one of five conditions holds: AI referrals already appear in analytics, the catalog runs to hundreds of variants, resellers compete for the merchant row on your own products, price and stock change daily or a feed and review program needs before-and-after evidence tied to the same questions. Stores selling into more than one language need it sooner, because the source graph changes with the question.
A store with a few dozen products may not need a platform. A documented monthly panel of 20 to 40 questions, stored with its raw answers, can be enough until products, engines or markets make manual checks unreliable. Ask any provider, Canlah AI included, to show raw answer records for your own product questions before signing, and verify engine coverage, sample size and interface yourself rather than trusting this guide.
Frequently asked questions
What is GEO for ecommerce?
GEO for ecommerce is generative engine optimization applied to an online store: measuring how AI engines answer product and category questions, then fixing the product pages, structured data, merchant feeds, category content and off-site reviews behind those answers. It differs from brand-level GEO because much of the work sits in a feed rather than on a page, and because the unit of visibility is a product.
Is AEO for ecommerce different from GEO for ecommerce?
No. AEO for ecommerce and GEO for ecommerce describe the same store-side work under two labels, and vendors use them interchangeably. AEO wording leans on answer formatting, while GEO wording usually includes the feed and off-site layers. Read a provider’s written scope rather than its acronym.
How do I optimise product pages for AI search?
To optimise product pages for AI search, state the constraints shoppers ask about in the page text and add merchant listing structured data with name, image, price and currency as a minimum. Keep price and availability identical across page, schema and feed, and open each heading with the answer rather than a preamble. Then confirm that OAI-SearchBot and Googlebot can crawl the page.
How do I get my products into Gemini shopping results?
Products reach Gemini shopping results through Google’s Shopping Graph, so the route is Merchant Center rather than the website alone. Meet the free-listing requirements, which include a return policy on the site, shipping settings and the required product attributes. Keep the feed fresh against a graph that refreshes more than 2 billion listings an hour, and check Universal Commerce Protocol eligibility if you want a checkout button on Gemini.
Does AI SEO for ecommerce replace conventional ecommerce SEO?
No. AI SEO for ecommerce does not replace conventional ecommerce SEO, because product schema, crawlability, canonical handling and category architecture serve both channels and the feed work is additive. The difference is measurement: an AI visibility panel records mentions, tiles, merchant rows and cited sources, which a rank tracker does not.
Does Canlah AI do GEO for ecommerce stores?
Yes. Canlah AI measures product and category questions across ChatGPT, Gemini, Google AI Overviews and Google AI Mode (Starter covers ChatGPT, Growth adds Google AI Overviews and Full adds Google AI Mode and Gemini), with Perplexity and Chinese engines not part of standard tracking, and executes the schema, page and off-site work behind the result. It does not publish a self-serve SKU-level shopping dashboard, so a merchandiser who needs daily per-SKU tile positions across thousands of products should add a dedicated tracker. Confirm engines, buyer intents, cadence and interface before signing.
Method and sources
This guide was written on September 23, 2026 from OpenAI’s and Google’s public documentation, the primary source for how each engine reads product data, and from Canlah AI’s own probe archives for the citation shares. No vendor dashboard was tested through a paid account and no client outcome data is used. The ecommerce leg rests on 95 citations from 12 questions, enough to show the split between two engines and not enough to state a market rate.
- Google Search Central, merchant listing structured data. Required and recommended properties; purchase-page eligibility.
- Google Merchant Center Help, free listings for products. Surfaces, return-policy and shipping rules, required attributes.
- Google Merchant Center Help, about the Universal Commerce Protocol. Checkout in AI Mode and Gemini; country eligibility.
- Google, shop with AI Mode and agentic checkout. Shopping Graph size and hourly refresh volume.
- OpenAI Developers, product feed specification. The nine required feed fields.
- OpenAI, power product discovery in ChatGPT. Shopify and Etsy integration and direct feed applications.
- OpenAI Developers, crawlers. OAI-SearchBot behaviour and robots.txt timing.
- OpenAI Help, shopping with ChatGPT search. Review summaries generated from public websites.
- Shopify, Perplexity Shopping guide. Merchant Program cost and checkout, as described by a third party.
- Canlah AI, agent-readiness of 50 cross-border DTC brands. Schema, robots.txt and UCP manifest counts.
- Canlah AI citation study, September 2026. Sixty Singapore questions and 526 citations, of which the ecommerce leg supplied 95 citations across 59 domains and 19 pages read.
See where your products stand in AI answers. Run the free AI visibility audit →
Related articles
FREE WHITEPAPER
Marketing in the Agent Era
All 13 chapters public — no email wall. Includes an original dataset on the agent-readiness of 50 cross-border DTC brands.
Read it free →