How Do AI Assistants Choose Which Products to Recommend?

Three things decide whether an AI assistant recommends your product, and none of them is a writing trick. First, the product has to be retrievable: in the feed the assistant reads, ranked in Google Shopping, and on a page its crawler can open. Second, the numbers the assistant can see have to be good: price, star rating and review count explain most of the choice once a product is in the candidate set. Third, someone other than you has to be saying it: the review sites, retailers and videos each assistant cites for your category. What does not help, in every controlled test we found, is the familiar AI SEO list: llms.txt files, "AI-optimized" descriptions, extra schema, FAQ padding and chunked pages. Below is the evidence for each claim, and what a Shopify brand can do about it.

Start with scale, because the surveys and the server logs tell different stories. Surveys put AI use for product research at 26 to 45 percent of consumers, depending on who asks (Bain, IBM and NRF). Measured referrals are far smaller. Similarweb puts generative AI at about 0.4 percent of retail ecommerce visits, and under 1 percent of total traffic even for retailers where it is a quarter of all referrals (Similarweb).

Small, growing, and worth more per visit

The base is small but the trend and the quality are not. Adobe's retail panel had AI referrals up 62 percent year over year in July 2026, after four-digit growth through 2025, and those visitors converted 60 percent better than other traffic for the eleventh month running (Adobe via Digital Commerce 360). Shopify reported AI-driven visits to its stores up about 8x and orders from AI searches up about 13x in its first quarter of 2026 (PYMNTS). And the click is not the whole value. Similarweb found that when an AI assistant names a brand, that brand's 7-day visit rate runs two to four times its rival's, because most shoppers who research in AI still go on to search or type the brand in (Similarweb via PPC Land). The honest frame for a merchant: a sub-1 percent channel that converts well, grows fast, and influences more than it refers.

The assistants do not share a pipeline, and the difference changes what you should work on. If your store is on Shopify, your products are already syndicated to ChatGPT, Google AI Mode and Gemini, Microsoft Copilot and Meta through Shopify Catalog, which is on by default for eligible stores and cannot be opted out of (Shopify). We covered the setup side in AI SEO for Shopify: what Shopify handles and what you need to do. This article is about what happens after you are in.

ChatGPT reads feeds and product pages

ChatGPT has two product surfaces. Its product carousels are built from shopping-style searches, and 83 percent of carousel products match Google Shopping's organic top 40 for the same query (Peec AI). Its buy offers come mostly from crawled product pages, about 88 percent in Profound's sample of a million offers, but when a merchant feed supplies the offer it takes the first slot almost every time, and the feed share of retrievals rose from 4 percent to around 20 percent in eight months (Profound). The text answers around those products cite editorial review sites and retailers, not Reddit: in one Shopify-store panel of 58,846 ChatGPT citations, Reddit appeared zero times. OpenAI says results are organic, that eligibility "does not guarantee display", and that among sellers of the same item it looks at availability, price, quality and whether you are the maker (OpenAI).

Google reads the Shopping Graph, YouTube and Reddit

Products in AI Overviews, AI Mode and Gemini come from Google's Shopping Graph, which is fed by Merchant Center and by Product structured data on your pages. Google's "no special schema" guidance applies to web pages; for product listings, a complete Merchant Center feed and correct Product and Offer markup are the route in (Google). In May 2026 Google added feed attributes it says are "primarily intended for use in conversational experiences such as AI Mode", including Q&A, product highlights and variant options, and a Merchant Center report that flags missing color, material and style data. For the editorial side, Google's AI surfaces lean on YouTube and Reddit, which appear in roughly 30 to 35 percent of AI Overview product answers, with brand and manufacturer pages at about 16 percent (Uberti-Bona Marin et al., 2026). Perplexity sits between the two, with its own free merchant program and a heavy Reddit habit.

