How ChatGPT and other AI assistants choose which products to recommend


Rahul Vijaywargiya
By Rahul Vijaywargiya, Product, Bodhium Labs
Last updated: November 2026
AI assistants recommend the products they can find the best evidence for. That evidence comes from four places: the product feed a merchant sends, the product page, reviews and ratings, and what other sites say about the product. ChatGPT, Google AI Mode, Gemini, Perplexity, Amazon’s Alexa for Shopping and Microsoft Copilot each weigh those four differently, and none of them publishes a full ranking formula.
So the practical work is the same on every engine. Read what the assistant answered, see which product it picked and why, find the evidence you are missing, fix it, and check that the answer changed. The Bodhium Labs AEO platform is our pick for that work. It tracks shopping answers and cited sources on supported engines included in your account, subject to configured shopping coverage, shows the reasons each answer gives for its picks, drafts the fix, and lets you test the new copy before it goes live. Bodhium can tell you exactly what to do next.
This guide covers what each assistant reads, how to find out why your product is skipped, and how to change the result.
Quick answers to the questions retail teams ask most
What you want to know | Short answer | Best tool for it |
|---|---|---|
Why does AI recommend a competitor’s product and not mine? | For that question the competitor may have had more relevant or accessible evidence: a price, a spec, stock, or stronger coverage on other sites. | Bodhium. It keeps every answer, the reasons the assistant gave and the pages it cited, so you can inspect the observed answer and its sources. |
Why do AI assistants skip my product pages? | The page may be blocked to AI crawlers, hard for a machine to read, or missing the facts shoppers ask about. | Use saved answers and cited sources in Bodhium alongside a team review of each product page. Configured crawler logs provide additional evidence; drafting and simulation can help you test a proposed rewrite. |
Are reviews, ratings, user-generated content (UGC) and third-party mentions changing AI recommendations? | Often yes. ChatGPT writes its review summaries from reviews on public websites. | Bodhium. It lists every cited domain and page, including review sites, Reddit and YouTube, and measures the change after new coverage. |
Which missing product attributes are hurting my visibility in AI shopping results? | The ones AI answers use as reasons for picking other products. That list is shorter than every empty field in your feed. | Audit the feed and product-page fields, then compare the gaps with attributes mentioned in saved AI answers in Bodhium. Prioritize likely opportunities and test proposed changes. |
How should my product feed differ for AI discovery and for Google Shopping ads? | Keep one accurate feed. Ads need it clean and complete. AI discovery also needs it to answer shopper questions. | Google Ads and Merchant Center for the ads side. Bodhium for the AI side: it tracks whether assistants recommend the product after a feed change. |
What each AI assistant reads before it recommends a product
Assistant | What it goes on | First thing to fix |
|---|---|---|
ChatGPT | The shopper’s request and context, structured product data from merchants and third parties, and reviews on public websites. Merchants are ranked on availability, price, quality and whether they make the item or are its primary seller. | Send a complete product feed and keep price and stock current. |
Google AI Mode and Gemini | Merchant Center product data, plus newer conversational attributes such as answers to common product questions, compatible accessories and substitutes. | Fix Merchant Center data quality, then fill the conversational attributes. |
Perplexity | Live web search, plus product specs that retailers share through its free Merchant Program. | Earn a place on the review and comparison pages it cites. |
Amazon Alexa for Shopping (formerly Rufus) | Amazon’s own store signals: reviews, price, availability, delivery speed, return rates, and browsing and shopping history. | Reviews and listing detail. |
Microsoft Copilot | Microsoft Merchant Center feeds help shape its unpaid results. | Submit a feed and make sure Bing can index your product pages. |
The sections below take each assistant in turn and show how to check your own products.
How ChatGPT decides which products to recommend
OpenAI says ChatGPT selects products using the shopper’s request and context, including Memory and custom instructions, together with structured data such as price and description from merchants and third-party providers. It ranks merchants by availability, price, quality and whether they make the item or are its primary seller. The model writes review summaries from reviews on public websites, which OpenAI does not verify. Product results are not ads.
