What AI answers actually cite
We classified every source behind 17,801 answers from ChatGPT, Perplexity, Google AI Overviews, AI Mode and Gemini: who owns each cited site, what kind of page it is, and who wrote it.
From 17,801 AI answers collected with Firemark between 5 October and 5 November 2026. Published November 2026.
AI answers mostly cite company websites
Across five engines and 27 brands, company websites take most citations and no single site dominates. The Reddit effect people talk about shows up mainly in Google's AI Overviews.
- 69%of 41,942 citationsFinding 1 · Composition →
Company websites take most citations
Two thirds of cited pages sit on company websites, and 80% of those belong to firms that are neither the brand asked about nor a named competitor.
- 11%top 10 of 11,948 sitesFinding 2 · Concentration →
No site owns AI search
The 10 most-cited sites take 11% of citations, and 65% of cited sites appear in exactly one answer.
- 47%ChatGPT 1.9% → AI Overviews 47%Finding 3 · Engines →
Each engine cites a different web
Google AI Overviews cites social or video in 47% of answers. The Reddit story is mostly a Google story: community sites appear in 21% of its answers.
- 22%company pages 63%Finding 4 · Retrieval →
Engines read community pages, then cite something else
Perplexity cites 22% of the community pages it retrieves, against 63% of company pages.
- 38%listicles 25%Finding 5 · Page formats →
Product pages beat listicles
On unbranded “who should I use” questions, 38% of citations go to product pages and 25% to listicles.
- 2.5%of all citationsFinding 6 · Open publishing →
Parasite publishing is small
Pages companies wrote on other platforms are 8.7% of third-party citations, mostly their own videos and social posts.
What we measured, and how
Every brand's questions were asked once a week for five weeks. Shares use each brand's latest week (7,910 answers), so no question is counted twice.
How each citation was classified
Every cited source was coded on three things. Two separate AI coders agreed strongly on most of them.
Two thirds of what AI cites is company websites
Most citations go to company websites, and most of those go to firms other than the brand asked about or its named competitors.
Share of answer–page citations by the type of cited site, latest run per brand.
How we classified companies
A site counts as a vendor or brand site when the company behind it sells a product or service, whatever it publishes: product pages, blogs, guides, docs and help centres all count.
A company's profile on G2, LinkedIn or YouTube counts under that platform. Analyst firms count as editorial. Sites that rank products but sell none of their own count as listicle sites.
No site owns AI search
Citations spread across thousands of sites. Only a handful show up for most brands.
Share of all answer–page citations, latest run per brand.
The most-cited sites
| Site | Type | Answers citing it | Share of answers | Brands it was cited for |
|---|---|---|---|---|
| youtube.com | Social & video | 537 | 6.79% | 27 |
| reddit.com | Community | 405 | 5.12% | 25 |
| google.com | Other | 371 | 4.69% | 25 |
| linkedin.com | Social & video | 318 | 4.02% | 24 |
| facebook.com | Social & video | 231 | 2.92% | 25 |
| g2.com | Review & directory | 145 | 1.83% | 18 |
| nih.gov | Academic | 112 | 1.42% | 7 |
| texas.gov | Government | 105 | 1.33% | 5 |
| quora.com | Community | 104 | 1.31% | 20 |
| gartner.com | Review & directory | 102 | 1.29% | 14 |
| sourceforge.net | Review & directory | 97 | 1.23% | 12 |
| instagram.com | Social & video | 97 | 1.23% | 19 |
| guideflow.com | Vendor | 89 | 1.13% | 9 |
| amazon.com | Marketplace | 82 | 1.04% | 12 |
| hubspot.com | Vendor | 81 | 1.02% | 7 |
Sites that appear across five or more brands. Sites cited for fewer brands, including every brand's own site and its competitors, are left out of this list.
Each engine cites a different web
The five engines draw on different kinds of site. Google AI Overviews stands apart on social, video and community pages.
| ChatGPT | Perplexity | AI Overviews | AI Mode | Gemini | |
|---|---|---|---|---|---|
| Vendor & brand sites | 66.8 | 90.8 | 82.3 | 73.0 | 75.9 |
| Government | 25.4 | 12.0 | 8.5 | 7.2 | 6.5 |
| Social & video | 1.9 | 2.1 | 47.2 | 6.4 | 0.7 |
| Review sites & directories | 8.3 | 13.3 | 15.3 | 7.2 | 11.1 |
| Affiliate & listicle sites | 9.0 | 19.3 | 12.6 | 5.5 | 9.9 |
| Editorial & trade press | 5.8 | 15.3 | 14.5 | 5.6 | 7.1 |
| Marketplaces & app stores | 4.1 | 4.2 | 7.3 | 6.5 | 3.5 |
| Other | 6.2 | 5.6 | 6.1 | 4.4 | 4.5 |
| Academic & research | 4.8 | 5.4 | 6.3 | 3.3 | 3.4 |
| Associations & nonprofits | 6.4 | 6.6 | 5.6 | 3.4 | 4.4 |
| Community (Reddit, Quora, forums) | 0.5 | 0.5 | 21.0 | 7.7 | 4.0 |
| Open publishing (Medium, LinkedIn articles, wires) | 2.5 | 5.4 | 14.3 | 5.6 | 3.8 |
| Reference | 5.6 | 1.6 | 2.3 | 1.7 | 1.8 |
Share of each engine's answers that cite at least one page of each type. Bases: ChatGPT 1,620, Perplexity 1,598, AI Overviews 1,474, AI Mode 1,620, Gemini 1,598 answers. These are lower bounds: sites cited once and outside the coded sample are not counted.
