TALK TO US: We analyzed every recommendation and every source link from 13 AI-visibility tracking programs: 948 buyer-style prompts, 96 buyer personas, 4,872 brand recommendations and 5,376 citations. Here’s which content AI engines cite, which pages win the recommendation, and what your website needs.
Key takeaways
- AI search cites company websites, not the press. 84% of all citations went to the websites of the companies being recommended. News and media sites earned 1.0%, forums 0.8%, and video 0.6%.
- If your site isn’t cited, you’re rarely recommended. When a tracked brand’s own site was cited, it was recommended 96% of the time and ranked #1 in 47% of answers. When it wasn’t cited, it was recommended just 18% of the time and ranked #1 in 4%.
- Commercial pages get cited, not blog posts. On company sites, homepages, service pages, About pages and product pages earned most citations. Blog content made up only 6.9%, and on-site “comparison” pages just 1.7%.
- Every industry has its own “citation page.” Product pages in e-commerce (44%), program pages in education (43%), service pages in professional services (38%), agent bio pages in real estate (39%) and sample itineraries in travel (25%).
- Who’s asking matters more than how they ask. A single brand’s visibility swung by as much as 80 points depending on the buyer persona. Prompt phrasing moved it by only a few points.
- AI engines barely agree. On the same 32 prompts, the four engines’ brand shortlists overlapped by just 7–18%, and they never once agreed on the #1 pick.
- Top recommendations are written with specifics. #1 recommendations were nearly 4x as likely as #6+ picks to include a concrete number (33% vs. 9%), and far less likely to include a caveat (2% vs. 13%).
About the data
This study pools 13 AI-visibility tracking programs that cover 12 brands. Each program tracks a fixed set of buyer-intent prompts written from the point of view of 4–8 detailed buyer personas, such as a “budget-conscious small-business owner” or a “risk-averse enterprise IT leader.” For every answer, we capture which brands were recommended (and in what order), plus every source URL the AI engine cited.
The programs span 8 broad industries: Consumer e-commerce, B2B e-commerce, Education, Financial services, Professional services, B2B software, Real estate and Travel & hospitality. Where an industry has more than one program, they are labeled A, B and C.
Most answers came from ChatGPT (948 answers across three GPT model versions). One B2B software program ran the same 32 prompts through ChatGPT, Gemini, Perplexity and Claude, which gives us a head-to-head comparison of the engines. Throughout this study, the “tracked brand” is the brand being monitored in each program.
Finding 1
AI search cites company websites, not the press
The biggest surprise was how little of the cited web is “earned media.” Across 5,376 citations, 84.2% pointed to the websites of the companies being recommended: 75.3% to other vendors and 8.9% to the tracked brands’ own sites. Combined, news publications, forums, video, online encyclopedias and universities made up less than 3%.
Where AI-search citations point
Share of all 5,376 citations, by type of site cited
“Websites of the companies being recommended” means the cited domain belongs to a brand that appears as a recommendation somewhere in the dataset.
This isn’t true of every engine. ChatGPT leaned hardest on company sites. Gemini and especially Perplexity pulled in far more third-party content.
Share of citations going to company websites, by engine
% of each engine’s citations that link to a recommended company’s own site
Gemini and Perplexity figures come from a single B2B software program (32 prompts each). ChatGPT figures span all 13 programs.
Finding 2
Citations and recommendations go together
AI engines mostly cite the sites of brands they’re already recommending. 95% of company-site citations belonged to a brand that was recommended in that same answer. Brands were rarely cited and then left off the list.
Tracked brand outcomes when its own site was cited vs. not
All 1,044 answers
Own site cited (385 answers)
96% recommended
47% ranked #1
Own site not cited (659 answers)
18% recommended
4% ranked #1
This shows correlation, not causation. Engines likely retrieve sources for the brands they plan to recommend, so citation and recommendation reinforce each other.
The pattern holds for every brand, not only the tracked ones. 74% of all 4,872 recommendations were backed by a citation to that brand’s own website, and the higher the rank, the more likely it was.
How often a recommended brand’s own site was cited, by rank
% of recommendations at each position where the brand’s website appeared among the answer’s sources
Finding 3
The pages that get cited: commercial pages beat blog posts
We classified every cited URL by page type. On company websites, citations went overwhelmingly to commercial and trust pages: homepages, service pages, About/team pages, product pages and program detail pages. The blog content most SEO strategies are built on made up only 6.9% of company-site citations. On-site “comparison” or “best of” pages were rarer still, at 1.7%.
Third-party citations looked completely different. There, articles and guides, “best of” listicles and directory or review-site profiles dominated.