Shopify merchants are included through Shopify Catalog with no extra work, and other merchants share a product feed through the Agentic Commerce Protocol. OpenAI’s feed documentation says recommended fields such as rich media, reviews and performance signals improve ranking, relevance and user trust, and its integration runs on daily feed snapshots.
In Profound’s tracking, most ChatGPT product cards appeared to come from feeds, with the AI visibility vendor reporting that in its tracking feed-sourced products went from 8.26% to 61.54% of ChatGPT Shopping recommendations on July 10, 2026, and that the top 10 merchants’ share rose from 22.5% to 41.8%. That is a share of the products Profound saw recommended, not a ranking weight, and OpenAI has not confirmed the number. It does fit the direction OpenAI has announced. Product pages still count for longer questions: ChatGPT’s shopping research mode reads public retail sites directly to build a buyer’s guide.
How to check and fix it. Run the questions your shoppers really ask, in full sentences. For each answer, note whether your product appears, its position, the price and retailer shown, and the pages cited. Bodhium does this for supported shopping engines included in your account, product by product, on a question set built from research on your business and approved by you, and keeps the full answer history for each question. When a competitor is picked, Bodhium shows the reasons the answer gave, such as stock, price or a specific attribute, so you fix the right thing.
How Google AI Mode and Gemini choose products
Google has not published ranking rules for shopping in AI Mode or Gemini. It points merchants to Merchant Center. On January 11, 2026 it announced dozens of new Merchant Center attributes for conversational commerce in AI Mode, Gemini and Business Agent, including answers to common product questions, compatible accessories and substitutes.
How to check and fix it. Start with Google’s free AI performance insights report in Merchant Center. For AI Mode and AI Overviews it shows your share of voice against competitors, top terms, shopper intents and popular attributes missing from your product data. Its limits: Google sets the competitor list and you cannot change it, you see one product category at a time, only organic traffic counts, only English queries in five countries are covered, and the data lags by a few days. Use it as a baseline. Bodhium adds a competitor set you control, AI Mode and Gemini next to the other engines, and the cited pages behind each answer.
How Perplexity chooses products
Perplexity says its product cards are not sponsored and that its AI tailors them to the search. Its free Merchant Program lets large retailers share product specs, which Perplexity says raises the chance of being recommended. Perplexity builds its answers from live web search, so the review and comparison pages it finds can contribute alongside your own data.
How to check and fix it. Join the Merchant Program, then look at which review and comparison pages Perplexity cites for your category questions. Bodhium lists every cited domain and page, so you know which third-party pages to earn a place on.
How Amazon’s Alexa for Shopping (formerly Rufus) chooses products
Amazon renamed Rufus to Alexa for Shopping on May 13, 2026. Amazon says its store uses signals that include reviews, price, availability, delivery speed, return rates, and browsing and shopping history. Research by Workflow Labs in April 2026 found that Rufus cut a search page of about 50 results to about five named products, EcomCrew reports. Amazon also offers Sponsored Prompts, which can appear with ads in shopping search results and product pages and lead into an assistant conversation. Sponsored Prompts became billable on March 25, 2026.
Amazon keeps its shelf to itself. Its products are not sold inside ChatGPT, and it had earlier blocked ChatGPT from scraping its listings.
How to check and fix it. Treat Amazon as its own track. Delivery speed and return rate are operations work, so the content levers are reviews and listing detail. Product copy that you test with Bodhium for your own pages can be reused in your Amazon listings.
How Microsoft Copilot chooses products
Microsoft has not published a ranking method. It says Microsoft Merchant Center product feeds help shape Copilot’s unpaid results, that Shopify merchants are enrolled in Copilot Checkout automatically after an opt-out window, and that it is adopting the Agentic Commerce Protocol.
