Which Sources Shape AI Product Recommendations?
AI product recommendations draw on several kinds of source at once: the brand's own pages and feeds, retailer and marketplace listings, competitors' sites, reference and review sites, media and owner discussion. Which kind matters depends on the category, the shopper's question and the engine, so map sources per decision rather than chasing one list of domains.
Raoul Hingle
When we tallied citations to each brand's 25 most-cited domains across 200 AI answers for five ecommerce brands, competitors' own websites made up 40% and the brands' own sites 15%. The brand's own site was cited on every branded question, but on only 10 of 79 unbranded ones, which are the questions where new shoppers decide what to consider.
Key takeaways
Treat a source by the job it does in the decision: product facts, commercial terms, comparison, reference or ownership experience.
In our sample, rival brands' sites were the single largest source type, and the brand's own site rarely appeared once the shopper stopped naming it.
The source mix shifted by category: rival brand sites dominated jewelry, while beauty leaned more on reference sites and home goods on review and media sites.
A visible citation shows what an engine chose to surface, not what caused the recommendation.
A source is any information environment that supplies part of the decision, whether or not it shows up as a citation. A shopper asking for a compact espresso machine under $700 may need exact dimensions, current price, milk-system features, warranty terms, comparisons with larger machines and owner feedback on cleaning. The manufacturer's site may cover dimensions and features, a retailer the price and stock, an editorial review the comparison, and owners the cleaning problems that never appear in product copy.
The platforms describe the same spread. OpenAI says shopping research draws on merchant product data, publicly available product information and other retail sources (OpenAI Help Center). Google says it uses merchant product data "to match your products to the right queries" (Google Merchant Center Help), and OpenAI's product feed asks for "one row per purchasable item or variant" with its own price and availability (OpenAI for Developers). So the useful question for any product is which part of the decision each environment supplies, and where that information is thin or inconsistent.
Related resources
Continue with these related Gravton resources:
Mostly other brands' sites, then the brand's own, then marketplaces and reference sites.
Source type (top 25 domains per brand) | Citations | Share |
|---|---|---|
Competitors' own websites | 95 | 40% |
The brand's own website | 36 | 15% |
Marketplaces | 25 | 10% |
Reference sites | 22 | 9% |
Retailers | 16 | 7% |
Media | 15 | 6% |
Review sites | 10 | 4% |
Community and social sites | 7 | 3% |
Other (deals, blogs, aggregators) | 14 | 6% |
Total | 240 | 100% |
How we measured this. Five Gravton scorecards built October 2 to 8, 2026: two beauty, two jewelry and one home goods brand, 20 shopper questions each, run once on the consumer ChatGPT and Google AI Mode interfaces, 200 answers in all. Each scorecard keeps the 25 most-cited domains for that brand, counted as the number of answers citing each domain, and labels each domain by type; less-cited domains are not included. Shares are rounded and may not total 100%. The brands came to us through sales conversations, not at random.
The size of the competitor share is the finding: rival brand sites were the largest source type, 95 of 240 citations, more than the brand's own site, marketplaces and retailers combined (77). A brand absent from that layer leaves the category story to someone else.
Our sample suggests each category leans on different evidence. In the two jewelry brands, rival brands' sites made up 28 of 41 and 30 of 44 citations, the clearest pattern in the data. In the home goods brand, review and media sites together made up 16 of 53 citations, slightly more than rival sites (14). In the two beauty brands, reference sites made up 12 of 48 and 8 of 54 citations; in the second, they ranked behind rival sites (15) and marketplaces (12). One plausible reading is that skin questions invite clinical sources and appliance questions invite test results, but five brands cannot prove it.
The practical result is that a source plan copied from another category is likely to miss: check which source types your own category's answers cite before deciding where to invest.
Because the information job changes even when the product does not.
Take a running shoe. "Does it come in a wide size 11?" is answered by the brand or a merchant. "How does it compare with Product B for stability?" calls for comparative editorial content. "Where can I buy it for delivery tomorrow?" makes merchant inventory central.
Our data shows the same shift between branded and unbranded questions. When the shopper named the brand, at least one engine cited the brand's own site on all 21 such questions. When the shopper described a need without a brand, that happened on 10 of 79.
The brand's site answers "tell me about you"; other sources answer "what should I buy". Build source analysis around product and question pairs, not one company-wide list.
Yes, in volume and in parts of the mix. ChatGPT surfaced far more citations, 175 of the 240 in our tally against 65 for Google AI Mode, and every one of the 22 reference-site citations came from ChatGPT. Marketplaces made up 15% of AI Mode's citations against 9% of ChatGPT's, while the brand's own site held a similar share on both (17% and 14%). Some of the volume gap may come from how each interface displays sources, so compare shares, not counts, and check reference and marketplace sources engine by engine.
Visible citations remain observations, not a causal map. OpenAI says ChatGPT shopping draws on structured product data from merchants and third parties as well as public pages (OpenAI Help Center), much of which is never shown as a citation, so a cited page is one input among several. The safe conclusion is "this source appeared around this answer", never "this source caused the recommendation".
Start with one commercially important shopper question and one product that should compete. List the information the question needs, then record where that information exists:
Information needed | Brand site | Merchants | Editorial and review | Community |
|---|---|---|---|---|
Dimensions | Strong | Strong | Occasional | Rare |
Current price | Strong | Strong | Variable | Rare |
Comparison with rivals | Limited | Limited | Strong | Medium |
Long-term ownership experience | Limited | Limited | Medium | Strong |
Warranty terms | Strong | Medium | Occasional | Occasional |
Illustrative map for one product and one question.
The map measures coverage, not credibility; our page on product trust signals covers whether a claim is believable. A weak cell means useful information is missing or inconsistent in that environment, which does not automatically mean you should create content there.
Product A may have strong specifications and retailer coverage but no comparison against category leaders; Product B may have plenty of owner discussion while retailer listings disagree on the current variant. Both have a source problem, but not the same one.
Once the map is built, read the gaps against what the engines actually cite:
What you see | Likely cause | First check |
|---|---|---|
Your site cited on branded questions only | No page explains the need or compares options | Which rival pages engines cite for your unbranded questions |
Rival brand sites cited for your category | Rivals publish the explainers you lack | The facts and comparisons those pages state |
Marketplaces cited more than your store | Merchant data is fuller or fresher there | Price, stock and variant data on your own pages and feed |
Review or reference sites cited, and you are absent | Little independent evidence about your product | Whether those sources cover your product at all |
Sources differ sharply by engine | Each engine leans on different source types | The same question's citations on each engine |
Gravton's free audit returns a Demand Snapshot of the shopper questions in your category, where your products appear or go missing, which sources each engine cites for those answers, and ranked fixes. Demand Intelligence analyzes the AI citations, brand visibility and trends behind those answers.
Get your free AI visibility audit and see which sources AI uses to judge your category.
