Computational Marketing

The shortlist is the new shelf: what a 56-person ChatGPT shopping study means for brands that want to be on the list

A small ChatGPT study shows buyers choose from the AI's list.

· 8 min read
Cover illustration for “The shortlist is the new shelf: what a 56-person ChatGPT shopping study means for brands that want to be on the list”
Features · November 13, 2026 · 8 min read · 1,794 words

The 56-person ChatGPT shopping study points to a change in where brands compete. When buyers choose from an AI-generated shortlist without looking further, that answer is the shelf they shop from. Earning a place there requires more than familiar branding or strong search rankings. It requires being retrievable, being described in ways that fit the buyer’s needs, and measuring whether that position holds.

What the 56-person ChatGPT shopping study actually measured

5W’s study examined shopping through ChatGPT, using standardized prompts. Its scope was limited to that platform. The analysis ran a large volume of golf shopping queries across two months, with a substantial share producing product recommendations. That window allowed patterns to emerge, but the design does not establish how every shopper behaves across every AI engine.

5W found a meaningful correlation of 0.57 between brand visibility and consumer selection. Callaway appeared more often across both months, while PING ranked higher on average on the shortlists it joined. The top three manufacturers also stood well apart from the rest in visibility.

How buyers actually behave when ChatGPT returns a shopping shortlist

5W found that the overwhelming majority of analyzed tasks ended without a click to an external web page. Participants read the answer and declared their finalists from it. The comparison step happened inside the conversation. A website could therefore lose the chance to explain a product before the buyer had even decided which brands deserved further attention.

What does that change about the competitive moment? It moves it ahead of the visit that brands have traditionally measured. If a shopper accepts the returned list as the candidate set, absence from that answer means absence from that particular comparison.

Brand recognition can still influence a choice within the list. Familiarity may give a buyer a reason to prefer one suggested product over another, even when it appears lower down. But recognition and inclusion perform different jobs. A known brand cannot rely on the buyer choosing it from an answer that does not name it.

Why AI shopping expands shortlists rather than just filtering them, especially in software and considered purchases

Calling it filtering assumes that buyers already have brands in mind. Often, the AI is building that set instead. Its answer can introduce a brand that the buyer had not planned to investigate, putting an unfamiliar option into direct comparison with an established one. Being well known therefore protects a brand less than it would if the system only sorted the buyer’s existing choices.

The software buyer findings give this distinction useful texture. Among buyers who chose a vendor, most said that vendor appeared in the first AI answer, while a further significant portion encountered it after follow-up prompts. The first answer has particular weight. Yet follow-ups can open another route into consideration, so presence across a conversation also has value.

For considered purchases, this makes the shortlist more than a final filter before checkout. It can shape what the buyer learns about the category and which alternatives seem relevant. The same distinction applies to consumer shopping. A golf shopper may begin with a need rather than a manufacturer, leaving the AI to supply the names against which that need gets judged.

The platform changes that rewired how ChatGPT assembles shopping recommendations

The 2026 platform changes moved ChatGPT shopping away from a web-crawl-dominant model toward one that also draws on structured product data. Merchant feeds became a direct input to whether a product could appear in a product card. This changes what information can qualify a product for presentation.

A feed supplies product information, while Business Hub gives brands a way to address how ChatGPT represents them and see a defined portion of their appearances. Neither should be confused with a promise of recommendation. OpenAI says plainly that claiming a Business Hub listing does not guarantee more visibility.

A brand may get plenty of coverage on the web and still need accurate product data for feed-integrated retrieval. Conversely, a complete feed does not settle how the broader web describes the brand.

What actually determines whether a brand gets cited — and why citations and brand mentions are not the same thing

Kevin Indig studied a large Semrush dataset across US categories and found a slightly negative correlation between the number of citations and the number of brand mentions. The most cited domain was also the most mentioned brand in only about one in five cases. The distinction is easy to overlook. A citation identifies a source used in the answer, while a brand mention identifies a name the answer discusses.

That separation challenges a familiar optimization habit. Links, domain authority and Google rankings can look like natural proxies for AI presence because they already feature in search measurement. But a page can supply evidence for an answer without its owner becoming a recommended brand.

Brand web mentions correlate far more strongly with AI Overview visibility than backlinks do. The effect of any additional mention on visibility remains unproven. CiteLens also found that Google AI Mode and Perplexity draw the large majority of their citations from Google’s conventional top results, while ChatGPT draws only around 30% from that pool. Google performance is therefore a weaker proxy for ChatGPT presence than a team might assume. The engine changes the relationship between search visibility and answer visibility.

