{
  "id": "A06",
  "url": "https://cartforagents.com/guides/how-ai-shopping-agents-choose-products",
  "title": "How AI shopping agents choose products: what the evidence shows",
  "description": "Separate official selection inputs, controlled experiments and vendor observations in a stage-by-stage ledger with methods and limitations.",
  "answer": "Shopping agents choose in stages: a product has to be eligible, get retrieved for the shopper's request, make the shortlist, win the seller slot and be purchasable. Platforms document their inputs (relevance to the query and the shopper's context, structured product data, reviews, price, availability, seller status), and controlled studies show that stated facts, ratings, list position, model version and manipulative text can each move a model's pick. The input you control most directly is the facts, so start there.",
  "take": "Of everything that moves an agent's pick, stated facts are the one a merchant fully controls. In our simulations, vague listings were almost never picked, while factual ones almost always won when they answered what the shopper asked and their neighbours didn't.",
  "summary": "Platforms document their inputs, labs measure effects inside sandboxes, and vendors report what they saw. Here is each kind of evidence, sorted by the stage of the purchase it actually covers.",
  "claims": [
    {
      "text": "OpenAI says a product appears in the ChatGPT shopping carousel when ChatGPT perceives it as relevant to the shopper's intent, considering the query and context such as Memory or Custom instructions. It lists structured metadata from first- and third-party providers, other third-party content, model responses generated before new search results, and OpenAI safety standards and product policies as inputs, and says a stated budget makes price more relevant.",
      "source_title": "OpenAI Help Center: Shopping with ChatGPT Search, How product results are selected",
      "source_url": "https://help.openai.com/en/articles/11128490-shopping-with-chatgpt-search",
      "source_date": "2026-10-27",
      "method": "documented",
      "limitation": "Living help page; the date is our check date. It names inputs, not weights or thresholds, and it is not a causal study."
    },
    {
      "text": "OpenAI says merchants for a product are ranked on factors like availability, price, quality and whether they are the maker or primary seller, that the first price shown typically comes from the first merchant listed and may not be the lowest, and that price or shipping changes can take time to appear.",
      "source_title": "OpenAI Help Center: Shopping with ChatGPT Search, How merchants are selected; Pricing",
      "source_url": "https://help.openai.com/en/articles/11128490-shopping-with-chatgpt-search",
      "source_date": "2026-10-27",
      "method": "documented",
      "limitation": "Describes seller ranking for one assistant. OpenAI says it expects this to evolve, including more personalized merchant rankings."
    },
    {
      "text": "To be automatically included in Shopify Catalog, a store must be on the Starter plan or higher and not in private mode, except that Agentic plan stores that have not published an online store to a paid plan can still be included through agentic storefronts. Products need a title, at least one image, a price above 0 and an identifiable product URL, and cannot be Unlisted or hidden from search engines. Shopify also reviews account standing and operating history over time.",
      "source_title": "Shopify Help Center: Requirements for being included in Shopify Catalog",
      "source_url": "https://help.shopify.com/en/manual/shopify-catalog/requirements",
      "source_date": "2026-10-27",
      "method": "documented",
      "limitation": "Admission requirements only. They do not say how an eligible product is ranked, and Shopify publishes no numeric thresholds for account standing."
    },
    {
      "text": "Shopify's search preview breaks listing quality into description word count, number of images, average rating and review count from trusted verified sources, variant and option completeness with in-stock availability, and store policy completeness. Shopify says search relevance also depends on popularity, customer engagement and brand recognition, and that agentic storefronts often re-rank Catalog results with their own logic, so the preview is a directional signal, not an exact prediction.",
      "source_title": "Shopify Help Center: Managing agentic storefronts, Reviewing your products' discoverability and ranking",
      "source_url": "https://help.shopify.com/en/manual/online-sales-channels/agentic-storefronts/agentic-home",
      "source_date": "2026-10-27",
      "method": "documented",
      "limitation": "Describes Shopify Catalog search, not the final ranking inside ChatGPT, Copilot or any other assistant. No weights are given."
