ChatGPT now names 3.87 B2B tech brands per answer, down from 5.36, and picks most of them before it searches.
ChatGPT named 5.36 brands per answer on a fixed panel of non-branded B2B tech prompts in March 2026. On September 13, 2026, the same prompts returned 3.87 brands per answer, a 27.9% drop (BrightEdge AI Catalyst, published September 21, 2026). The AI brand shortlist your buyers see is getting shorter, and four studies released since late July show the list forms inside the model before it runs a single web search.
Here's what changed, why the shortlist now matters more than citation counts, and what your team should do about it this month.
The AI brand shortlist is the set of vendors an AI engine names when a buyer asks a category question without naming a brand. "Best SOC 2 compliance software for a 50-person startup" is a shortlist prompt. "Is Vanta good?" is not.
Two things moved in September. BrightEdge measured the shortlist itself shrinking: ChatGPT held the same prompts constant for six months and named 1.49 fewer brands per answer at the end. Northwestern University researchers published a paper on September 14 (arXiv 2609.16304) showing that six frontier models, with web search switched off, leave out large and established brands across whole categories. A week later, a UK study of 8,609 AI responses showed that a user's own chat history shifts which brands make the list.
Put together, the data describes a funnel with a gate at the top. If the model does not carry your brand into the answer, your website, schema and content never get a vote.
ChatGPT's B2B tech shortlist lost 27.9% of its slots in six months. BrightEdge tracked a fixed panel of non-branded US prompts on March 15 and September 13, 2026, and counted distinct brands named per response.
BrightEdge used only prompts that kept the same intent label on both dates. About 15% of prompts changed stage between March and September, so the analysts excluded them. A cross-check across all prompts, at each date's own labels, still showed a 25.1% decline. The shrink holds under both cuts.
BrightEdge also reports that AI search referral volume kept climbing over the same window. More buyers ask, and each answer names fewer vendors. Each remaining slot carries more weight than it did in spring.
Informational prompts gave up the most room on the AI brand shortlist, falling from 5.15 to 3.50 brands per answer. These are the "what is zero trust architecture" and "how does reverse ETL work" questions buyers ask months before they build a vendor list.
That stage matters for B2B because it plants the names buyers carry into later research. A brand that appears in an early explainer answer gets searched by name later. A brand that drops out of it has to win the buyer back at the comparison stage, where competition runs hottest.
Consideration prompts still name the most brands, 6.39 per answer, but they lost 22.1% of their slots. These are the "Zoho vs QuickBooks for a 10-person agency" questions. Buyers at this stage compare options, so a shorter list here means fewer vendors reach the demo request.
BrightEdge held the transactional stage out of its conclusions because a third measurement reversed the March-to-September direction. Treat any single two-date comparison of buying-intent prompts with the same caution.
Big platforms lost share on the shorter list. Single-purpose tools and institutional sources gained it.
On informational prompts, Microsoft fell 9.5 points, Google 6.7 and Amazon 6.4. Developer and data names such as Apache, dbt, Snowflake, Elastic and HashiCorp fell with them. IBM gained 9.9 points. NIST.gov gained 8.1 points and TechTarget 6.9, with Cisco, FBI.gov and FTC.gov also rising.
On consideration prompts, every gainer BrightEdge listed does one job:
When ChatGPT has fewer slots, it fills them with brands it can tie to one specific use case. "Does everything" loses to "does this one thing" on a short list.
Track this metric for your brand → nobori.ai
Yes, in most category queries. ChatGPT writes brand names into its own first search query before it fetches any page.
Suganthan Mohanadasan, co-founder of Snippet Digital, read the network traffic behind ChatGPT conversations on a logged-in Plus account in July 2026 and published the results on August 10. The model stores its self-written searches under a search_queries key in the response your browser downloads. For "best AI note taking app," the first query ChatGPT wrote before any results came back named seven products: Granola, Notion AI, Otter, Fireflies, Fathom, Mem and Limitless. The user named none.
Across 27 conversations, 21 first queries contained brands the user never typed. Across 13 unrelated categories, 11 did the same. When Mohanadasan asked for a recommendation without naming any vendor, ChatGPT reached into memory in 10 of 11 cases. A complaint with no shopping intent, "our meeting notes are a mess," became a search for Granola's pricing page.
The follow-up searches then ran one site: probe per named brand. ChatGPT never went looking for new candidates. It walked down a list it already held.
A brand named in ChatGPT's own query reached the final answer 68.9% of the time. A brand that ChatGPT only fetched during the search reached it 2.1% of the time. That is a 33x gap.
Mohanadasan also logged 86 cases where ChatGPT recommended a brand without fetching its website in that conversation. A mention does not require a crawl.
The second filter runs after the shortlist forms. Across 3,554 retrieved pages in 57 conversations, ChatGPT cited 110, or 3.1%. Pages in first position within a domain group earned a 5.2% cite rate. Pages in sixth place or lower earned 0.3%. Two tight pages per intent from one domain converted best at 6.2%; six or more fell to 1.7%.
