How ChatGPT Chooses a Brand
ChatGPT picks brands in two steps: it builds a shortlist from Reddit, listicles and review sites, then verifies each shortlisted brand on its own website. We recorded 5,102 AI search sessions to test this. In a head-to-head activewear test, one brand was cited 5 of 5 times and a comparable competitor 0 of 5.
By SEOasis · ChatGPT search and Google Gemini · September 1 to October 23, 2026
AI search sessions recorded
named when verified vs never verified
sites read vs cited per answer
of sites read were never cited
Found, Then Verified, Then Named
- Brands that are found and then verified are named in 92% of answers and cited in 71%. Found but never verified: 45% and 37%.
- ChatGPT reads about 22 sites per answer and cites about 3.5. 81% of the sites it read were never cited.
- Reddit is read in 74% to 95% of sessions but cited in under 10%, except on reputation questions, where it becomes the answer.
- Brands lose for fixable reasons: claims the model can't confirm, missing credentials, name collisions, and facts hidden behind logins or widgets.
Two Months Watching ChatGPT Shop
A custom tool sends each prompt to ChatGPT's search engine through OpenAI's API, the closest programmatic match to ChatGPT search. It records every search query the model writes, every page the search returns (“consulted”), every link in the final answer (“cited”), and the answer text, so we can see which brands were “named” with or without a link.
Each prompt runs 3 to 10 times because ChatGPT answers differently each time. Gemini runs through a matching adapter on Google Search grounding.
| Scope | Detail |
|---|---|
| AI search sessions captured | 5,102 |
| Distinct prompts | 504, across about 15 industries |
| Engines | ChatGPT search (about 4,500 sessions) and Google Gemini (about 590) |
| Live experiment | A 124-page test site, rechecked against ChatGPT 8 times over 7 weeks |
| Window | September 1 to October 23, 2026 |
Recon, Then Verification
For a prompt like “best X for Y,” ChatGPT runs at least two searches, and neither looks like a human search.
Step 01
Recon search (discovery)
It expands the prompt into long-tail queries, often adding the year and the word “reviews.” It pulls Reddit threads and listicles to build a shortlist. Brand names from its training data get thrown out if they don't also appear in those results, so being in the training data is not enough.
Fed by: Reddit, guest posts, listicles, PR
Step 02
Verification search
It checks shortlisted brands against their own sites, sometimes with site-restricted searches. It favors terms like “official” and “price,” plus any certifications, return policies or competitor comparisons mentioned in the recon results.
Fed by: your own pages
Verification is where brands get named
| Outcome | Mentioned | Cited |
|---|---|---|
| On shortlist and passes verification | 92% | 71% |
| On shortlist, never verified | 45% | 37% |
The two phases look very different
| Round 1 | Round 2 and later | |
|---|---|---|
| Verification-style queries (site:brand.com, "official," price) | 7% | 64% |
| Independent third-party sources | 71% | 45% |
| Brand-official websites | 5% | 22% |
From the 596-session base study (93 prompts, 5 verticals).
Other Patterns From the Base Study
- Queries average 8.3 words and read like SEO strings, such as "best solid cologne for men 2026 reviews." 13% contain the current year and 9% say "reviews."
- 41% of answers take more than one search round, at about 2.5 queries per round.
- A brand's own site that ChatGPT pulls into its results is cited at roughly 6 times the rate of other pages. The verification search is the commitment signal: ChatGPT verifies brands it has already shortlisted and almost always keeps them.
- Mentions outrun citations by 13 to 21 points, so counting only links undercounts how often a brand appears in answers.
22 Sites Read, 3.5 Cited
Of 2,603 distinct domains ChatGPT read, 81% were never cited. The USPTO trademark database appeared in 136 sessions and was cited 0 times. Each category has its own authority stack: health draws on PubMed and professional associations, fragrance on GQ, Esquire and Fragrantica, apparel on brand sites plus certification bodies, and services on Reddit, brand sites and licensing boards.
Reddit builds the shortlist, and answers reputation questions
Only 1.6% of ChatGPT's queries contain the word “Reddit.” It shows up because it saturates the results, not because the model goes looking for it. In one service-business audit, a single Reddit thread was the model's main independent source on the client and was cited 31 times.
| Prompt type | Reddit cited |
|---|---|
| Category or discovery ("best X") | 3.5% |
| Explainer ("what is...") | About 0% |
| Transactional ("where can I buy X") | About 50% |
| Brand reputation ("is X legit?") | 43% to 73% |
412 prompts, 2,292 sessions, September 2026.