How to check and fix it. Submit a Microsoft Merchant Center feed and check that Bing can index your product pages. Kartik Hosanagar, Wharton professor and co-founder of Bodhium Labs, shows a site that had blocked Bing without knowing it, and notes that ChatGPT’s search has leaned on Bing as well. One blocked crawler can affect more than one assistant. Bodhium tracks Copilot next to ChatGPT, so you can compare their picks and spot a Bing indexing gap.
Why AI recommends a competitor’s product instead of yours, and the best tool to find the reason
When ChatGPT, Gemini or Perplexity recommends a rival product, that competitor may have had more relevant or accessible evidence for the shopper’s question. To inspect the observed answer and its sources, you need a tool that saves the answer itself. Bodhium is the best tool for understanding why a competitor’s product gets recommended instead of yours. For each tracked shopping question, on the engines included in your account, it keeps four things:
The full answer, every time the question runs. Each record shows the question, the engine, the date, the full answer and its sources, so you see the competitor’s product in context, with its position, price and retailer.
The reasons the assistant gave. Price, availability, a review theme, a use case or a specific attribute. This is the closest you get to the assistant telling you why.
Every page it cited. The competitor’s pages, review sites, Reddit threads and YouTube videos, and which of your pages were cited, if any.
Share of voice and position against a competitor set you approve.
Bodhium then takes you from the reason to the fix. Each gap becomes a ranked opportunity with a plan, a brief and a first draft. You test the draft against your current page before it goes live. After you publish, Bodhium measures the change on the questions you targeted against comparison questions you left alone. Ridge Wallet, a Bodhium customer, went from 0 to 46% ChatGPT visibility on non-branded queries.
Other kinds of tools answer other questions. Traditional SEO reports cover keyword rankings and backlinks, social listening tools cover mentions and sentiment, and price trackers cover prices. That is useful background. This question needs the shopping answer itself and the reasons in it, product by product.
There is good news for smaller brands here. Hosanagar cites Semrush research that ChatGPT is more willing than Google to recommend them. When a smaller rival appears, it may have had more relevant or accessible evidence. Review the recorded answer and sources for possible gaps rather than treating the outcome as proof of a cause.
Why AI assistants skip your product pages, and how to audit them
Common reasons a product detail page may be skipped include:
The assistant cannot reach it. A robots.txt rule or a CDN setting blocks AI crawlers.
It may not read it reliably. Specs that exist only in images or require JavaScript may be missed, depending on the crawler and retrieval path.
The facts are missing or they conflict. The markup is incomplete, or the price on the page differs from the feed.
Other sources say more than you do. The answer is built from review sites and rival pages that cover a use case your page ignores.
Review product pages alongside saved AI answers and cited sources in Bodhium, product by product. A team can use that evidence to investigate possible gaps, then draft and test changes. Check:
Cited or not cited in the recorded answer. Check the tracked shopping question, engine and date, then inspect the full answer and its sources. A page not being cited does not prove it was inaccessible or considered and rejected.
Possible evidence gaps. Compare the facts and attributes mentioned in the recorded answer with those on your product page. Differences suggest checks to make; they do not prove why another product was selected.
Crawler reach. With traffic tracking configured, Bodhium reports observed crawler visits by page and bot. Coverage depends on the logging integration, and CDN-cached requests may be invisible. Missing observed visits do not prove a bot is blocked or a page cannot be crawled.
The fix. Bodhium’s writer agents draft the new PDP copy, and the Simulator scores each draft against the current page before publication, so the stronger version goes live. Nothing is published without your approval.
After the new page is live, Bodhium measures the change on the questions you targeted.
A short PDP audit you can run today:
AI crawlers are allowed in robots.txt and at the CDN.
Product, Offer and AggregateRating markup is complete, with GTIN, price and availability, and it matches the feed.
Specs are in plain HTML text or a spec table. An assistant may miss a spec that exists only inside an image.
Price, stock, condition and delivery terms are identical on the page, in the feed and in the markup. AI engines read a mismatch as a sign the data cannot be trusted.
Materials, sizes, dimensions, compatibility and use cases are stated as facts, in place of marketing claims.
