Ecommerce GEO: how to get your products seen and recommended by AI assistants

A shopper asks ChatGPT for “a waterproof jacket for a rainy hiking weekend, under $150”. The answer names three products, shows a carousel with prices and stock, and sometimes a sponsored unit underneath. Your brand is either in that answer or it isn’t, and no rank tracker will tell you which.

Ecommerce GEO (generative engine optimization) covers the work that raises your odds of appearing in those answers: in ChatGPT, Google AI Overviews and AI Mode, Gemini, Perplexity, and the shopping agents that are starting to place orders. This guide explains how these engines pick a product, where SEO ends and GEO or paid placement begins, what the agentic protocols change, and where to start and how to measure.

Where your products can appear in AI answers

Each assistant mixes several placements. Telling them apart keeps you from funding the wrong lever.

ChatGPT: the answer, shopping carousels and ChatGPT Ads

In ChatGPT, a product can be named in the answer text, appear in a shopping carousel (image, price, merchant), or show up in a sponsored unit below the answer. OpenAI describes the carousel as an organic result, selected independently of advertising.

Ads are the third lever. After a US pilot in February 2026, they reached 31 European markets on 24 August 2026, shown to Free and Go users only, with contextual targeting at launch in Europe. Since 31 August 2026, OpenAI’s Ads Manager has been self-serve across those markets, including product feed campaigns. We cover the setup in our article on OpenAI Ads Manager and product feeds. Buying an ad unit does not put your product in the answer text.

Google: AI Overviews, AI Mode and Gemini

AI Overviews show a summary at the top of some search results. AI Mode offers a conversational search inside Google, and the Gemini app answers shopping questions with product cards. Google powers these experiences with its Shopping Graph, which draws on Merchant Center feeds among other sources.

In May 2026, Google added six optional conversational attributes to Merchant Center, aimed at AI Mode, Gemini and Business Agent. We walk through each field in our guide to Merchant Center conversational attributes.

Perplexity and shopping agents

Perplexity cites its sources and shows product cards on shopping queries. Shopping agents go a step further: they compare, build a cart and, in some markets, complete the purchase for the user. That is where the UCP and ACP protocols come in, covered below.

How AI engines pick a product

No vendor publishes its recommendation logic. Public studies and our own audits point to four mechanisms.

They query existing product indexes

For a shopping question, the assistant generates sub-queries and searches existing indexes. In March 2026, a Peec AI study reported by Search Engine Land analyzed more than 43,000 products from ChatGPT shopping carousels: 83% matched Google Shopping organic results. Your Merchant Center feed works for ChatGPT too. We break down the study in our article on ChatGPT carousels and Google Shopping data.

OpenAI also accepts product feeds submitted by merchants, and Google reads Merchant Center feeds for its own AI experiences. Either way, the feed is the way in.

They read structured product data

Once a product is found, the model has to understand it: category, material, use case, compatibility, size, price, stock. It finds those facts in your feed and in the schema.org Product markup on your page. A title like “Neve jacket orange” says nothing to a question about rain. A title and attributes that carry “waterproof”, “breathable” and “hiking” give the answer something to work with.

They lean on third-party sources

Models repeat what others say about you. Omniscient Digital classified more than 23,000 sources cited on branded queries: 57% were reviews, listicles, forums, social posts and case studies, against 12% for product pages. Verified reviews, a place in your category’s comparison articles and consistent product claims count as much as your own pages.

The flip side is documented. A fictional deodorant brand was recommended first by ChatGPT on two long-tail queries, after a month online and a three-page website. We tell that story in the fake brand that ranked first in ChatGPT. When your own data is thin, the model fills the gaps with other people’s.

They drop what they cannot verify

An agent that recommends a product stakes its credibility on price and availability. If that information is missing from the served HTML, contradicts the feed or only appears after JavaScript runs, the product becomes less likely to be cited. The 2026 edition of the Agentic Commerce Index, our benchmark, measured it on 645 product pages from 194 French brands: price was machine-readable on only 42% of pages, availability on 39%.

SEO vs GEO vs paid AI placements

SEO remains the base layer: assistants rely in part on search indexes to choose their sources. GEO (some teams say AEO, answer engine optimization) aims for citation and recommendation inside the answer. Paid placements buy a unit next to that answer. We go deeper on the first distinction in SEO vs GEO for ecommerce.