Crawler access comes first
The best-known product-review brand in the world appeared in none of 500 sessions, consistent with NYT blocking OpenAI's crawlers. RTINGS sits at the other extreme: data-heavy, independent, specific, and cited almost every time it is read.
| Site | Consulted (of 500 sessions) | Cited |
|---|---|---|
| 95% | 2 | |
| RTINGS | 108 | 70 (65%) |
| NYT Wirecutter | 0 | 0 |
500 sessions on 100 crowded product prompts. In the same test, 75% of sessions put a winning brand's name into the very first query: the model starts with candidates and searches to confirm them.
One Listicle Is One Vote
On prompts where ChatGPT's top pick changed between runs, the split followed the listicles it read.
- Mattress: Tom's Guide crowns Saatva and Sleep Foundation crowns Helix. ChatGPT went 2 Saatva, 2 Helix, and each answer cited the listicle that crowned its pick.
- Wireless earbuds: Sony won 4 of 5. The one Technics answer was the one that cited TechRadar, the only big publisher crowning Technics.
Where the major listicles agree, ChatGPT gives the same answer every time. Where they disagree, one new listicle with a different winner can flip individual answers.
“Best Non-Toxic Activewear Brands”
We ran one prompt five times. Each run made two searches, for 10 search calls in total. Brand B is a comparable brand with slightly better prices, so the gap is not about product quality or price. It is about surviving ChatGPT's investigation-then-verification process.
| Brand | Cited | Why |
|---|---|---|
| Brand A | 5 of 5 | On the shortlist every run; site states certifications clearly |
| Brand B | 0 of 5 | Never reached the shortlist |
Why Brand B lost: it failed both steps
Discovery. It did not appear in the Reddit threads that ChatGPT's first search surfaced, so it never entered the shortlist.
Verification. Even on the shortlist, its site would likely have lost. ChatGPT's verification searches kept returning to certifications and standards. Brand A's site covers each of these directly, so ChatGPT can confirm them in one pass. Brand B's site says its factories are “pursuing” GOTS, its restricted-substance details sit in an FAQ with no dedicated page, and it blocks OpenAI's training crawler (GPTBot).
| Search term | Times searched |
|---|---|
| GOTS | 46 |
| OEKO-TEX | 45 |
| "official" | 43 |
| PFAS | 8 |
Counts across the runs are approximate; they come from an auto-generated transcript.
Where Brand B Did Win
The same brand shows how page-level vocabulary changes results.
- Bio-based elastane: cited 4 of 5 times, because its site has a page that owns that vocabulary.
- Made in one US city: it beat Brand A because its factory is entirely local, while Brand A's FAQ names factories in several places and ChatGPT flagged the contradiction.
Ten Ways Brands Lose in ChatGPT
We ran the same method on 15+ businesses and kept seeing the same failure patterns. Businesses are described by vertical only.
01 · Men's fragrance brand
Named first, never linked
Named #1 in 5 of 5 "best solid cologne" answers with 0 citations; links went to Esquire, GQ and Men's Journal. On 40 non-branded questions it was named 5 times, vs 22 for the category leader.
02 · Wholesale fashion supplier
Wins the shortlist, loses the checkout
Named in 11 of 15 answers, #1 every time. All 3 reputation answers sent buyers to a marketplace instead, because the marketplace published the minimums, payment terms and returns the supplier hid behind a login. The marketplace out-cited the supplier's own site about the supplier, 22 to 17.
03 · Fire-protection distributor
Site claims fail verification
The brand prompt failed 5 of 5 ("provisionally legitimate but unverified"). ChatGPT caught a distributor claim the manufacturer contradicted, a flow-rate spec about 1,000 times off, and certification badges with no documents. All free to fix.
04 · Interstate moving broker
ChatGPT argues against the purchase
0 citations across 50 non-brand answers. On brand prompts, 5 of 5 answers told the buyer not to book, citing a suppressed Trustpilot rating and a BBB complaint pattern. Reputation had to be fixed before traffic.
05 · National mortgage lender; telehealth brand
Name collision
The lender was invisible on 40 of 40 category answers and was confused with two same-name lenders. Its site listed three license numbers and two parent companies, and 411 reviews sat in a widget the crawler couldn't see. The telehealth brand's "is it legit" prompt cited a scam warning for a different, similarly named site in 5 of 5 runs.
06 · Nitrate supplement brand
The name is the liability
On a low-oxalate prompt, ChatGPT told buyers to avoid "spinach-based" products, and the brand has "spinach" in its name. On the athlete prompt its criteria matched the product exactly, yet it never retrieved it. The site never mentioned the certification body or the unit ChatGPT searched for.
