AI Search Cites Vendor Pages, Not Blogs: What 5,376 Citations Reveal

The GEO industry has two competing narratives. One says earned media drives AI citations — journalism, Reddit, review sites. The other says your own website is the evidence layer AI engines trust. A new study from Kern Media, analyzing 1,044 AI answers and 5,376 citations across 8 industries, lands firmly on the second side — with important caveats about which pages on your site actually get cited.

What Kern Media measured

Kern Media pooled 13 AI-visibility tracking programs covering 12 brands across Consumer e-commerce, B2B e-commerce, Education, Financial services, Professional services, B2B software, Real estate, and Travel & hospitality.

Each program tracked a fixed set of buyer-intent prompts written from 4–8 detailed buyer personas — "budget-conscious small-business owner," "risk-averse enterprise IT leader," and similar. For every answer, Kern captured which brands were recommended (and in what order) plus every source URL cited.

Most answers came from ChatGPT (948 across three GPT model versions). One B2B software program ran the same 32 prompts through ChatGPT, Gemini, Perplexity, and Claude for head-to-head engine comparison.

Finding 1: Company websites dominate citations

Across all 5,376 citations, 84.2% pointed to the websites of companies being recommended:

News, forums, video, encyclopedias, and universities combined: less than 3%.

This is not uniform across engines. ChatGPT leaned hardest on company sites:

With ChatGPT, your website is your evidence. With Perplexity and Gemini, third-party sources matter far more. This aligns with Separate AEO Strategies for ChatGPT and Perplexity.

The tension with other studies is real. Evertune's analysis of nearly 400 million citations found listicles dominate citation formats. Kern's finding is about whose domain gets cited, not what page format third-party sites use. Both can be true: AI cites company websites, and when it cites third parties, those are often listicles and directories.

Finding 2: Citations and recommendations reinforce each other

95% of company-site citations belonged to a brand recommended in the same answer. Engines rarely cite a brand's site and then leave it off the recommendation list.

For the 12 tracked brands specifically:

When a tracked brand's own site appeared among sources, it was recommended 96% of the time and ranked #1 in 47% of answers. When its site was absent, it was recommended just 18% of the time and ranked #1 in 4%.

This is correlation, not causation. Engines likely retrieve sources for brands they plan to recommend. But the practical implication is clear: if your site is not retrievable, you are unlikely to be recommended.

Finding 3: Commercial pages beat blog posts

On company websites, citations went overwhelmingly to commercial and trust pages:

Blog content — the format most SEO strategies are built on — made up just 6.9% of company-site citations. On-site comparison pages: 1.7%.

Third-party citations looked completely different: articles, guides, "best of" listicles, and directory profiles dominated.

Industry-specific "citation pages"

Every vertical has a page type that earns disproportionate citations:

There is no universal "best page for AI citations." There is a best page for your industry.

Finding 4: Persona matters more than prompt phrasing

A single brand's AI visibility swung by as much as 80 percentage points depending on which buyer persona asked the question. Changing how the prompt was phrased moved visibility by only a few points.

This matters for measurement. Tracking one generic prompt ("best CRM software") misses the variance that persona-driven prompts reveal. A brand invisible to the "enterprise IT leader" persona may dominate for the "startup founder" persona — and you will not see either if you only test one phrasing.

Finding 5: Engines barely agree on recommendations

On the same 32 prompts across four engines, brand shortlists overlapped by just 7–18%. The four engines never once agreed on the #1 pick.

Combined with Brain Buddy's September 2026 finding that ChatGPT and Google AI Overviews share only 8% of sources on the same question — see AI Engines Do Not Share Citation Lists — the picture is clear: there is no single AI ranking to optimize for.

Finding 6: Specificity wins top recommendations

#1 recommendations were nearly 4× as likely as #6+ picks to include a concrete number (33% vs 9%). They were also far less likely to include a caveat (2% vs 13%).

Vague recommendations ("a solid option for teams of all sizes") lose to specific ones ("serves 2,400+ enterprise clients with 99.9% uptime SLA"). AI engines reward evidence they can extract and repeat.

What this means for your strategy

Fix your commercial pages before your blog calendar

If ChatGPT is your primary surface, the pages that describe what you do, who it is for, and why you are the right choice are your citation layer. Homepages, service pages, product pages, and About pages — not the blog archive.

That does not mean stop blogging. It means blogging is not where ChatGPT looks for evidence about your brand.

Match page type to industry

A real estate agent should optimize bio pages. An e-commerce brand should optimize product pages. A university should optimize program detail pages. Copying a generic "publish listicles" playbook without checking your industry's citation pattern wastes effort.

Test with personas, not just keywords

Build prompt sets from real buyer personas — the risk-averse CFO, the technical evaluator, the end user. Visibility can swing 80 points between them. One generic prompt is not a measurement program.

Split strategy by engine

ChatGPT: own-site evidence. Perplexity: third-party listicles, directories, and fresh content. Gemini: mix of both. A single GEO checklist applied uniformly will over-invest in the wrong layer for at least one engine.

How Obsurfable fits

Obsurfable runs buyer prompts across AI engines and records which brands get named, which pages get cited, and how answers shift over time. Kern Media's study shows the aggregate pattern across 8 industries. Obsurfable shows your pattern — which of your pages Copilot and ChatGPT retrieve, which competitors displace you, and which personas see you versus do not.

Bottom line

Across 5,376 citations in 1,044 AI answers, 84% pointed to company websites — not press, forums, or video. On those company sites, commercial pages dominated and blogs made up just 6.9%. Citations and recommendations moved together: own-site cited meant 96% recommendation rate versus 18% without. Persona swung visibility by up to 80 points; prompt phrasing moved it only a few. Engines agreed on the #1 pick zero times out of 32 shared prompts.

Your website is your evidence — but only if the right pages on it are crawlable, specific, and matched to what your industry’s buyers actually ask.