Where AI Engines Go to Decide About Your Brand
I sorted the 32,000 sources ChatGPT, Claude, Gemini, and Perplexity cited while they advised a buyer. About four out of five came from somewhere other than the brand being asked about.
When a marketing team wants to know what AI says about them, the first thing they usually do is go look at their own website. Is the product page clear, is the pricing up to date, did somebody finally fix that FAQ. I get why. It's the one thing you control, and it feels like the obvious place an engine would go to learn about you.
It's one of the places. It's a smaller one than most people assume, and it depends a lot on which question the buyer is asking at the moment.
I know that because I already had the data. Earlier this month I ran the AI Verdict Study, where I ran 432 scripted buyer conversations across four AI engines to see which brands survive a real back-and-forth and which ones get talked out of the deal. Every one of those conversations logged the sources the engine cited, turn by turn. That's 32,089 citations to 3,241 different websites, and nobody had looked at them yet, including me. So I did.
How I sorted them
Quick version, and the full method is at the bottom for anybody who wants to check it.
The conversations covered nine categories, each with a challenger brand and the category leader it competes with (Pipedrive and HubSpot, Huntress and CrowdStrike, Omnisend and Klaviyo, Helix and Tempur-Pedic, and five more). The buyer in every conversation used the same seven questions in the same order, starting with "what are the best options for someone like me" and ending with "should I choose this brand?"
Gemini hides its sources behind Google redirect links, so I unwound all 5,436 of those back to the real pages (all but 278 still resolved). Then I sorted the websites into eight types: the brand being asked about, its direct rival, other companies (software vendors, retailers), review marketplaces like G2 and Capterra, independent review and affiliate sites, mainstream and trade media, community sites like Reddit and YouTube, and health, government, and nonprofit sources. I hand-labeled the 500 most-cited domains plus every study brand's own site, which covers 76% of all citations. The long tail after that is too long to label by hand, so the shares below are shares of that 76%.
The shortlist gets built from other people's pages
The first question the buyer asks never mentions a brand. It just describes the situation and asks what the options are. For that question, the brand's own website accounted for 4% of the classified citations.
The rest came from other companies' sites (39%), independent reviewers (21%), and media (19%). So when an engine decides who makes the first list, it's mostly reading pages written about the category by somebody else, and a lot of those somebodies are your competitors.
Who the engines read, question by question
Every conversation asked the same seven questions in the same order. Each column shows where that question's citations came from, and the blue number under it is how often the engine searched at all on that question. The yellow at the bottom is the brand the buyer was asking about. Tap a column for the breakdown.
Source: AI Verdict Study v1, jarredsmith.com. 32,089 citations from 432 search-on conversations, Sept 18 to 20, 2026. Shares are of the 76% of citations from the 500 most-cited domains plus every study brand's own site.
That chart is worth clicking through engine by engine, because the shape changes a lot. But the overall pattern holds for all four. Your share of the conversation is small at the start, jumps when the buyer asks about you specifically, drops again when the buyer asks what people compare you with, and comes back partway at the end.
Your site gets read when the buyer asks about you
Turns three and four are where the buyer gets specific. "What are the main tradeoffs or limitations I should know about before choosing this brand," and then "would it still be a good fit if my needs grow." On those two questions the brand's own site jumped to 35% and 37% of citations.
The part I didn't expect was which pages. Across every turn, help center and support documentation made up 29% of all the citations to a brand's own site, more than the blog, the pricing page, and the comparison pages put together. On turn three, the tradeoffs question, it was 46%. When a buyer asks "what are the limitations," the engine goes and reads your help docs, finds the article that explains how to work around the thing your product doesn't do, and repeats it.
I don't think most teams write their help center knowing it's going to be quoted to a prospect as the downside of their product. I'd write it that way now.
Your competitors write a lot of the comparison pages
Turn five asks "which tools do people like me most often compare with this brand, and what do they say about the differences." On that question the brand's own site dropped to 7%, and other companies' sites and independent reviewers each took about 29%.
Some of that is expected, and some of it is your competitor's marketing team. I flagged every cited page on another vendor's site that was a "best X" list, an "X vs. Y" page, or an "alternatives to X" page. Those made up 13% of all classified citations, and 17% in the software categories.