Once a product is in the candidate set, the choice is mostly arithmetic. In the largest controlled experiment, Columbia and Yale researchers built a mock marketplace and let six models shop it across eight categories. Rating, price and review count drove the decisions; a tenth of a star was worth a 25 to 67 percent price premium depending on the model, and doubling the review count was worth 17 to 37 percent (Allouah et al., 2025). A second study with ten-product sets found product attributes explained 82 percent of the variance in the pick, list position 6.5 percent, and brand identity 1.2 percent. A fictional brand beat a real incumbent half the time with 0.075 more stars, a 7.3 percent lower price, or 1.6 times the reviews (Chu and Hou, 2026).

What that means for your catalog

Those are simulations, so treat the exact ratios as directional. The practical point survives: the assistant decides on the numbers it can read, so the numbers have to be present wherever it reads them. That is the Merchant Center and Catalog feed for the carousel, the server-rendered HTML of the product page for the buy offer, and the review sites it cites for the text. A 4.6-star product with 40 reviews loses to a 4.5 with 2,000 because review count is weighted on a log scale but still counts. Google will not show product ratings in Shopping surfaces until a merchant has around 50 reviews across the catalog (Google). Explicit prices and recent dates on the page helped consistently in a 252,000-trial study, and formatting changes alone did little (Vishwakarma et al., SIGIR 2026).

This is where most "AI optimization" money is wasted. The only controlled study of product descriptions specifically took 52,000 Amazon listings, rewrote each one fifteen different ways, and had a model re-rank them. A storytelling rewrite cost four rank positions. A minimalist rewrite cost nearly two. Jargon and unusual vocabulary both lost ground. The rewrites that gained were the ones that stated concrete advantages and quality facts, and an FAQ format was neutral (E-GEO, 2025). A full-pipeline study with real retrieval found shopping was the one domain where every body-text rewrite reduced citations (SAGEO Arena, 2026).

Write the page the shopper's question is looking for

The one description edit with a measured upside is matching the words the shopper used. In the marketplace experiment, front-loading "Office" in the title of an office lamp raised its share by 7 to 80 points across models, because the buyer had asked for an office lamp (Allouah et al.). The same study warns that a few edits reduced share, and a second experiment found that the payoff from description optimization falls to nothing once most competitors do it too (Chu and Hou). So the rule is: specs, materials, dimensions, compatibility, who it is for and what it is not for, in the words a shopper types, with the price and the date visible. Put the brand story below the fold. It is for the human who has already decided to click.

When an assistant answers from memory, it reaches for incumbents. In a 500,000-response study for Lightspeed, a national chain was named 63 to 70 percent of the time with web search off; turning search on dropped that to 46 to 58 percent and let small stores in (Vaer AI for Lightspeed, 2026). A 37,000-run academic audit found about half of specialist and regional brands never surfaced in any run (Jack et al., 2026). Yet concentration at the top is lower than SEO intuition suggests: category leaders took only 2 to 8 percent of recommendations across 11,573 answers, and 85 percent went to brands outside the top five (Fractl).

It is corroboration, not fame

What separates the brands that appear from the ones that do not is not size. A pre-registered study of 95 DTC brands found branded search volume had essentially no relationship with being recommended; Drunk Elephant, at 100,000 searches a month, was never mentioned for premium skincare, while Manduka and Solly Baby at a tenth of that were recommended in three quarters of runs (Digital Commerce Partners). What did track prominence was search interest and online conversation; ad spend and press volume contributed little (Malthouse et al., 2026). Across ChatGPT answers, the number of times a product was mentioned in the sources the assistant cited correlated 0.4 with its visibility (180 Marketing).