Why most brands cannot see how often they actually appear in AI answers

Business Hub offers a meaningful step toward direct measurement, but its definition of a mention is narrower than a shopper’s experience of seeing a brand. A mention counts when ChatGPT references the brand’s website. The brand name alone does not qualify. An answer can therefore place a product in front of a buyer without that appearance entering the dashboard’s count.

Retail distribution makes the gap wider. When a recommendation links to a retailer rather than the brand’s own domain, the retailer receives the counted mention. A brand sold mainly through those channels could have a substantial presence in answers that its own dashboard does not fully reflect.

The resulting number is useful as a floor. It is not a complete account of shortlist presence, and claiming the listing does not itself increase how often the brand is recommended. Correcting how a brand is represented and increasing how often it is recommended remain separate tasks. That distinction creates a need for independent observation of the answers themselves, including appearances with no link to the brand’s site.

How framing within the answer shapes which shortlisted brand the buyer actually picks

5W found that comparison grids held shopper attention longest in the largest share of evaluated tasks. Their structure does practical work for the buyer. A grid makes differences visible in one place, allowing someone who has delegated the comparison to choose without reconstructing it across several websites. The product’s description within that structure can be as consequential as its place in the list.

A caveat can also outweigh a prominent position. The study includes a participant who immediately dismissed a brand because of a caveat surfaced by the AI. That example does not establish that every warning produces rejection.

Positioning tags offer the positive version of that mechanism. A label such as “best for beginners” connects a product to a buyer’s stated need, giving that buyer a shortcut to selection. For brand marketers, the practical lever is product framing across web descriptions and product data. They can influence the information available to the system, though they cannot dictate its final wording or ranking.

The external signal architecture that makes a brand citable — and why a brand's own site can only do so much

Source-diversity logic limits how many citations can come from a single domain owned by the brand, so the brand needs to appear in third-party sources. Beyond a well-written product page, the brand needs independent sources that give answer engines material from which to retrieve and compare it.

Comparison and ranking pages have a particular role here. They put the choices side by side, describe how they differ and show how they fit what buyers need. That makes third-party comparisons relevant to the task the AI is performing when it assembles a shortlist. Descriptions on a brand’s own pages need to be complemented by an external record.

The practical response is to build both sides of that record. A complete, accurate merchant feed supports product card eligibility, while specific, attributable and machine-readable content gives engines usable information. More content is not automatically better. Content that lacks those traits can add more words without adding evidence that can be trusted, whether a person or a model wrote it.

In the Deana Burke case, a fake deodorant brand appeared in ChatGPT recommendations within three weeks, showing that the visibility playbook can work even when there is no real product behind it. That is also its warning. A citation without product and sales infrastructure is a vanity metric.

Brands with AI-assisted content pipelines should have a person review the content before it enters the publish queue. A fabricated statistic or misattributed claim can damage credibility with both AI systems and human readers, even if it earns attention first. Publication is the point at which that risk becomes public.

What a brand needs to track, and in which engines, to know whether its shortlist position is holding

Shortlist positions vary across runs, engines and platform changes. A single appearance cannot establish durable visibility. Only a small fraction of brands remain present across multiple consecutive runs of the same prompt, so a snapshot may capture a favorable answer. Checking once a quarter provides too sparse a sample to monitor that variability adequately.

The practical measurement unit is a consistent panel of purchase-intent prompts. Across that panel, a brand needs to track mention frequency, position when present, sentiment and framing, and the third-party sources used in citations. Each measure answers a different question. Frequency shows how often the brand appears, position shows how high it ranks, and tags or caveats show whether the answer describes it as a fit.

In his broader work on measurement, Indig found no reliable link between AI visibility and direct traffic or branded search across four brands and 632 individual correlation tests. That finding supports measuring mentions and citations directly. It warns against assuming that downstream traffic reveals how often a brand enters an answer.

Measurement also needs to extend beyond ChatGPT. At least eight GEO software platforms were available in 2026, but no single tool covered ChatGPT, Claude, Gemini and Perplexity in full with the depth of prompt-panel testing needed. The operating model is therefore to run the same purchase-intent panel across engines on a cadence, connect changes in presence and position to changes in content and feed coverage, and use Business Hub as corroborating evidence.

More in Features