07 · Career counselling practice; medical weight-loss clinic
Missing the credential the model checks
A competitor won 10 of 10 because ChatGPT treats "registered psychologist" as a protected title, and the client's site said "psychologist-led." The clinic was #1 on its own name but named 0 of 9 times on need-based prompts, and its site named no supervising physician.
08 · Rave apparel retailer; European proptech startup
Found, read, never verified
The retailer's plus-size page was read 3 of 3 times and cited 0, because it showed no size range while ChatGPT checked six competitors' size charts for 3XL and 4XL. The proptech's "99% accurate, cross-referenced against 40 million properties" claim from a sales call appeared nowhere on its site.
09 · Local self-storage operator; fine-jewelry boutique
The shortlist lives somewhere unexpected
The operator went 0 of 5 on "best storage in [city]." The winner took 5 of 5 on a 5.0 rating from 13 Google reviews. National storage directories got 0 citations; the shortlist came from Reddit, review aggregators, the chamber of commerce and city government pages. The boutique was cited 3 of 45 times vs the leader's 20, and the designers it stocks won on their own sites.
10 · European proptech, 7 markets
Language follows geography
Naming a country in an English prompt switched ChatGPT's searches into that country's language: 91% local-language queries and 88% local-domain sources, even from a US location. The client won only the pan-European prompt, through one English blog page cited 3 of 3 times.
Two More Things Kept Repeating
- Qualifiers flip results. The wholesale supplier scored 0 of 15 on head terms like “best fashion wholesaler” and ranked #1 once qualifiers were added. The rave retailer went from 0 of 9 consulted on broad prompts to 13 of 18 consulted and 8 of 18 cited on qualified ones.
- City pages matter more for AI than for Google. “Long distance movers [city]” gets about 20 searches a month in Phoenix on Google. In ChatGPT, city-plus-service prompts cite ordinary company websites directly, while national prompts cite aggregators.
Can a Brand-New Site Get Cited?
We built a site for a fictional skincare product on a new domain, launched September 1, 2026. It held 120 articles titled to match search queries ChatGPT had written for real prompts, in three tiers of 40:
- High competition: generic buyer prompts (“best NAD face cream”)
- Medium: condition-specific prompts (“NAD face cream for rosacea”)
- Low: very specific audiences (“NAD face cream for lighthouse keepers”)
| Day | High | Medium | Low |
|---|---|---|---|
| 3 to 5 | 0 / 0 | 0 / 0 | 0 / 0 |
| 9 | 25 / 0 | 58 / 0 | 88 / 33 |
| 26 | 20 / 0 | 60 / 1 | 91 / 44 |
| 48 | 20 / 0 | 51 / 0 | 89 / 39 |
Consulted / cited, out of 100 sessions per tier. 300 ChatGPT sessions per recheck, 8 rechecks over 7 weeks.
A 9-day-old site was cited in a third of low-competition answers
With one backlink. On the lighthouse-keepers prompt it was cited 10 of 10 times and became the #3 cited domain in that tier, behind only the American Academy of Dermatology and PubMed.
Competition decided the result
Low-tier pages were cited, medium-tier pages were read but not cited, and high-tier pages were barely read.
ChatGPT found the site before Bing did
Bing showed zero indexed pages while ChatGPT retrieved the site in 88% of low-tier sessions, through OpenAI's own crawler. AI search visibility is not the same as Bing SEO.
Query behavior follows specificity
On crowded prompts, over 98% of ChatGPT's queries were unique. On very specific prompts it searched the prompt word for word: "NAD face cream lighthouse keepers" appeared in 10 of 10 runs.
What Didn't Work
We added PubMed-backed evidence sections to 80 medium and high articles and left the low tier untouched as a control. Citations stayed flat at 0 to 1. The reason: across every capture since launch, ChatGPT opened one of our pages 0 times out of more than 260 page-open actions. Every citation came from the search-result snippet, which is the top of the page, and our research sat around word 900 of 1,250. Moving the evidence to the top didn't change results in the first week, and the test is still running.
On crowded topics, fresh exact-match content can't break in on relevance alone; it needs authority, mentions and third-party coverage. On specific long-tail topics, a new site can win within days.