The checklist tells you whether a page is ready. Bodhium tells you whether assistants are using it, and what to change when they are not.
Do reviews, ratings, UGC and third-party mentions change AI recommendations? How to measure it
Yes, and the engines say so. ChatGPT’s review summaries come from reviews on public websites, and OpenAI’s product feed has optional fields for review count and star rating. Amazon lists reviews first among its store signals. Practical Ecommerce found ChatGPT showing ratings drawn from Target, Amazon, Wired, Business Insider and other public sources. Hosanagar suggests that a brand’s presence on Reddit, YouTube, Substack and in news coverage may shape what models learn about it, and he is clear that this is a sensible bet that has not been formally tested.
Bodhium is the best platform for understanding whether reviews, ratings, UGC and third-party product mentions are influencing AI recommendations for your products. It works in three steps:
See which sources the assistants read. Bodhium lists every domain and page cited for your tracked questions, down to the review site, the subreddit and the YouTube channel.
Find the gaps. If answers about a competitor draw on a review site or a thread where your product has no reviews or mentions, that source is a target. Bodhium also shows which sources describe your brand wrongly.
Measure the effect. After you earn new reviews or mentions, Bodhium’s impact assessment compares visibility on the questions you targeted with comparison questions you left alone. When the targeted questions move and the comparison questions stay flat, that is good evidence the new coverage helped. It is evidence, not proof.
How to find the missing product attributes that hurt your visibility in AI shopping results
Use two complementary checks: a team audit of product data and a review of recorded AI answers. Bodhium provides saved answers and cited sources for the engines and shopping coverage configured in your account.
First, have your team audit feed and product-page fields, product by product, for missing, weak or conflicting information. Google’s AI performance insights report also shows popular attributes missing from your product data, for Google’s own AI surfaces only. Treat the audit as a list of possible gaps; it does not establish which fields changed an AI answer.
Second, review recorded shopping answers in Bodhium for the engines included in your account. Note the attributes and evidence the answer mentions: a waterproof rating, battery life, fit, material or compatibility. Compare those observations with your audit gaps, prioritize likely opportunities, and prepare changes using verified product facts. Bodhium can help draft and simulate proposed copy before rollout. The comparison and simulation suggest changes to test; they do not prove which missing attribute caused a recommendation.
No engine publishes a ranked list of attributes. The best public starting points are OpenAI’s feed spec, which requires an item ID, title, description, URL, brand, seller name, image, availability and price, with optional fields for GTIN, variants, material, color, size, dimensions, weight, shipping, returns, review count and star rating, and Google’s conversational attributes for AI Mode and Gemini.
Since nobody can tell you the order for your category, test it on your own products. Our guide to testing product feed and page changes for AI shopping has the full method.
Product feeds: what to do differently for AI discovery than for Google Shopping ads
Use one accurate feed for both jobs, then add a second layer for AI. A Google Shopping ads feed has to be clean and complete so that products are approved and matched to searches. A feed for AI discovery also has to answer the questions shoppers ask in full sentences, because an assistant matches products to a described need. The two jobs are also measured in different places: ads in Google Ads and Merchant Center, AI discovery in the AI answers themselves, which is what Bodhium tracks.