The specific cases are pretty striking. In the conversations where the buyer asked about Klaviyo, the engines cited omnisend.com 130 times, and 63 of those were Omnisend's own comparison pages, like their Omnisend vs. Klaviyo post. In the CrowdStrike conversations, huntress.com got cited 101 times, including Huntress's own Huntress vs. CrowdStrike page. Gusto's "Gusto vs. OnPay" pages showed up in the OnPay conversations. And in the project management conversations, Teamwork's "Asana alternatives" post and a whole series of monday.com "X vs. Asana vs. monday" posts kept coming up while the buyer was asking about Asana.
Klaviyo, for what it's worth, got rejected in 10 of its 24 conversations in the study, and more than once the engine told the buyer to go with Omnisend. I can't prove the comparison pages caused that, but if I worked at Klaviyo I'd want a few more of those pages written by us.
This varied a lot by engine. On Gemini, competitor-written comparison and list pages were 20% of citations. On Claude, 15%. On ChatGPT, 2%.
Each engine reads a different internet
Each engine reads a different internet
Where each engine's citations came from across all seven questions. Hover a segment for its label. Switch views to compare software buyers with consumer buyers.
Citations and domains are averages per seven-turn conversation. Overlap is the average share of cited websites two identical runs of the same conversation had in common. Source: AI Verdict Study v1, jarredsmith.com. 32,089 citations from 432 search-on conversations, Sept 18 to 20, 2026. Shares are of the 76% of citations from the 500 most-cited domains plus every study brand's own site.
This was the biggest surprise in the whole data set, and it lines up with what I saw in the verdicts.
ChatGPT reads company websites almost exclusively. Brand sites, rival sites, and other vendors added up to 84% of its classified citations, and it cited independent reviewers 2% of the time and media 3%. It also cited the fewest sources, about 33 per conversation from 14 different domains.
Perplexity is the opposite. It averaged 140 citations per conversation from 65 domains, and media was 23% of its mix, far more than any other engine. Forbes alone was 4% of everything Perplexity cited.
Gemini leaned on community sites more than anyone, at 11%, mostly Reddit and YouTube. It also cited the brand's own site the least (11%), and 33 of its 108 conversations never cited the brand's own site at all.
Claude cited review marketplaces like G2, Capterra, and PeerSpot more than the others (12%) and independent reviewers heavily (30%). It cited Reddit zero times in 8,193 citations.
About Reddit, since it comes up a lot. There's a lot of advice going around right now that Reddit is where AI visibility gets won. In this data Reddit was 2.1% of all citations, and G2 and Capterra together were 1.7%. That's not nothing, and in a couple of categories Reddit mattered a lot (it was the single most-cited site for the home espresso buyer). But it's one source among thousands, and one of the four engines ignored it entirely.
Software buyers and consumer buyers get sent to different places
The split between B2B software and consumer products was bigger than the split between engines in some cases.
For the six software categories, the brand's own site was 23% of citations and other vendors were 32%. For the three consumer categories (running shoes, mattresses, espresso machines), the brand's own site dropped to 9%, and independent reviewers (39%) and media (28%) took over. In mattresses, the three most-cited sites were Forbes, the Sleep Foundation, and Tom's Guide. Helix and Tempur-Pedic's own sites were a rounding error by comparison.
The ten sites the engines read most, by category
Pick a category. These are the websites all four engines cited most while advising one buyer about a challenger and its category leader, as a share of all citations in that category.
Source: AI Verdict Study v1, jarredsmith.com. 32,089 citations from 432 search-on conversations, Sept 18 to 20, 2026. Shares are of the 76% of citations from the 500 most-cited domains plus every study brand's own site.
Pick your closest category in that tool. The list of sites is a pretty good starting map of who you'd want to be talking to, and it's different in every category. For project management it's mostly vendor blogs and a handful of independent reviewers. For payroll it's the two brands' own sites plus business.com. For security it's the vendors, PeerSpot, G2, and a couple of security news sites.
The sources change almost as much as the answers
In the main study, 21 of 72 brand and engine pairings (29%) gave the same buyer, asking the same questions, a yes on some runs and a no on others. The sources move around even more.