    },
    {
      "text": "Shopify says ChatGPT and Microsoft Copilot find Shopify products through Shopify Catalog, Google AI Mode and Gemini through the Google & YouTube sales channel, and Meta surfaces through the Facebook and Instagram by Meta channel; that blocking AI crawlers in robots.txt affects only open-web discovery and does not stop Catalog from sending product data; and that every store has /agents.md, /llms.txt and /llms-full.txt URLs that merchants can customize with theme templates.",
      "source_title": "Shopify Help Center: Shopify Catalog and product discovery for agentic storefronts",
      "source_url": "https://help.shopify.com/en/manual/online-sales-channels/agentic-storefronts/products",
      "source_date": "2026-10-27",
      "method": "documented",
      "limitation": "Living help page checked on this date; Shopify notes Google AI Mode and Gemini are still rolling out to eligible stores. Describes routes and files, not ranking."
    },
    {
      "text": "Google says supporting links in AI Overviews and AI Mode need to be indexed and eligible for a snippet, and that no new machine-readable files, AI text files or special markup are needed to appear in these features.",
      "source_title": "Google Search Central: AI features and your website",
      "source_url": "https://developers.google.com/search/docs/appearance/ai-features",
      "source_date": "2026-10-27",
      "method": "documented",
      "limitation": "Google Search AI features only; not a product-ranking specification and not a statement about other assistants."
    },
    {
      "text": "Google Shopping says results labelled \"Top recommendations\" are selected based on relevance to the search, ratings, price and product features. Microsoft says Copilot bases shopping options on the prompt, likely engagement, available merchant data and other factors.",
      "source_title": "Google Shopping Help: Understand how shopping results are generated; Microsoft Support: Shopping with Microsoft Copilot",
      "source_url": "https://support.google.com/googleshopping/answer/15456476?hl=en",
      "source_date": "2026-10-27",
      "method": "documented",
      "limitation": "Each statement covers one named surface. Neither publishes weights. Copilot source: https://support.microsoft.com/en-us/microsoft-copilot/shopping-with-microsoft-copilot (checked 2026-10-27)."
    },
    {
      "text": "Every reports that changing one article's title and description from abstract positioning to a plain statement of its contents moved it from 2 of 240 to 240 of 240 citations in a supplied-snippet experiment, and that adding only \"(Updated 2026)\" to the original title scored 0 percent, against 2 percent before.",
      "source_title": "How to Get AI Agents to Recommend Your Articles (Every)",
      "source_url": "https://every.to/also-true-for-humans/how-to-get-ai-agents-to-recommend-your-articles",
      "source_date": "2026-10-27",
      "method": "controlled-simulation",
      "limitation": "Author-reported simulation about articles, not products; models picked from supplied snippets and never searched. The date is our check date. Title and description were changed together."
    },
    {
      "text": "In the ACES sandbox, one seller's description edit produced average selection gains of +3.66, +8.37, +14.79, +7.38, +14.89 and +0.32 percentage points across six buyer models, the last not significant, with significant increases in 33% of category and model experiments. The same study reports that randomized listing position causes large, model-specific changes in selection, and that one product's selection share moved from 45% to 77% between Claude Sonnet 4 and Opus 4.5 but from about 25% to 6% between GPT-4.1 and GPT-5.1.",
      "source_title": "What Is Your AI Agent Buying? (arXiv v3)",
      "source_url": "https://arxiv.org/html/2508.02630v3",
      "source_date": "2025-12-15",
      "method": "controlled-simulation",
      "limitation": "Controlled sandbox with small assortments. Selection shares are not market shares, and the study does not measure deployed assistants."
    },
    {
      "text": "In the E-GEO benchmark, which holds ten retrieved products fixed per query, simple rewriting improved mean rank on 2,000 held-out queries by +0.56 and +0.57 with GPT-5 and Claude rankers but averaged -0.36 and -0.03 with GPT-4o-mini and GPT-4.1, and 11 of 15 heuristic rewriting prompts performed worse than the simple prompt.",
      "source_title": "E-GEO: A Testbed for Generative Engine Optimization in E-Commerce (arXiv v2)",
      "source_url": "https://arxiv.org/html/2511.20867v2",
      "source_date": "2026-07-14",
      "method": "controlled-simulation",
      "limitation": "Post-retrieval benchmark: products are already in the candidate set. Not live ChatGPT rank, and rewriters did not see the future shopper query."