The author flags the limits: one account, a few hundred conversations, a software-heavy query mix. Treat the percentages as direction. The mechanism, brand names injected before retrieval, you can reproduce in DevTools in two minutes.
Models search for brands they already remember at 3.2 times the rate of brands they don't. geoSurge measured this across nine industries, 66 US buyer prompts, 3,960 responses and 13,281 fan-out queries between May 29 and June 9, 2026 (published July 28).
geoSurge measured memory on one model and search behavior on Gemini 3.5 Flash, so the effect had to survive a model switch. Brands in the model's top-10 recall got searched 55.7% of the time (274 of 492 cases). Brands outside it got searched 17.4% of the time (161 of 924). Brands in the top-5 recall reached 67%.
Most fan-out queries stayed generic: only 31% named a brand. But 63% of those brand-named queries pointed at one of the model's five most remembered brands. The gap held in all nine industries, with remembered brands searched 41% to 82% of the time and unremembered brands 9% to 23%.
Category structure changes the strength of the effect. In automotive, 82% of brand-named queries hit a top-5 memory brand. In finance, 77% did. In fitness and wellness, 50% did, which leaves more room for a brand the model never recalled. geoSurge did not isolate B2B software, but any category where three or four vendors dominate analyst coverage is likely to behave like finance.
Traditional brand strength buys recognition. It does not buy a spot on the AI brand shortlist.
The Northwestern paper (Malthouse, Lee, Yang, Pal and Feng, arXiv 2609.16304, September 14, 2026) tested GPT-5.5, GPT-5.4 Mini, Gemini 3.1 Pro Preview, Gemini 2.5 Flash, Claude Opus 4.7 and Claude Sonnet 4.6 with web search off. The researchers built each competitive set from Kantar BrandZ and Statista market data before running any prompts, so brands the models skip still count as candidates. They ran 1,200 category-only lists and 1,200 needs-based lists.
Large brands disappeared. L.L.Bean, Eddie Bauer and REI Co-op never appeared for hiking jackets. Craftsman and Black+Decker never appeared for cordless drills. Braun and Philips never appeared for coffee makers. BrandZ salience showed "little systematic relationship" with how high a brand ranked in the lists.
What did track with prominence was marketplace visibility. Google Trends search interest correlated with ranking prominence at 0.63, and search interest carried the largest coefficient in every model the authors fit. Brandwatch online conversation came second. Ad spend and Wikipedia pageviews landed near zero. The authors label the findings exploratory, not causal, but the direction is clear: models reward brands people search for and talk about, not brands that buy reach.
Yes, when the prompt supplies the right cues. The Northwestern team calls this "conditional retrievability."
Craftsman and L.L.Bean scored 0% presence in the top five on plain category prompts. When the researchers rewrote prompts around specific buyer needs, such as budget, experience level and intended use, Craftsman rose to 35.4% presence and L.L.Bean to 5.4%. When the researchers used wording lifted from each brand's own positioning, presence jumped to 81.3% for Craftsman and 88.5% for L.L.Bean.
The models know these brands. They file them under attributes buyers rarely type. For a B2B team, that splits the problem in two. If your brand only surfaces when a prompt uses your own tagline, you have a positioning-vocabulary gap, not an awareness gap. The fix is to get third parties to describe you with the words buyers use, so the association forms around real buyer language.
geoSurge found the same escape hatch in live search. For "what payment provider should a startup use," Gemini searched Stripe, PayPal and Square from memory, then searched Lemon Squeezy, a brand it never recalled. Live web presence can pull an unremembered brand into the fan-out, but it is the less reliable route.
Recognition and recommendation run on different inputs. Victorious tested 175 brands across legal, healthcare, SaaS, financial services and ecommerce on eight AI platforms for its Q2 2026 Quarterly Search Report. The engines described 96% of the brands accurately when asked by name. 89% of the same brands never appeared in category answers.
The two signals that moved with mention rates both sit off your website: referring domains (r = 0.49) and third-party web mentions (r = 0.45). Victorious sells SEO services and did not publish an independent replication, so read the correlations as a benchmark.
FrictionAI and BrilliantSEO ran 14,140 controlled queries and found a sharper version. New Balance held the highest Google Knowledge Graph score in their sample, 64,235, and appeared in 3.4% of "best athleisure brands" answers. Lululemon, with a score of 810, appeared in 92.5%. Knowledge Graph strength correlated with the gap between recognition and recommendation at -0.10 across all brands. Within brands the models filed in the same category, the correlation rose to +0.68. The engines sort you into a category first. Strength only helps inside the right one.
Our earlier piece on entity optimization for AI search covers how to fix the identity layer that feeds recognition. The shortlist sits one layer above it.
Yes, on three of the four engines tested. Cassie Wilson Clark, host of the Found in AI podcast, and Joao da Silva, co-founder of friction AI, published "The Personalization Gap" on September 23, 2026, after analyzing 8,609 responses from ChatGPT, Gemini, Claude and Perplexity across six personas, three consumer categories and 30 prompts in the UK.
