ChatGPT Product Recommendations: How the Model Actually Chooses

Type "best running shoes for flat feet" into ChatGPT with no shopping mode toggled on, no product carousel in sight, and you'll still get two or three named brands back in plain text. That's the part almost nobody writes about. Most of what's published under "ChatGPT product recommendations" is actually about the dedicated Shopping surface, the card-based interface with images and prices. But a huge share of real purchase-influencing conversations never touch that surface at all. Someone asks a normal question, gets a normal paragraph answer, and a brand name is or isn't in it. That's the mechanism this article is about.

I'm going to walk through what's actually documented about how that plain-text recommendation gets picked, not speculation dressed up as fact. Where a claim comes from a study with a stated methodology, I'll say so and link it. Where it's directional or contested, I'll flag that too.

Two systems are doing the work, not one

ChatGPT recommending a product draws on two genuinely different mechanisms, and conflating them is where most confusion about "AI SEO" starts.

• Parametric knowledge from training. This is what the model learned during pretraining, baked into its weights, frozen at a knowledge cutoff. If a brand was written about often enough, in enough varied contexts, before that cutoff, the model has some internal association with it. This is static. No amount of publishing new content this week changes what's already in the weights of a given model version.
• Live retrieval. When a query benefits from current information, ChatGPT can run a web search and pull in real-time results, then generate its answer grounded in what it fetched, a pattern generally described as retrieval-augmented generation. This is where fresh content, an updated review, or a new comparison post can actually move the needle inside days, not model-training cycles.

A plain conversational answer usually blends both. The model's baseline sense of "which brands exist in this category" comes from training. Which two or three it actually names, in what order, and with what supporting detail, is frequently shaped by what it retrieves live. That's also why the same question asked twice, a week apart, can surface a different brand list. The underlying knowledge didn't change. The live retrieval layer did.

What retrieval actually pulls from, and why that list looks nothing like Google's top ten

This is the part with real numbers behind it, from more than one independent source, not a single vendor's marketing claim. A Similarweb-anchored citation audit covering roughly 600,000 citation events from January and February 2026 found that only about 12% of URLs ChatGPT and other AI tools cite overlap with Google's own top-ten organic results for the same queries. The other 88% comes from sources that don't rank on page one at all. If your SEO team is optimizing purely for classic ranking factors and assuming AI visibility follows automatically, that gap is the reason it often doesn't.

A separate synthesis, the 5W Citation Source Audit for Q1 2026, which pooled nine separate datasets from Similarweb, SEMrush, Ahrefs, Peec AI, Profound and others covering citations from January 2025 through April 2026, put actual numbers on the split: Wikipedia accounted for 13.15% of ChatGPT citations in the US and Reddit for 11.97%, together over a quarter of everything ChatGPT cites. Reuters came in seventh at 2.27%. Forbes was the only US business publication to crack the top 20, at 1.38%. The Wall Street Journal, the New York Times, Bloomberg and the Financial Times didn't appear in the top 20 at all. If you're a brand assuming a placement in a major business outlet is your ticket into ChatGPT's answers, that data says otherwise.

A companion study specifically on recommendation-style prompts, tracking 1.21 million citations across seven AI platforms in Q1 2026, found Reddit was the single most-cited user-generated-content domain on four of the seven platforms, at 8.1% of ChatGPT citations and peaking at 17.9% on Perplexity. Reddit is not the same publisher every quarter, either. One data point worth sitting with: the same 5W audit noted Reddit's ChatGPT citation share collapsed from roughly 60% down to about 10% within a two-week span in September 2025, without a clear public explanation. If your content strategy leans entirely on one platform's current citation weight, that swing is the risk you're carrying.

Where OpenAI's own retrieval index actually comes from

ChatGPT's web search has never run on a single, static index. Reporting tracking the underlying source mix through early 2026 found it shifting meaningfully: alignment with Bing's results reportedly dropped from around 26% to about 8%, while alignment with Google's results rose to roughly 33% over the same window, according to the AI Citation Source Index research. Read that as evidence of a hybrid setup, licensed search data blended with OpenAI's own crawling and re-ranking, rather than a single fixed pipeline. Practically, that means you can't reverse-engineer ChatGPT visibility by only optimizing for one search engine's index. Whatever underlying data sources ChatGPT is drawing from at a given moment, it's applying its own relevance and trust filtering on top, not just forwarding a search engine's ranking.

The credibility signals that actually seem to matter

A few patterns show up consistently across independent analyses of what makes a brand show up in an AI-generated recommendation, distinct from the citation-source data above.

• Being named inside someone else's list, repeatedly, across independent sources. One analysis from Onely attributes roughly 41% weight to appearing in authoritative third-party "best of" or comparison lists, more than any single other factor it tracked. The logic tracks with how retrieval-augmented generation works: if five independently written buying guides all mention your brand in the same category, that's a stronger signal than one glowing page you wrote about yourself.
• Reviews and structured comparison content, not raw brand messaging. The same analysis found online reviews contributing roughly 16% weight, and separately noted that editorial and press coverage outweighed a brand's own website as a trust signal. The model appears to treat what other people and outlets say about you as more decisive than what you say about yourself, which is a reasonable thing for a system trained to avoid regurgitating marketing copy verbatim.
• Recency, but not obsessively. The same source pegged roughly 71% of citations as coming from content published in 2023 through 2025, suggesting a real but not absolute recency bias, older evergreen content still gets cited, it's just outweighed on average by fresher material.
• A hard ceiling on how many brands make the cut. Multiple sources converge on the observation that a typical ChatGPT answer names somewhere around 3 to 4 brands, not a long list. That's a genuinely different competitive dynamic than a Google results page with ten blue links. If you're not in that short list, you're not in the conversation at all for that specific query, there's no page two.