    },
    {
      "text": "ABxLab evaluated 17 models across more than 80,000 trials and reports that higher ratings increased selection by 30 to 80 percentage points in 14 of 17 models in its original condition, and that preference instructions strongly changed how models weighted attributes.",
      "source_title": "A Framework for Studying AI Agent Behavior: Evidence from Consumer Choice Experiments (arXiv v2)",
      "source_url": "https://arxiv.org/html/2509.25609v2",
      "source_date": "2026-02-24",
      "method": "controlled-simulation",
      "limitation": "Binary choices with controlled web content. A rating response is not evidence of product quality, and the human comparison sample was small."
    },
    {
      "text": "SafeGEO reports that its strongest realistic seller-source attack moved a flawed target product's top-three placement from 12.7% to 95.9% for one model (Devstral Small 2), relative to a truthful-rewrite control, and that evidence-breakdown defenses reduced harmful promotion by up to 39.2 points without restoring baseline performance.",
      "source_title": "SafeGEO: Understanding Generative Engine Optimization Risks in Recommendation Agents (arXiv v2)",
      "source_url": "https://arxiv.org/html/2606.28356v2",
      "source_date": "2026-10-27",
      "method": "controlled-simulation",
      "limitation": "Constructed recommendation cases with deliberately flawed targets; a security finding, not the prevalence of manipulation in the wild and not a tactic. The date is our access date for this preprint version."
    },
    {
      "text": "Bias Beware reports that social-proof wording often raised a product's recommendation rate or rank across several models, while scarcity and exclusivity wording often reduced it. In ACES, sponsored tags were generally penalized and platform endorsements rewarded.",
      "source_title": "Bias Beware: The Impact of Cognitive Biases on LLM-Driven Product Recommendations (EMNLP 2025)",
      "source_url": "https://aclanthology.org/2025.emnlp-main.1140/",
      "source_date": "2026-10-27",
      "method": "controlled-simulation",
      "limitation": "Controlled black-box experiments with no sales validation. Not evidence that fabricated popularity and genuine reviews are equivalent. Published in the EMNLP 2025 proceedings (November 2025); the date shown is our check date."
    },
    {
      "text": "A 2026 study of agent payment flows reports credential, cart-mismatch and descriptive-upsell attacks succeeding in 90%, 56% and 73.3% of its sample-agent experiments.",
      "source_title": "Signing the Transaction but Not the Decision (arXiv v1)",
      "source_url": "https://arxiv.org/abs/2609.11757",
      "source_date": "2026-09-10",
      "method": "controlled-simulation",
      "limitation": "Rates for specific attacks and test harnesses in a preprint; not the incidence of fraud in real checkouts."
    },
    {
      "text": "SAGEO Arena, a benchmark that restores retrieval and reranking to the evaluation, reports that earlier generative-engine optimization approaches often degrade performance in retrieval and reranking, while structural information can help.",
      "source_title": "SAGEO Arena (arXiv v1, work in progress)",
      "source_url": "https://arxiv.org/abs/2602.12187",
      "source_date": "2026-02-12",
      "method": "controlled-simulation",
      "limitation": "A work-in-progress preprint, not specific to e-commerce. It does not show that structured data improves deployed shopping rank."
    },
    {
      "text": "Ahrefs tracked 1,885 pages that added JSON-LD against about 4,000 matched controls and found +2.4% AI Mode and +2.2% ChatGPT citation differences, neither distinguishable from zero.",
      "source_title": "We Tracked 1,885 Pages Adding Schema (Ahrefs)",
      "source_url": "https://ahrefs.com/blog/schema-ai-citations/",
      "source_date": "2026-05-11",
      "method": "vendor-claim",
      "limitation": "Observational matched comparison by a vendor, not randomized. Measures page citations, not product recommendations, and does not argue for removing Product or Offer markup."