They compared history-free API calls, temporary logged-in sessions and accounts primed with user histories, then subtracted normal run-to-run variation. What remained was the history effect on the recurring brand set:
Geography drove the clearest shift. Primed UK accounts saw UK and EU brands gain 6.6 points of share, US brands lose 4.9 points, and .uk citations rise 6.0 points. Mohanadasan saw the same thing from Dubai: meal kit and insurance prompts returned UAE providers he never asked for.
Measurement takes a hit too. Across 173 matched comparisons, history-free API sampling recovered 79% of the recurring brands seen in primed sessions. Temporary logged-in sessions recovered 95%. Your tracking method changes the shortlist you see.
Fewer slots per answer raise the value of each slot and the cost of missing it. Run the arithmetic on a category with 12 credible vendors. In March, ChatGPT's 5.36 names per answer covered 44.7% of that field. In September, 3.87 names covered 32.3%. A vendor that sat in sixth or seventh place in spring now falls off the answer.
The downstream effect shows up in branded search, not referral clicks. Similarweb's 2026 Brand Visibility Index, as reported by AuthorityTech, found brands recommended by ChatGPT were 2.5 times more likely to receive a site visit within seven days, and 55.9% of that traffic arrived through branded search. ChatGPT plants the name, the buyer types it into Google days later, and your analytics credits organic search.
That attribution gap hides the loss. A vendor dropped from the shortlist sees branded search volume slide with nothing in the referral reports to explain it. If your branded search trend softened this quarter while rankings held, check the AI shortlist before you blame seasonality.
Run five checks. Each one takes under an hour.
Our query fan-out optimization guide covers how to map the follow-up searches once you know you are on the list.
Sort your brand into one of two games, then fund the right one.
Game one: you never appear in the first query. Page work will not move you, because the decision happens before your server gets a request. Spend on the inputs that correlate with shortlist presence: referring domains, third-party mentions, search interest and online conversation. That means analyst briefings, category roundups, review-site profiles, podcast appearances with transcripts, and comparison articles on publications your buyers read. Our breakdown of off-page AEO and third-party mentions ranks those channels.
Game two: you appear in the first query. Now the 3.1% citation filter decides your fate. Pick one page per buyer intent and fold duplicates into it. Put the answer sentence at the top in plain HTML. GPO reported in September that ChatGPT's free-tier index stores a page title plus about 200 characters of body text, so boilerplate intros waste your most valuable space.
For both games, own one job. BrightEdge's consideration gainers all do one thing. Pick the use case you win and make every third-party description repeat it.
Week by week:
Expect shorter lists, faster model turnover and more personal answers.
Model updates now land every few weeks. Google moved Gemini 3.5 Flash into AI Mode in July and Gemini 3.7 Flash in September, one day after its public release. OpenAI renamed its query key in early August, and Mohanadasan's captures dropped from 12 searches per answer to 4. Each change can reshuffle which brands the model reaches for first.
Personalization will widen the spread between what a tracker sees and what a buyer sees. Gemini already shows a 33.3-point history effect. As assistants store more memory about users, the "one true answer" becomes a range of answers, and visibility turns into a probability you measure across many runs.
The durable asset sits outside any one engine: a brand that people search for, write about and tie to one clear job. Our analysis of AI citation overlap found engines share 9% of sources but most of the same brands. Brand-level presence carries across engines in a way source-level tactics do not.
What is an AI brand shortlist?
An AI brand shortlist is the set of vendors an AI engine names in answer to a category question that names no brand. ChatGPT's B2B tech shortlist averaged 3.87 brands per answer in September 2026, down from 5.36 in March (BrightEdge 2026).
Does ChatGPT choose brands before it searches?
In most recommendation queries, yes. One network-traffic analysis found 21 of 27 first search queries contained brands the user never typed, and brands named in those queries reached the answer 68.9% of the time versus 2.1% for brands only fetched (Mohanadasan, August 2026).
Does a strong brand guarantee a spot on the shortlist?
No. Northwestern researchers found BrandZ salience showed little relationship with LLM recommendation prominence, while Google search interest correlated at 0.63. Victorious found engines recognized 96% of 175 brands but 89% never appeared in category answers.
Can on-page SEO get my brand onto the AI shortlist?
On-page work helps once you are on the list. The data gives it little power to put you there. Shortlist presence tracks referring domains, third-party mentions, search interest and online conversation, all of which live off your website.
Why do different people see different AI brand recommendations?
Chat history and location change the list. A September 2026 UK study of 8,609 responses measured 33.3 points of history-driven brand divergence on Gemini, 16.0 on ChatGPT, 7.2 on Perplexity and none on Claude.
How often should I re-check my AI shortlist position?
Monthly at minimum, and after every major model release. Run each buyer question five times per check, because a single answer can drop or add vendors between runs.
Is AI recommending your competitors or you?
Nobori tracks your brand across ChatGPT, Gemini, Perplexity, Google AI Overviews and Claude every day, and shows which competitors made the shortlist when you did not.
Get daily AI visibility alerts for your brand → nobori.ai