All three run on the same input: product data. An enriched feed improves the relevance of a ChatGPT Ads campaign and your odds of an organic citation. A paid unit sitting under an answer that recommends a competitor, though, creates a mismatch that budget won’t fix.

Agentic commerce: what UCP and ACP change

Agentic commerce describes journeys where an AI agent discovers, compares and buys on the user’s behalf. Two protocols frame the topic. Below, we separate what has been announced from what we infer.

Google UCP (Universal Commerce Protocol)

Google introduced UCP in January 2026. It is an open standard, built with Shopify, Etsy, Wayfair, Target and Walmart, that gives agents and merchants a shared language: merchant capability discovery, cart, discounts, payment and order tracking. It works alongside AP2 (Agent Payments Protocol) for payment authorizations given to an agent. The merchant stays the merchant of record. In May 2026, Google announced Universal Cart, a cart shared across Search, Gemini, YouTube and Gmail. UCP-based checkout is rolling out in the US first.

OpenAI ACP (Agentic Commerce Protocol)

OpenAI and Stripe released ACP in September 2025, alongside Instant Checkout in ChatGPT. In March 2026, OpenAI moved Instant Checkout toward apps and refocused ChatGPT on product discovery, sending shoppers to the merchant’s site to buy. ACP remains the open specification, maintained by OpenAI and Stripe, for agent-driven transactions in ChatGPT.

What we take from it

For a European retailer, agentic checkout is still limited today. Agents reading your catalog is already happening: ChatGPT carousels, AI Mode, Gemini. Our reading: the best return comes from making your catalog readable and consistent now, then plugging into the protocols when they reach your market. Our UCP and ACP checklist covers the technical workstreams.

What to fix first in your product data

Across the catalogs we audit, the order of priority rarely changes. Start with your top products (by traffic or margin), then widen.

1. Synced price and availability. The same values in the feed, on the page and in the JSON-LD, refreshed at least daily. A feed-to-page mismatch is a common cause of Merchant Center disapprovals, and an agent will drop a product whose price it cannot confirm.
2. Identifiers and category. GTIN, brand, MPN, google_product_category and product_type tie the product to a known entity and a comparable category.
3. Descriptive titles. Product type, brand, key attribute, material, use case. The title should answer a buying question, without superlatives or keyword stuffing.
4. Use-case attributes. Material, season, audience, compatibility, dimensions, care. These are the constraints shoppers put in a prompt.
5. Pages readable without JavaScript. Price, stock, description and Product markup in the served HTML, with aggregate rating and review count.
6. Conversational attributes. question_and_answer, related_product and document_link through a supplemental feed, once the basics are clean.

If you are starting from scratch, our complete Google Merchant Center guide covers account and feed setup.

Conversational attributes: the Merchant Center layer built for AI answers

Once the basics are clean, conversational attributes are the Merchant Center work most directly tied to GEO. Google launched six of them at Google Marketing Live in May 2026. The classic feed was built to match a title to a short query. AI Mode, Gemini and Business Agent field something else: a buying question loaded with constraints, such as “a variable-temperature kettle for green tea, with a manual in English”. These fields give them facts to answer with.

• question_and_answer: up to 30 question and answer pairs per product, factual, with no prices, promotions or delivery times.
• related_product: typed relationships to other products in the feed (accessory, part, same range), 30 at most.
• document_link: up to five public PDFs per product, such as a manual, assembly guide or spec sheet.
• item_group_title: the name of the product family that groups the variants.
• variant_option: the axes a group varies on, such as color, size or capacity.
• popularity_rank: a popularity percentile within your own catalog, calculated from sales or orders.

Google documents three things. The attributes are optional, available in every country, and supplying them has no effect on offer approval. UCP checkout is rolling out in the US first, but a European merchant can fill these fields today. And Google does not present them as a Shopping ranking lever: treat them as a comprehension layer for AI experiences, with no documented effect on ranking.

One rule saves wasted effort. Don’t repeat in these fields what already sits in description, product_highlight or product_detail, and don’t send Q&A whose content is already in a PDF listed under document_link. Google recommends a supplemental feed, joined on id, so your primary feed stays untouched. Our article on Merchant Center conversational attributes goes through each field with formats and examples.