I'd treat the exact percentages above as directional rather than gospel; they come from single-source analyses rather than peer-reviewed methodology, and OpenAI hasn't published its own ranking weights publicly for obvious competitive reasons. But the qualitative pattern, third-party validation beating self-promotion, recency mattering somewhat, and a tight cap on named brands, shows up consistently enough across separately run analyses that I'd treat it as a reasonably safe operating assumption for a content strategy, not just one agency's pitch deck.

How this is genuinely different from Instant Checkout and Shopping Research

It's worth being precise here because the terminology gets blurred constantly. ChatGPT's dedicated Shopping surface and its Shopping Research mode are a different technical pathway: a structured product-card interface, populated in part from merchant feeds submitted through OpenAI's commerce infrastructure, with its own ranking logic tuned for comparison shopping. Instant Checkout is different again, a transaction layer that only matters once a product has already been selected.

The organic conversational recommendation this article covers, the plain-text "you might want to look at X or Y" answer with no shopping UI involved, doesn't touch any merchant feed. It's pulled from the model's general knowledge and live web retrieval the same way any other factual question gets answered. That distinction matters for strategy: submitting a product feed to OpenAI's commerce program does nothing for whether your brand gets mentioned in a normal chat about, say, the best noise-canceling headphones under $200. Those are two separate visibility problems requiring two separate playbooks, and a lot of brands are currently investing entirely in the feed side while ignoring the much more common conversational side.

What actually moves the needle for conversational recommendations

Given everything above, here's where I'd actually put effort if the goal is showing up in ordinary ChatGPT conversations, not the Shopping surface specifically.

• Get named in independently written comparison and "best of" content. This is the single highest-weighted signal across the studies cited above. That means genuine outreach to bloggers, review sites and niche publications who write real comparisons, not a pile of sponsored placements that read like ads, since the model appears to weight editorial framing differently from promotional framing.
• Show up in real Reddit and forum discussions, unpaid. Given Reddit's outsized citation share, a genuine, non-astroturfed presence in relevant subreddits, answering questions honestly as a person rather than a brand account, has a plausible path to influencing what the model has seen and retrieves. Fake engagement here is also easier to detect and penalize than most brands assume.
• Build or maintain an accurate Wikipedia presence if you're eligible. Given Wikipedia's consistent double-digit citation share across multiple studies, a well-sourced, neutral company or product page there is one of the more durable levers available, though Wikipedia's own notability and conflict-of-interest rules mean this isn't available to every brand and shouldn't be gamed.
• Keep review volume and quality current on third-party platforms, not just your own site. The retrieval layer favors evidence from sources other than you. G2, Trustpilot, category-specific review sites and even product-specific Reddit threads all feed that evidence pool more directly than your own testimonials page does.
• Track your actual visibility instead of guessing. Since the same query can return different brand mentions week to week as retrieval sources shift, a one-time manual check tells you almost nothing about your trend. This is the kind of monitoring AI Peekaboo is built around, and it's also worth reading our breakdown of how to measure AI visibility if you want the fuller methodology.

A worked example of how fast this can shift

Similarweb's downstream-impact research, covering the finance, travel and beauty categories, found that brands recommended by ChatGPT were 2.5 times more likely to get a site visit within seven days compared to brands in the same competitive set that weren't recommended in a similar conversation. That's a real, measured downstream effect, not a hypothetical. It's also exactly why the citation-share swings mentioned earlier matter commercially and aren't just an academic curiosity. If Reddit's citation weight can drop from 60% to 10% in two weeks for reasons nobody's fully explained publicly, a brand whose visibility depended heavily on Reddit mentions during that window would have seen a real, unexplained traffic dip with no internal cause to point to.

The honest limits of what's known

None of the studies cited here come from OpenAI publishing its actual ranking algorithm, because that data isn't public and there's a reasonable argument OpenAI wouldn't want it to be, the same way Google has never published its full ranking formula. Everything above is inferred from large-scale citation tracking, third-party audits and OpenAI's own general statements about how retrieval and training data combine, not confirmed internal weighting. Treat the exact percentages as directionally useful rather than a precise formula you can plug numbers into, and be skeptical of anyone selling you a service that claims to have reverse-engineered the exact algorithm. Nobody outside OpenAI has.

The bottom line

ChatGPT choosing which product to mention in an ordinary conversation isn't one decision, it's a blend of what the model absorbed during training and what it pulls from the live web at answer time, filtered through a real and measurable preference for independent, third-party evidence over brand-authored content. Reddit, Wikipedia and genuine comparison content carry more weight than most marketing teams assume, and traditional press coverage carries less. None of that requires a merchant feed or a Shopping-surface integration, it's a separate, arguably harder, visibility problem than getting listed in a shopping card, and most brands aren't working it at all yet.

If you want a clear read on whether your brand is actually showing up in these ordinary ChatGPT conversations, and what's realistically fixable versus what needs a longer earned-media push, email me at filipe@aipeekaboo.com or grab 30 minutes on my calendar. AI Peekaboo tracks exactly this kind of citation and mention behavior across ChatGPT, Perplexity and Google's AI answers, and I'm happy to walk through what we're seeing in your category.