Two runs of an identical conversation, same engine, shared on average 38% of the websites they cited. Perplexity was the most consistent at 53%. Gemini was the least at 28%. And the share of websites an engine cited in every single run of a conversation was small, anywhere from 5% (Gemini) to 25% (Perplexity).
The long tail is long, too. The 10 most-cited websites covered 14% of all citations, and it took the top 100 to get to 44%. There's no short list of five sites that everybody reads.
Most engines stop searching before the final answer
The first chart has a second number under each question, which is how often the engine searched at all. On the last question, "should I choose it," ChatGPT searched in 13% of conversations, Claude in 2%, and Gemini in none of them. Perplexity searched in 95%. So for three of the four engines, the verdict comes from whatever they already read earlier in the conversation, plus whatever they believed going in. If your brand didn't come across well in the first five answers, there's usually no second look before the engine makes the call.
What being cited did for the verdict
I checked whether a brand's own site being cited had anything to do with how the conversation ended, and the connection is weaker than I'd have guessed.
The brand's own site got cited somewhere in 88% of conversations. Challengers whose sites got cited five or more times were rejected 12% of the time, which looks good. Category leaders whose sites got cited five or more times were rejected 40% of the time, which doesn't. So getting read is the minimum, and what the engine decides after reading you seems to depend a lot more on how well you fit the buyer and on what everybody else wrote about you.
What I'd do with this
Number one, go read your help center like a skeptical buyer. On the "what are the limitations" question, it was the most-cited part of your site. Make sure the articles about what your product doesn't do explain who it's not for and what to do instead, in plain terms, because that's the passage that's going to get quoted.
Two, write your own comparison pages, and write honest ones. Your competitors already wrote theirs, and the engines are reading them in conversations about you. A comparison page that says clearly who should pick the other product is more believable, and I'd bet more quotable, than one that says you win at everything.
Three, find the independent reviewers in your category and get in front of them. Use the category tool above as a starting point. For consumer products especially, those sites plus the big media outlets are most of what the engines read.
Four, check each engine separately. ChatGPT reads company websites, Perplexity reads media, Gemini reads Reddit and YouTube, and Claude reads review marketplaces. One blended "AI visibility score" hides all of that, and the fix for each one is different.
Five, check more than once. If two identical runs only share about a third of their sources, one check tells you about one draw.
Six-question source check
Answer for your own brand. You'll get a short list of what to fix first, based on what the engines read in this data. Nothing you click here is stored or sent anywhere.
Recommendations draw on the citation analysis in this post. Source: AI Verdict Study v1, jarredsmith.com.
Method and limits
Sources come from the 432 search-on conversations in the AI Verdict Study (September 18 to 20, 2026), run through the ChatGPT (gpt-5), Claude (Haiku 4.5), Gemini (2.5 Flash), and Perplexity (Sonar Pro) APIs with no system prompt. That's 32,367 citations; 278 Gemini links no longer resolved, leaving 32,089. I hand-labeled the 500 most-cited domains and every study brand's own domains into eight types, which covers 24,331 citations (76%). "Comparison and list pages" were flagged from the URL (paths containing "vs," "alternative," "compare," "best," or "top"), so that count is an estimate and probably a low one. These are API results; the consumer apps use their own search setups and may cite differently. It's nine categories and three days, so treat the category-level numbers as a map and not a census. The full dataset, including every cited URL by conversation and turn, is on the study page.
I spend a good chunk of Explainable on why the pages that get cited aren't always the ones a brand would pick, and this data put some numbers on it I didn't have when I wrote the book. If you want to know where the engines are reading about your brand specifically, that's most of what an AI Visibility Audit covers, or you can grab 15 minutes with me and I'll tell you whether it's worth doing. I'd also love to hear what you're seeing in your own category, especially if it doesn't match this.
Jarred Smith is the author of Explainable: Why AI Recommends Some Brands & Ignores Others, an Amazon bestseller on AEO, GEO, and SEO. He's a marketing leader with nearly 20 years of experience across healthcare, public media, retail, and environmental services. Find him at jarredsmith.com.