Company-website citations
Page type share · 4,528 citations
Third-party citations
Page type share · 848 citations
Page types were classified from URL patterns. About 18% of company-site URLs (and 22% of third-party URLs) were specific deep pages that didn’t fit a type, so they’re excluded from the bars above.
Finding 4
Every industry has its own “citation page”
The single most-cited page type changes completely from industry to industry. Plan your AI-search content around the page type that works in your industry, not a generic template.
Most-cited page types on company websites, by industry
% of each industry’s company-site citations (rows don’t sum to 100%; uncategorized deep pages are omitted)
| Industry | Home | Service / solution | Product / category | Program / package | About / team / bios | Blog / guide | FAQ / policy | Audience / use-case | Itinerary | Area guide | |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Consumer E-commerce | 14 | – | 44 | – | 7 | 3 | 12 | – | – | – | 1 |
| B2B E-commerce | 19 | 3 | 12 | – | 21 | 4 | 12 | 1 | – | – | 3 |
| Education | 11 | – | 4 | 43 | 11 | 4 | 5 | – | – | – | 2 |
| Financial Services | 13 | 28 | 8 | – | 8 | 5 | 1 | 8 | 2 | – | 6 |
| Professional Services | 8 | 38 | 3 | – | 4 | 19 | 1 | 5 | – | – | 7 |
| B2B Software | 15 | 15 | 7 | – | 4 | 11 | 13 | 4 | – | – | 3 |
| Real Estate | 27 | 1 | – | – | 39 | 4 | 1 | 1 | – | 10 | 1 |
| Travel & Hospitality | 32 | 2 | 1 | 1 | 6 | 3 | 8 | – | 25 | – | 2 |
- Consumer e-commerce: product and collection pages (44%), plus FAQ and shipping/policy pages (12%).
- Education: program and package detail pages with scope, sequence and what’s included (43%).
- Professional services: individual service pages (38%) and in-depth articles (19%).
- Financial services: service pages (28%), plus individual advisor and team microsites.
- Real estate: agent and team bio pages (39%), homepages (27%) and area guides (10%).
- Travel: homepages (32%) and sample itinerary pages (25%).
- B2B e-commerce: About and company pages (21%), homepages (19%), plus product and policy pages.
- B2B software: a broad mix of homepages, solution pages, documentation, FAQs and industry pages.
Finding 5
The About page is underrated
Looking only at citations to the tracked brands’ own sites (479 citations), About, team and bio pages were the second most-cited page type, at 16.7%. Only program and package pages (18.6%) beat them, and they edged out product pages (14.0%) and homepages (9.6%). About pages were cited in 7 of the 8 industries. In financial services, they accounted for 17 of the tracked brand’s 23 citations.
That makes sense. When a buyer asks “who is the most trusted firm for X,” the About page is where a model finds founding details, credentials, team expertise, scale and who the company serves.
Finding 6
Crowded categories are harder, and fragmented ones are up for grabs
The number of brands an engine names per answer ranged from 3.2 to 6.8, depending on the industry. In programs where answers listed more brands, answers also carried more citations (r = 0.76), and the tracked brand’s visibility tended to be lower (r = −0.68 with brands per answer, −0.77 with citations per answer).
Category crowding and tracked brand visibility
Sorted by brands named per answer. “Different #1 brands” counts how many distinct brands took the top slot across the program’s prompts.
| Industry (program) | Brands named per answer | Citations per answer | Tracked brand visibility | Different #1 brands |
|---|---|---|---|---|
| Education C | 3.2 | 2.8 | 56% | 11 |
| Consumer E-commerce A | 3.2 | 2.8 | 63% | 13 |
| B2B Software B | 3.5 | 3.4 | 48% | 13 |
| Education A | 4.0 | 2.5 | 87% | 9 |
| Education B | 4.0 | 3.2 | 96% | 8 |
| B2B E-commerce | 4.4 | 2.8 | 53% | 18 |
| B2B Software A* | 4.6 | 7.4 | 59% | 35 |
| Professional Services | 5.2 | 8.3 | 8% | 16 |
| Travel & Hospitality | 5.3 | 6.7 | 35% | 30 |
| Consumer E-commerce B | 5.4 | 5.9 | 39% | 17 |
| Real Estate | 5.5 | 7.2 | 5% | 29 |
| Financial Services A | 6.6 | 6.7 | 20% | 20 |
| Financial Services B | 6.8 | 6.6 | 23% | 16 |
*B2B Software A includes four engines (128 answers). Other programs have 60–80 ChatGPT answers each. Visibility reflects each tracked brand’s own strength as well as category crowding, so treat the correlations as directional.