Two Engines, Two Playbooks
We ran the same 100 prompts 5 times on each engine. Gemini leans on what it already knows, so brand recognition counts for more there. Where the engines split on flagship products, Gemini favored the household name and ChatGPT favored the reviewers' pick.
| ChatGPT | Gemini | |
|---|---|---|
| Answered without searching | About 0% | 38% |
| Queries per answer, when searching | 5.8 | 1.0 |
| Top sources | Wikipedia, Consumer Reports, TechRadar, Tom's Guide | YouTube, Reddit, Facebook |
| Wirecutter appearances | 0 | 0 |
| Prompt | Gemini | ChatGPT |
|---|---|---|
| Wireless earbuds | AirPods Pro | Sony WF-1000XM6 |
| Noise-cancelling headphones | Bose QC Ultra | Sony WH-1000XM6 |
| Streaming device | Apple TV 4K | Roku Ultra |
| Home printer | Brother | Epson EcoTank |
- Gemini reads a video's title, description and captions, and many of its grounded YouTube sources were faceless, AI-narrated “Top 10” channels, so clean captions can get a video grounded.
- Wirecutter reaches Gemini through YouTube even though both engines are blocked from its website.
- The engines shared only 47 Reddit threads. Those are the canonical threads both search backends rank first, and they carry weight on both.
A brand optimized for ChatGPT is not automatically visible in Gemini or AI Overviews.
Be in Both Places
A brand needs to win discovery and verification. Reddit gets it onto the shortlist, and its pages must match what ChatGPT searches for to verify it. Start by diagnosing which gate the brand is failing.
| Symptom | Failing gate | Typical fix |
|---|---|---|
| Never consulted or named on category prompts | Shortlist | Earn placement on the specific sources the model reads for the category |
| Consulted but never cited | Verification or citability | Put the citable fact in the first lines, specific and checkable |
| Named but not linked | Citability | Become the most linkable source for the fact |
| Brand prompt ends in "verify" or "avoid" | Verification | Fix contradictions, publish proof, resolve name collisions |
| Reputation prompts cite Reddit or Trustpilot negatively | Reputation layer | Do reputation work before traffic work |
Step 01
Get on the shortlist
- Earn presence in the Reddit threads and listicles that ChatGPT's recon search surfaces.
- Add guest posts and PR mentions in the same sources.
Step 02
Pass verification
- Run the target prompt many times in fresh ChatGPT sessions and log the second-search queries. We recommend 10 to 15 runs; 5 is enough to see patterns but thin.
- Build collection pages around the recurring terms, mapped to real search queries rather than your internal catalog structure.
- Put every checkable fact where ChatGPT looks first (see the table below).
- Build dedicated pages for materials, restricted chemicals and manufacturing claims, and state certifications as held, not "pursuing."
- Use any differentiator everywhere it can be found, like a single-city manufacturing claim.
Put every checkable fact where ChatGPT looks first
| Page element | Job |
|---|---|
| Title tag and H1 | Main SEO term plus the facts ChatGPT verifies (materials, certifications) |
| One or two sentence answer block under the H1 | State every verifiable fact up front, since ChatGPT may read only the top of the page |
| Product grid or item-list schema | Price, size, availability, description |
| Comparison table, FAQ, proof strip | Optional; add when the logged searches show comparisons, FAQs, or reviews and press links |
Technical Requirements
- Allow crawlers. Cloudflare blocks LLM bots by default, and Brand B blocks GPTBot.
- Put key facts in raw HTML, since ChatGPT does not render JavaScript.
- Add structured data (schema markup) near the top of the HTML.
- Use keywords in URLs and link to collection pages from blog posts.
Also From the Research
- Go after qualified long-tail prompts before head terms. Platforms and established brands own head terms, while the long tail can be won in days.
- Put the claim you want cited in the first lines of the page, since ChatGPT often judges a page by its search snippet.
- Treat Reddit as a shortlist and reputation lever, not a link source, except on reputation questions, where it decides the answer.
- Measure mentions as well as citations. Links alone undercount visibility by 13 to 21 points.
- Plan for each engine separately: ChatGPT reads publishers, and Gemini reads YouTube and social.
Limitations
- The data comes from OpenAI's and Google's APIs, which are close proxies for the consumer apps but not identical.
- Most audits use 3 to 5 runs per prompt: enough to see patterns, not enough for statistical certainty on a single prompt.
- The activewear head-to-head is one prompt and five runs; the rest of the study covers 5,102 sessions.
- AI answers change daily, so every number is a snapshot from September to October 2026.
- The live experiment is one site in one category.
More Case Studies
Lillie&Lee: Our First PR Campaign for AI Visibility
One survey, three press releases, 1,265 new backlinks.
Bronson Vitamins: Unlocking AI Visibility
+264% AI-referred revenue.
Dr. Brandt: First 4 Months in the GEO Era
A new AI-referred revenue channel built from zero.
Find Out Which Gate Your Brand Is Failing
Book a free audit and we'll run your buying prompts through ChatGPT, show you whether you're missing the shortlist or failing verification, and scope a plan against the gaps.
Book Your Free Audit