Beyond Google, here is our reading. Nothing suggests ChatGPT or Perplexity read these Merchant Center fields. They do read your product pages. The same questions and answers, published in plain text on the page, give those assistants content they can extract, phrased the way your customers write prompts. Write them once, from support tickets, manuals and reviews, then carry the same facts into the feed and onto the page. One piece of groundwork then serves AI Mode and the other assistants.

Automating feed optimization without losing control

A question we see phrased as a prompt: what automation setup uses AI to optimize a product feed for a merchant listing platform and reduce disapprovals? The setup that works has three layers. Deterministic rules fix the mechanical issues (formats, units, categories, recoverable missing fields). An AI model enriches what needs understanding (titles, use-case attributes, Q&A) from your catalog and pages. Your team approves the output before it goes live, through a supplemental feed, without touching the source feed.

That is the job of Feed Enrich, the Dataïads product feed optimization tool: it audits the catalog, applies transformations, generates missing attributes and publishes the enriched feed to Google Merchant Center, Meta and ChatGPT Ads. To see where you stand, the free feed audit scores your catalog on about 40 criteria, including the conversational layer.

Images and visuals for AI engines

Product search is turning multimodal. Google Lens identifies an item from a photo and pulls prices, reviews and sellers from the Shopping Graph. Shopping ads appear one by one in the Images tab on mobile, and matching a photo to an offer relies on the structured data in your feed. Since April 2025, Google has also opened AI Mode to questions that start from an image, through Lens. We map these placements in our article on product data beyond Shopping ads.

Answers cite text, images supply the proof

Citation studies show the other side. Out of 5,772 citations collected across six models and ten categories, 97.5% pointed to text. YouTube is the exception because models read its titles, descriptions and transcripts. We dig into those numbers in why answer engines cite text, almost never your videos.

Our interpretation: images matter for visual recognition and for the shopper choosing from a carousel, while text decides the citation. A polished visual that nothing on the page describes leaves a written answer nothing to quote.

What to provide

• Several angles and an in-use shot. A compliant packshot as the main image, then details, back view, scale and lifestyle photos through additional_image_link. When a shopper asks for a bag “big enough for a 16-inch laptop”, a photo of the bag in use answers better than a white background.
• The same image in the feed and on the page. The feed image should show the variant being sold and match the one on the product page. A mismatch muddies the link between offer and page.
• Descriptive alt text. Product type, color, material and visible use, in one sentence. It is the readable version of the image for a model that processes text.
• The image property in Product markup. List the main image URLs in your JSON-LD, served in the initial HTML alongside price and stock.

AI-generated visuals: fidelity and transparency

Generating lifestyle scenes from a packshot saves weeks on a large catalog. Two frameworks apply. Google asks for AI-generated images to carry IPTC metadata that says so, such as TrainedAlgorithmicMedia. In the EU, Article 50 of the AI Act has required machine-detectable marking of generated images since 2 August 2026, with a grace period to 2 December 2026 for systems already on the market; basic retouching such as background removal or upscaling is out of scope. Our article on the EU AI Act and AI-generated ecommerce visuals covers the obligations.

Fidelity is our position more than a written rule: a generated visual should show the product as sold, including color, proportions and what comes in the box. A shopper who receives something other than the picture sends it back and leaves a review, and reviews are among the sources assistants quote.

Feed Enrich, the Dataïads product feed optimization tool, handles this visual layer too. It adapts images to the formats some placements require, such as square for Demand Gen, and enriches the visuals published through the supplemental feed, with your team approving them before they go live.

How to measure AI visibility

No console gives you a rank in ChatGPT today. Measurement relies on prompt panels you build and rerun.

Build a prompt panel

Start from the questions your customers ask: support tickets, on-site search, long queries in Search Console. Group them by intent (discovery, comparison, budget constraint, alternative to a named brand). Run each prompt several times per assistant: answers vary from one run to the next, and a single run proves nothing. Test with and without web browsing where the assistant allows it.

Track a handful of metrics

• Citation rate: share of answers where your brand or product appears.
• Share of voice: your presence relative to the competitors named on the same panel.
• Carousel presence: how many of your SKUs show up in shopping carousels.
• Cited sources: the domains the assistant relies on when it cites you, or a competitor.
• Stability: how consistently you are cited from one run to the next.

The intents where competitors get cited and you don’t become your roadmap: content to create, attributes to add, reviews to collect.