AI reads Reddit, then cites something else
Engines retrieve far more community and social pages than they end up citing. Company and government pages get through most often.
| Type of site | Perplexity: cited of retrieved | ChatGPT: shown of pages read |
|---|---|---|
| Vendor & brand sites | 63% | 35% |
| Government | 64% | 31% |
| Social & video | 27% | 21% |
| Review sites & directories | 47% | 32% |
| Affiliate & listicle sites | 58% | 21% |
| Editorial & trade press | 57% | 21% |
| Marketplaces & app stores | 56% | 32% |
| Other | 66% | 31% |
| Academic & research | 62% | 21% |
| Associations & nonprofits | 74% | 34% |
| Community (Reddit, Quora, forums) | 22% | 10% |
| Open publishing (Medium, LinkedIn articles, wires) | 50% | 13% |
| Reference | 62% | 22% |
Perplexity retrieves 10 results per answer and marks the ones it uses. ChatGPT's search results show what it read before choosing its sources.
Product pages are cited more often than listicles
Product and feature pages are the most-cited format, including on unbranded “who should I use” questions.
Share of citations by page format, from a weighted sample of 1,760 fetched pages.
Companies writing on other platforms are a small share of citations
Pages companies write on other people's platforms are a small part of what AI cites, and most of them are the company's own videos and social posts.
Where company-written pages on third-party platforms sit, as a share of those citations.
Where we agree and disagree with other studies
The single-site rankings agree: YouTube and Reddit lead. Concentration differs, because long, situational B2B questions spread citations across thousands of company sites. Our AI Overviews figure for user-generated content falls inside the published academic range.
| Published study | Their figure | This study |
|---|---|---|
| Peec, most cited domains (Mar 2026) | Reddit, YouTube, LinkedIn, Wikipedia lead | YouTube, Reddit, Google, LinkedIn lead |
| Everything-PR citation index (May 2026) | Top 15 sites ≈ 68% of citations | Top 15 = 13.0% |
| arXiv 2605.14021; Hagar (Northwestern) 2026 | AI Overviews: user-generated content 14–24% | AI Overviews: community, social and open publishing ≈ 19% of citations |
| AirOps (Oct 2026) | Reddit leads in Perplexity and AI Mode | Community appears in 0.5% of Perplexity answers |
How every citation was classified
Each cited source was coded on three layers: the site, the page and the question.
Who controls the cited site
13 types, from company sites and government to Reddit-style communities
How content gets onto the site
Open, listing, contributor, paid, editorial or owner only
What kind of page was cited
17 formats, e.g. product page, listicle, explainer, how-to, video
Who wrote it
The company on its own site, a company on another platform, or an unaffiliated author
What the buyer asked for
Recommendation, comparison, evaluation, how-to, explainer, branded check
Where the question sits
TOFU, MOFU or BOFU, set when the questions were written
κ is agreement between two separate AI coders (Claude Sonnet 5.5 and Claude Opus 5.5) on 150–300 items per layer; above 0.8 is strong. A site counts once per answer.
How we measured it
1,620 buyer questions written to track 27 brands, mostly B2B, in US English. They are long and situational, the way buyers actually ask. They are not a sample of all AI search traffic, and no search volumes sit behind them.
ChatGPT: sources shown with the answer. Perplexity: results the answer marks with [n]. AI Overviews, AI Mode and Gemini: the listed references. A site counts once per answer.
Every site cited in two or more answers was coded against a written codebook; sites cited once are represented by a random sample of 1,000, weighted back. A second AI coder checked each layer separately (κ 0.76–0.95). Page format and authorship come from a weighted sample of 1,760 fetched pages.
90% of a random 500 cited links were live; under 1% returned 404. Removing any one brand moves headline shares by about a point.
Each brand's most recent run between 8 Oct and 5 Nov 2026. Engines change their behaviour often; treat every figure as dated.
Tenpoint Labs, a B2B content and AI-search agency. The answers come from Firemark, the AI-visibility tracker Tenpoint Labs built, which asks a fixed set of buyer questions of five AI engines every week and stores every answer and source.
Tenpoint Labs (2026). "What AI Answers Actually Cite." November 2026. Data: 17,801 AI answers collected 5 Oct–5 Nov 2026, 27 brands. tenpointlabs.com/research/what-ai-answers-cite
See how often AI answers name your brand
Get the free AI visibility audit