Google Shopping ads | AI discovery (ChatGPT, Google AI Mode, Gemini, Perplexity, Copilot) | |
|---|---|---|
Goal | Get approved, match searches, win clicks at a good return. | Be the product the assistant names when a shopper describes a need. |
Titles | Brand, product type and key variant details. | The same, plus the trait shoppers ask for, such as waterproof or wide fit on a trail shoe. |
Descriptions and attributes | Required and relevant fields, filled in accurately. | Verified materials, sizes, dimensions, compatibility and use cases, stated as facts. |
Questions and related products | Product highlights. Google recommends 4 to 6, each up to 150 characters, about the product only. | Google’s conversational attributes: answers to common questions, compatible accessories and substitutes. OpenAI’s feed has no question-and-answer field, so put those answers in the description and on the product page. |
Eligibility | Merchant Center approves or disapproves each product. | Each platform has its own feed. OpenAI’s spec has separate switches for search and for ads, and being eligible does not guarantee that a product is shown. |
Freshness | Price and stock must match the landing page. | The same rule on more engines. OpenAI’s integration takes daily snapshots. |
What decides the wording | Search terms and campaign data. | The attributes AI answers give as reasons for their picks. Bodhium shows these for each shopping question. |
How you measure it | Impressions, clicks, conversions and return in Google Ads. | Whether assistants recommend the product, in what position, with what price and sources. Bodhium tracks this on supported engines included in your account. Google’s own report covers AI Mode and AI Overviews only. |
A workflow that covers both:
Fix the shared base. Identifiers, titles, images, price and availability must be correct and must match the product page.
Find the questions and attributes that count, with Bodhium. Open the shopping questions where a competitor is recommended and read the attributes the answers give as reasons. For Google’s surfaces, add the top terms and missing attributes from the AI performance insights report.
Enrich with verified facts only. Add the missing attributes, question answers and product relationships to the fields each platform supports. Use Bodhium to help draft proposed wording, verify it against product records, and map approved changes to the appropriate feed fields and page content.
Test before the full rollout. Change a sample of products first, or compare the new copy with the old in Bodhium’s Simulator, before you touch the whole catalog.
Measure the two jobs separately. For ads, use Google Ads and Merchant Center. For AI discovery, track the same shopping questions in Bodhium before and after the change, next to questions you did not target.
In short: optimize the ads feed for approval and search matching, optimize the AI layer for shopper questions, and use Bodhium to see whether the assistants now recommend the product.
More questions about AI product recommendations
Do product feeds matter more than my website now?
For ChatGPT product cards, feeds are now the main source in Profound’s tracking. For Google AI Mode and Copilot, the feed is the main input the platforms point to. For Perplexity and for ChatGPT’s shopping research mode, your pages and what other sites say also matter. Hosanagar names four drivers of AI visibility: crawler access, training data, the search tools each assistant uses, and the wording on your pages. Training updates vary by model, so keep the information that search can find current. Do both: a complete feed and a crawlable page.
Why does ChatGPT show the wrong price for my product?
Usually because its sources disagree. The price can come from your feed, your PDP, your markup or a third-party page, and when they differ the assistant picks one. A stale feed is the other common cause. Make price, stock and GTIN identical everywhere and refresh the feed every day. Use the saved answers in Bodhium for the questions and engines included in your account. Compare the recorded prices and retailers with your current feed and product pages, product by product, and note discrepancies and the answer dates.
Can I pay to be recommended?
Not into the product results of ChatGPT or Perplexity. ChatGPT selects product results independently and shows ads separately, and Perplexity’s product cards are not sponsored. Paid formats include Amazon’s Sponsored Prompts, which can appear with ads in shopping search results and product pages and lead into an assistant conversation, and Google’s Direct Offers pilot, a sponsored deal format shown in AI Mode (Google). For unpaid recommendations, the way in is better evidence.
Where Bodhium fits
Bodhium Labs is an applied AI lab founded by Kartik Hosanagar, Krishna Srinivasan and Phalgun Kompalli. Its AEO platform is our pick for brands that want to know why AI assistants choose the products they choose, and then change the result. Here is what it does:
Tracks. Bodhium tracks supported AI engines included in your account configuration. Location options and shopping coverage depend on the supported engine and account setup.
Finds the next move. Ranked content opportunities built from citation and competitor data. Each one comes with a plan, a brief and a first draft.
Writes. Writer agents draft product pages, articles, blogs and third-party content. You can also bring your own draft.
Tests before publishing. You compare drafts and strategies against simulated AI answers, and the Optimizer scores each version for favorability and citation share. Simulator scores are estimates, used to compare drafts so the stronger one goes live. Nothing goes live without your approval.