Benchmark with the Agentic Commerce Index

The Agentic Commerce Index measures the other side: how well agents can technically read a brand’s product pages. The 2026 edition gave 194 French brands an average score of 48 out of 100, below the readiness threshold of 65. Only 44% of product pages carried usable schema.org Product markup, and 15% of sites had an llms.txt file. Comparing your score with your sector tells you whether your gap sits on the demand side (you are not cited) or the supply side (your page is hard to read).

Action plan: the ecommerce GEO checklist

1. Build a panel of 50 to 200 prompts by intent and record your baseline on ChatGPT, Gemini, AI Mode and Perplexity.
2. Audit your Merchant Center feed: identifiers, titles, attributes, price, stock, disapprovals.
3. Align feed, page and JSON-LD, and make pages readable without JavaScript.
4. Enrich use-case attributes, then add conversational attributes through a supplemental feed.
5. Strengthen third-party proof: verified reviews, aggregate ratings, presence in cited comparison articles.
6. Run ChatGPT Ads and Google ads from the same enriched feed, and measure paid separately.
7. Rerun the panel every month and point new fixes at the intents where you are still missing.

How Dataïads supports ecommerce GEO

The Dataïads Agentic GEO offer connects measurement and activation. Our audit and monitoring tool, in alpha, combines two views: visibility in LLMs (real prompts by intent, several runs, score, share of voice, untapped intents) and a technical and semantic audit of your product pages as an agent reads them. Fixes then go through product data: Feed Enrich for the feed, structured data and pages for consistency, then a new measurement.

Run an AI visibility audit to find the intents where your products are missing from AI answers, and why.

FAQ

How do I get ChatGPT to recommend my products?

Start with your Google Merchant Center feed and your organic ranking in Google Shopping, the source of most products in ChatGPT carousels. Add titles and attributes that describe the use case, serve price and stock in your page HTML, and collect verified reviews. Then track your presence on a prompt panel.

What is the difference between SEO and GEO for an ecommerce site?

SEO ranks a page in a list of results. GEO gets a product cited and recommended in a generated answer. They overlap, since assistants rely in part on search indexes, but GEO depends more on structured product data and third-party reviews.

Do ChatGPT Ads improve my visibility in ChatGPT answers?

No, according to OpenAI: the ad systems are separate from the model that writes the answer, and the sponsored unit appears below it. Ads only reach Free and Go users. Citations in the answer text and carousel placements stay organic.

Which product feed attributes matter most for Google AI Mode and Gemini?

The basics first: descriptive title, GTIN, brand, category, price, availability, images. Then use-case attributes (material, audience, compatibility) and the six Merchant Center conversational attributes, including question_and_answer and related_product, sent through a supplemental feed.

Should I implement UCP or ACP now?

In Europe, agentic checkout is still limited. Your priority is a readable, consistent catalog, which already pays off in carousels and AI Mode. Plan the technical integration with your ecommerce platform, which often handles part of it, for when the protocol reaches your market.

How can I measure my brand’s visibility in ChatGPT and Perplexity?

Build a prompt panel by intent, run each prompt several times per assistant, then track citation rate, share of voice against competitors, carousel presence and cited sources. Rerun the panel every month to see the effect of your fixes.

What AI tool can optimize a product feed and reduce Merchant Center disapprovals?

A product feed optimization tool that combines automated rules with AI enrichment, human approval and publication through a supplemental feed. Feed Enrich works that way: it fixes missing or malformed fields, enriches titles and attributes, and publishes to Merchant Center, Meta and ChatGPT Ads without changing your source feed.

Do Merchant Center conversational attributes help my products rank in AI Mode?

Google documents no ranking effect. These six optional fields, available in every country, give AI Mode, Gemini and Business Agent facts to answer buying questions with: Q&A, related products, PDF manuals, product families, variants and internal popularity. They do not affect offer approval. Start with question_and_answer on your top products, through a supplemental feed, without copying your description.

Which product images help AI assistants recognize and recommend my products?

A compliant main image, several angles and at least one in-use photo, identical in the feed and on the page. Add descriptive alt text and list the images in your Product markup. Written answers mostly cite text, so every visual needs attributes and a description that say what it shows. An AI-generated visual should stay true to the product and carry the marking Google and the EU AI Act require.