    },
    {
      "text": "Peec reports that 83% of more than 43,000 ChatGPT carousel products strongly matched Google's top-40 organic Shopping results, and, in a later trace analysis, a 50-token description cut at one scoring stage.",
      "source_title": "Peec AI: How ecommerce managers can optimize ChatGPT product rankings; ChatGPT Shopping optimization: US vs Europe",
      "source_url": "https://peec.ai/blog/how-ecommerce-managers-can-optimize-chatgpt-product-rankings-step-by-step",
      "source_date": "2026-03-17",
      "method": "vendor-claim",
      "limitation": "Vendor observations. Matching is not proof of sourcing, and a trace interpretation is not an OpenAI specification. Second source: https://peec.ai/blog/chatgpt-shopping-optimization-in-2026-us-vs.-europe (2026-10-16)."
    },
    {
      "text": "In an unpersonalized Shopify Catalog snapshot of 162 searches, the native checkout flag was false on all 1,620 first returned variants.",
      "source_title": "Cart for Agents: Receipts cited by our guides (Shopify agent-catalog snapshot, Oct 26, 2026)",
      "source_url": "https://cartforagents.com/research/agent-shopping-research#receipts-cited-by-our-guides",
      "source_date": "2026-10-26",
      "method": "observed",
      "limitation": "One unpersonalized response per search, first variant only. It does not show whether anyone could buy inside an assistant."
    },
    {
      "text": "In our supplied-snippet simulation, an invented humidifier with vague copy was picked in 3 of 180 Claude sessions and 0 of 60 on GPT-6.1-Sol, against 25% for random picking; with its facts as a spec line it was picked in 134 of 180 and 20 of 60. On GPT-6.1-Sol and two Gemini models it won every session with a detailed shopper brief that its facts answered and none with a one-line ask.",
      "source_title": "Cart for Agents: Vague vs specific product facts, simulated picks by Claude, GPT and Gemini models",
      "source_url": "https://cartforagents.com/research/vague-specific-product-facts-simulation",
      "source_date": "2026-10-27",
      "method": "controlled-simulation",
      "limitation": "Invented product, supplied listings, API models. The detailed brief was written to fit the test product. Not live ChatGPT results, visibility or sales."
    }
  ],
  "sources": [
    {
      "title": "OpenAI Help Center: Shopping with ChatGPT Search",
      "url": "https://help.openai.com/en/articles/11128490-shopping-with-chatgpt-search",
      "accessed": "2026-10-27"
    },
    {
      "title": "Shopify Help Center: Requirements for being included in Shopify Catalog",
      "url": "https://help.shopify.com/en/manual/shopify-catalog/requirements",
      "accessed": "2026-10-27"
    },
    {
      "title": "Shopify Help Center: Managing agentic storefronts",
      "url": "https://help.shopify.com/en/manual/online-sales-channels/agentic-storefronts/agentic-home",
      "accessed": "2026-10-27"
    },
    {
      "title": "Shopify Help Center: Shopify Catalog and product discovery for agentic storefronts",
      "url": "https://help.shopify.com/en/manual/online-sales-channels/agentic-storefronts/products",
      "accessed": "2026-10-27"
    },
    {
      "title": "Shopify: Global Catalog MCP",
      "url": "https://shopify.dev/docs/agents/catalog/global-catalog",
      "accessed": "2026-10-27"
    },
    {
      "title": "Google Search Central: AI features and your website",
      "url": "https://developers.google.com/search/docs/appearance/ai-features",
      "accessed": "2026-10-27"
    },
    {
      "title": "Google Shopping Help: Understand how shopping results are generated",
      "url": "https://support.google.com/googleshopping/answer/15456476?hl=en",
      "accessed": "2026-10-27"
    },
    {
      "title": "Microsoft Support: Shopping with Microsoft Copilot",
      "url": "https://support.microsoft.com/en-us/microsoft-copilot/shopping-with-microsoft-copilot",
      "accessed": "2026-10-27"
    },
    {
      "title": "How to Get AI Agents to Recommend Your Articles (Every)",
      "url": "https://every.to/also-true-for-humans/how-to-get-ai-agents-to-recommend-your-articles",
      "accessed": "2026-10-27"