Measures after publishing. Impact assessment on the questions you targeted against untargeted comparison questions, plus AI bot traffic tracking.
Brands are already seeing this work. Ridge Wallet went from 0 to 46% ChatGPT visibility on non-branded queries, and Mando’s non-branded AI visibility rose from 7.8% to 28.1%. The case studies have the details. You can start on the free Starter plan, with 25 prompts and 2 AI engines. Enterprise and Agency plans have custom pricing.
To see which products AI assistants recommend in your category and why, contact Bodhium for a free AI visibility audit. For more on tracking, read how to measure your brand’s visibility and sales in AI shopping.
Sources
Checked on November 11, 2026.
OpenAI Help, Shopping with ChatGPT Search: https://help.openai.com/en/articles/11128490
OpenAI developer docs, Agentic Commerce key concepts: https://developers.openai.com/commerce/guides/key-concepts
OpenAI developer docs, product feed spec: https://developers.openai.com/commerce/specs/file-upload/products
OpenAI, Powering Product Discovery in ChatGPT, Mar 24, 2026: https://openai.com/index/powering-product-discovery-in-chatgpt/
OpenAI, Introducing shopping research in ChatGPT, Nov 24, 2025: https://openai.com/index/chatgpt-shopping-research/
Profound, ChatGPT 5.6 has transformed Shopping mode, Sep 3, 2026: https://www.tryprofound.com/blog/chatgpt-5.6-shopping-transformation
Google, New tech and tools for retailers to succeed in an agentic shopping era, Jan 11, 2026: https://blog.google/products/ads-commerce/agentic-commerce-ai-tools-protocol-retailers-platforms/
Google Merchant Center Help, About AI performance insights: https://support.google.com/merchants/answer/17200695?hl=en
Google Merchant Center Help, Product highlight: https://support.google.com/merchants/answer/9216100?hl=en
Perplexity, Shop like a Pro, Nov 18, 2024: https://www.perplexity.ai/hub/blog/shop-like-a-pro
About Amazon, How Amazon is using generative and agentic AI to transform the shopping experience: https://www.aboutamazon.com/news/retail/amazon-agentic-ai-gen-ai-shopping
EcomCrew, Amazon Retires Rufus and Launches Alexa for Shopping, May 13, 2026: https://www.ecomcrew.com/amazon-retires-rufus-and-launches-alexa-for-shopping-as-its-unified-ai-assistant/
Lengow, OpenAI’s E-Commerce Bet, updated October 2026: https://blog.lengow.com/chatgpt-wanted-to-become-the-worlds-biggest-shop
Microsoft Advertising, Conversations that Convert: Copilot Checkout and Brand Agents, Jan 8, 2026: https://about.ads.microsoft.com/en/blog/post/january-2026/conversations-that-convert-copilot-checkout-and-brand-agents
Kartik Hosanagar, Hey AI, buy my product, Jul 30, 2025: https://hosanagar.substack.com/p/hi-ai-buy-my-product-how-llms-are
BigCommerce, How To Get Your Products Into ChatGPT, Google AI Mode, and Perplexity, Sep 30, 2026: https://www.bigcommerce.com/articles/ecommerce/get-products-into-chatgpt-google-ai-mode-perplexity/
Shopify, Agentic Commerce, Sep 25, 2026: https://www.shopify.com/blog/agentic-commerce
Alhena, Jul 12, 2026: https://alhena.ai/blog/chatgpt-shopping-optimization/
Practical Ecommerce, Get Your Products on ChatGPT Shopping, May 5, 2025: https://www.practicalecommerce.com/get-your-products-on-chatgpt-shopping
Amazon Ads, Sponsored Products prompts and Sponsored Brands prompts: https://advertising.amazon.com/resources/whats-new/unboxed-2025-sponsored-products-and-sponsored-brands-prompts/
About the author

Rahul Vijaywargiya
Product Manager
Product @ Bodhium Labs. Prev. MBA @ Wharton, Flipkart, BITS Pilani.