Earn the pages your category's assistant already cites

The off-site work is therefore specific. Find which sites the assistants cite for your category and get into them legitimately. For electronics that is RTINGS, Tom's Guide and TechRadar; for beauty, Allure, Byrdie and the Sephora and Ulta listings; for kitchen, Serious Eats and Good Housekeeping; for outdoor gear, OutdoorGearLab (Fractl). Seed product to reviewers, pitch the round-ups, keep your retailer listings complete, and treat YouTube as the one user-generated source that every assistant reads. Amazon is not a prerequisite: it blocked OpenAI's crawlers in November 2025 and all but vanished from ChatGPT product answers, which is an opening for independent stores there (Modern Retail). The legal line is bright. The FTC's rule banning fake and AI-generated reviews, sentiment-gated incentives and review suppression took effect in October 2024 with per-violation penalties (FTC), and Reddit has been removing coordinated brand hype with its own models since July 2026.

Each of these is sold as AI visibility work and has null or negative evidence behind it. An llms.txt file: no assistant fetches it, and Shopify already serves one anyway (Ahrefs). Schema added "for AI" beyond the Product markup Google already uses: no citation lift on any platform in a matched test of 1,885 pages (Ahrefs). FAQ padding and FAQ schema: neutral at best. Chunking pages into assistant-sized pieces: Google has asked people to stop (Google). A "best brands in our category" post that ranks you first: assistants excluded the publisher from the recommendation 69 to 83 percent of the time while recommending the competitors it listed. Buying reviews or seeding Reddit: illegal in the first case, detected and filtered in the second. Chasing in-chat checkout: OpenAI stepped back from Instant Checkout in March 2026 and ChatGPT now acts as a referrer to your own store, so the page it lands on is what matters (TechCrunch). Blocking the training crawlers, GPTBot, ClaudeBot and Google-Extended, is a policy choice with no effect on shopping visibility; the search and user agents are the ones to keep open.

The same prompt returns the same brand list less than 1 percent of the time, and a single run carries roughly 33 points of error on a brand's presence rate (SparkToro, "Don't Measure Once", 2026). Whether a brand appears at all is far steadier than where it ranks, so the only honest metric is presence rate across seven or more runs per prompt, reported over a few weeks. Any dashboard that gives you a rank in ChatGPT is reporting noise.

The free stack

Shopify's Agentic reports and the agentic referring channel dimension are the only complete record of orders that start in an assistant, because GA4 and pixels do not fire inside direct checkout. GA4's AI Assistant channel, added in May 2026, catches ChatGPT, Gemini, Copilot and Grok referrers; a custom channel group placed above Referral picks up Perplexity and the long tail. Search Console's generative AI report, Bing's AI Performance report and Merchant Center's AI insights show impressions, not clicks, which is still enough to see a product's share of voice move. Expect undercounting: app traffic and stripped referrers land in Direct, and clicks from AI Overviews look like ordinary Google organic.

Do we need to be on Amazon to show up in ChatGPT?

No. Amazon blocked OpenAI's crawlers in late 2025 and ChatGPT's retailer citations shifted to Walmart, Target, Best Buy and independent stores. Inside Google's AI surfaces Amazon still leads, so an Amazon listing helps there, not in ChatGPT.

Do reviews matter more than the star rating?

Both matter and they trade off. In the controlled studies a tenth of a star was worth more than a doubling of reviews, but a product with very few reviews is discounted heavily, and Google will not show ratings at all below about 50 reviews across your catalog. Volume first, then rating.

Should we rewrite every product description?

Only the ones that are missing facts. Add specs, use cases, compatibility, materials and the shopper's own words. Do not add story, tone or jargon; those cost rank in the only controlled test. If a page already states what the product is and who it is for, leave it.

How long until we see a change?

Feed and page fixes show up within a crawl cycle, usually days to a few weeks. Off-site mentions take months because someone else has to publish them. Measure presence rate monthly, not daily.

Is any of this different from regular SEO?

Not much. Retrieval, relevance and corroboration were always the inputs. AI answers remove the middle of the funnel where a poorly ranked page could still pick up clicks, and they weight visible numbers more heavily than a search result does.

If you want to know which assistants already recommend your products, and which of your product pages are missing the numbers they read, talk to us.