    },
    {
      "title": "What Is Your AI Agent Buying? (arXiv v3, 2025-12-15)",
      "url": "https://arxiv.org/html/2508.02630v3",
      "accessed": "2026-10-27"
    },
    {
      "title": "E-GEO: A Testbed for Generative Engine Optimization in E-Commerce (arXiv v2, 2026-07-14)",
      "url": "https://arxiv.org/html/2511.20867v2",
      "accessed": "2026-10-27"
    },
    {
      "title": "A Framework for Studying AI Agent Behavior: Evidence from Consumer Choice Experiments (arXiv v2, 2026-02-24)",
      "url": "https://arxiv.org/html/2509.25609v2",
      "accessed": "2026-10-27"
    },
    {
      "title": "Bias Beware: The Impact of Cognitive Biases on LLM-Driven Product Recommendations (EMNLP 2025)",
      "url": "https://aclanthology.org/2025.emnlp-main.1140/",
      "accessed": "2026-10-27"
    },
    {
      "title": "Manipulating Large Language Models to Increase Product Visibility (arXiv v2, 2024-09-02)",
      "url": "https://arxiv.org/html/2404.07981v2",
      "accessed": "2026-10-27"
    },
    {
      "title": "SafeGEO: Understanding Generative Engine Optimization Risks in Recommendation Agents (arXiv v2)",
      "url": "https://arxiv.org/html/2606.28356v2",
      "accessed": "2026-10-27"
    },
    {
      "title": "SAGEO Arena (arXiv v1, 2026-02-12, work in progress)",
      "url": "https://arxiv.org/abs/2602.12187",
      "accessed": "2026-10-27"
    },
    {
      "title": "Signing the Transaction but Not the Decision (arXiv v1, 2026-09-10)",
      "url": "https://arxiv.org/abs/2609.11757",
      "accessed": "2026-10-27"
    },
    {
      "title": "We Tracked 1,885 Pages Adding Schema (Ahrefs, 2026-05-11)",
      "url": "https://ahrefs.com/blog/schema-ai-citations/",
      "accessed": "2026-10-27"
    },
    {
      "title": "Peec AI: How ecommerce managers can optimize ChatGPT product rankings (2026-03-17)",
      "url": "https://peec.ai/blog/how-ecommerce-managers-can-optimize-chatgpt-product-rankings-step-by-step",
      "accessed": "2026-10-27"
    },
    {
      "title": "Peec AI: ChatGPT Shopping optimization, US vs Europe (2026-10-16)",
      "url": "https://peec.ai/blog/chatgpt-shopping-optimization-in-2026-us-vs.-europe",
      "accessed": "2026-10-27"
    },
    {
      "title": "Cart for Agents: Receipts cited by our guides",
      "url": "https://cartforagents.com/research/agent-shopping-research#receipts-cited-by-our-guides",
      "accessed": "2026-10-27"
    },
    {
      "title": "Cart for Agents: Vague vs specific product facts, simulated picks by Claude, GPT and Gemini models",
      "url": "https://cartforagents.com/research/vague-specific-product-facts-simulation",
      "accessed": "2026-10-27"
    }
  ],
  "limitations": [
    "The four-grade scheme and the ledger are our proposed way to sort evidence. They are not an industry standard and not a ranking model.",
    "Official documentation describes stated inputs and requirements. None of the platforms we read publishes weights, thresholds or a complete ranking formula, and living help pages change; every doc on this page was checked on October 27, 2026.",
    "Controlled studies show what can move a model's choice inside a sandbox or a supplied candidate set. They do not measure what ChatGPT, Gemini, Copilot or Perplexity show your customers, and effects differ sharply by model and version.",
    "Vendor observations are reported by companies with commercial interests and were not independently reproduced by us.",
    "We cannot see inside any assistant. Our own catalog snapshot and simulations are one route, one evening and invented test products respectively."
  ],
  "related": [
    "https://cartforagents.com/guides/agentic-commerce",
    "https://cartforagents.com/guides/agent-ready-product-data",
    "https://cartforagents.com/guides/ai-shopping-platforms",
    "https://cartforagents.com/research/vague-specific-product-facts-simulation",
    "https://cartforagents.com/research/article-titles-agent-reading-simulation",
    "https://cartforagents.com/research/shopify-catalog-vs-chatgpt-results"
  ],
  "published": "2026-10-27",
  "updated": "2026-10-27",
  "checked": "2026-10-27",
  "kind": "guide"
}