Fragmentation is the other side of the story. In real estate, travel and B2B Software A, 29 to 35 different brands earned a #1 spot. Nobody owns those categories in AI search yet. Financial services and professional services answers were crowded and anchored by a few large, well-known firms (3–4 brands each appeared in at least half of the answers), so specialists had to compete for the lower slots.
Finding 7
Who’s asking matters more than how they ask
Each program tracks the same product category through several buyer personas. Within a single program, the tracked brand’s visibility swung by as much as 80 percentage points depending on the persona, and 5 of the 13 programs swung by 60 points or more.
Range of tracked brand visibility across buyer personas
Lowest to highest persona-level visibility within each program (each persona ≈ 10 prompts)
Prompt phrasing mattered far less. We sorted all 948 prompts by intent. Averaged within each program, tracked brand visibility varied by only about 7 points across prompt types. Nearly half of real buyer prompts were feature- or fit-specific (“which providers offer X for Y”), not generic “best of” questions. Trust and reputation prompts pulled in slightly more sources per answer.
Prompt types: share, citations and visibility
948 prompts classified by intent (1,044 answers)
| Prompt type | Share of prompts | Citations per answer | Tracked brand visibility‡ |
|---|---|---|---|
| Feature / fit-specific | 49% | 4.9 | 44% |
| "Best / top" lists | 27% | 5.3 | 47% |
| Price / value | 12% | 4.7 | 51% |
| Trust / reputation | 6% | 5.5 | 47% |
| How-to / informational† | 5% | 7.6 | n/a |
†How-to prompts appeared in only one program, so visibility isn’t comparable. ‡Average of each program’s visibility rate for that prompt type, which controls for industry mix. The average prompt was about 17 words long, far longer and more specific than a typical Google query.
Finding 8
AI engines barely agree with each other
In the one program tracked across four engines, we ran the same 32 B2B software prompts through each of them. The results were strikingly different:
- Brand shortlists overlapped by only 7–18% between any two engines (measured as shared brands ÷ all brands named).
- The engines never agreed on the #1 brand. All four picked the same top recommendation on 0 of 32 prompts.
- The tracked brand’s visibility ranged from 16% on ChatGPT’s smaller model to 91% on Claude.
- Perplexity cited 19.4 sources per answer, 6x as many as ChatGPT’s smaller model. Claude’s answers in this dataset came with no source links at all, so they reflect what the model already knows.
Same 32 prompts, four engines
B2B software program, one answer per prompt per engine
Perplexity’s citations were also the most diverse: 183 unique domains, with 15% going to thin listicle and “statistics” sites, 15% to other blogs, 6% to forums and 5% to video.
Finding 9
How #1 recommendations are written
For every recommendation, the engine writes a short explanation of why it’s on the list. We compared 4,872 of these blurbs by rank. The #1 pick reads like a confident, specific match. Lower picks read like hedged alternates.
Language in recommendation blurbs, by rank
% of recommendation blurbs at each rank containing the element
Models also got specific about products. 70% of recommendations named a particular product, program, package or service line, not just the company.
Finding 10
AI engines reuse a small set of canonical URLs, and favor fresh ones
Citations concentrate on a few URLs. 32% of all citations went to URLs that were cited five or more times, and in each program the top 10 URLs captured 15–46% of all citations. Once a page becomes the go-to source for a topic, it keeps getting cited.
Few cited URLs had a date in the path (214 of 5,376). Those that did skewed heavily toward recent content: 51% were from 2025 or 2026.
Year in cited URLs
Share of the 214 cited URLs that include a year in the path
Finding 11
Third-party sources depend on the industry
Only about 16% of citations went to third-party sites, but which ones matter depends entirely on the vertical:
- Professional services: government registries and authorization marketplaces (8%), standards bodies and professional associations (5%), review and directory profiles (5%).
- Travel & hospitality: review and directory platforms (6%) and official tourism boards (6%).
- Consumer e-commerce (Program A): big-box marketplaces and retailers (18%).
- Real estate: city, county and licensing-board sites (4%), the local chamber of commerce, and listing portals.
- B2B software (Program A): thin listicle/”statistics” sites (11%), niche blogs (13%), industry associations (4%), forums (4%) and video (3%), driven mostly by Perplexity and Gemini.
- Education (Program C): university pages (5%).
- Financial services: almost entirely firm websites (96–97%), including individual advisor-team pages hosted on large firms’ domains.
Bonus
Oddities worth knowing
- Lookalike domains get cited. Three ChatGPT citations in consumer e-commerce pointed to lookalike domains that used real retailers’ names on unusual top-level domains. Monitor citations for impersonators of your brand.
- Some answers cite nothing. 9% of answers from ChatGPT’s smaller model had no citations, which means the model answered from memory. Your brand’s long-term web footprint still shapes those answers.
- Listicle networks punch above their weight. In B2B software, a handful of low-authority, templated “top tools” and “statistics” sites earned more citations than every news publication combined.
- Engines cite what they recommend, but not always reputable versions of it. A foreign-language site and templated “statistics” pages appeared among Perplexity’s sources for niche software questions.
What your website needs for AI search: a checklist
Based on the patterns above, these are the content priorities we recommend:
- Rebuild the About page as evidence. Founding year, scale, credentials, awards, team bios and a plain statement of who you serve best. It was the second most-cited page type on tracked brands’ sites.
- Create one page per offering and per audience. Service, program and product pages for each offering, plus audience or use-case pages (“for healthcare,” “for families,” “for first-time buyers”). These match the persona-specific prompts buyers actually type.
- Put decision criteria on commercial pages. Answer the criteria buyers ask about (pricing approach, support, turnaround, flexibility, credentials, what’s included) on the pages models cite, not only on the blog.
- State facts in citable form. Use numbers, named offerings and explicit “best for” statements. #1 recommendations were nearly 4x as likely to include a concrete number.
- Build your industry’s citation page. Product and collection pages (e-commerce), program details (education), service pages (professional and financial services), agent bios and area guides (real estate), sample itineraries (travel).
- Publish FAQ and policy pages. Shipping, installation, returns, financing and onboarding pages earned 12–13% of company-site citations in e-commerce and software.
- Keep key pages fresh and stable. Update and date important content, and keep URLs stable so pages can become canonical sources.
- Complete the third-party profiles your industry uses. Registries, associations, review platforms, marketplaces, tourism or licensing bodies, whichever the engines cite in your vertical.
- Make sure AI crawlers can reach you. Recommendations track retrievability, so check robots rules, rendering, page speed and internal linking to your money pages.
- Track by engine and by persona. One blended visibility score hides 80-point persona swings and engines that rarely agree.
Methodology & limitations
- Data source: exports from an AI-visibility tracking platform covering 13 tracking programs (12 brands) in 2026. Each export includes the prompt, persona, engine, every recommended brand with its rank and explanation, and every cited URL.
- Scale: 1,044 answers to 948 unique prompts written for 96 personas, with 4,872 brand recommendations (861 unique brands) and 5,376 citations (924 unique domains).
- Engines: ChatGPT (948 answers across GPT-5.4, GPT-5.4 mini and GPT-5.6), plus Gemini, Perplexity and Claude (32 answers each, one program only).
- Classification: Cited domains were labeled “company website” if the domain belongs to a brand recommended anywhere in the dataset. Remaining domains were hand-classified into site types. Page types were assigned from URL patterns. Prompt types were assigned by keyword rules, with one primary type per prompt.
- Anonymization: No tracked brand, competitor or cited-site names are reported. Results appear only at the broad-industry level.
- Limitations: The prompts and personas were designed for each brand’s tracking program, not sampled randomly from real users. Each answer is a single snapshot. The cross-engine comparison is limited to 32 prompts in one category. Persona-level figures rest on about 10 prompts each. All relationships here are correlations, not proof of cause.
FAQ
What types of websites does ChatGPT cite most?
In our data, 84% of citations went to the websites of the companies being recommended. For ChatGPT specifically, 82–97% of citations went to company sites, depending on the model version. News sites, forums, video and encyclopedias together accounted for under 3% of citations.
Do blog posts help you get recommended by AI search?
Less than most people assume. Blog posts made up only 6.9% of citations to company websites. Commercial pages (homepages, service, product, program and About pages) earned most citations. In-depth articles still matter in some industries, such as professional services, where they were 19% of company-site citations.
Does being cited by an AI engine mean you’ll be recommended?
Mostly, yes. When a tracked brand’s own site was cited, it was recommended 96% of the time and ranked #1 in 47% of answers. When it wasn’t cited, it was recommended only 18% of the time. Engines tend to cite the brands they’re already recommending, so the relationship runs both ways.
Do ChatGPT, Gemini, Perplexity and Claude recommend the same brands?
Rarely. On the same 32 prompts, any two engines’ brand lists overlapped by only 7–18%, and all four never agreed on the #1 brand. Perplexity cited about 19 sources per answer, while ChatGPT’s smaller model cited about 3.
Which page should I optimize first for AI visibility?
Start with your industry’s “citation page”: product pages for e-commerce, program pages for education, service pages for professional and financial services, agent bios for real estate, sample itineraries for travel. Then strengthen your About page, which was the second most-cited page type on tracked brands’ own sites.
Want to know what AI search says about your brand?
Kern Media tracks brand visibility across ChatGPT, Gemini, Perplexity, Claude and Google AI Mode by persona and by engine, then builds the content that earns the citation.
