The tail is most of the domain list, and it is not most of the citations. In a six-engine panel running from May to September 2026, 46.9% of 23,280 cited domains were seen exactly once. The head, on the wide counts, is a few percent per domain. Promptwatch's June 2026 ChatGPT window had no domain above 4%. A circulating claim that 15 websites own 68% of citations does not survive those panels. Page concentration is a different fact: the top 1% of cited pages, in one B2B panel, earn 30% of citations.
- 46.9% of domains were observed once in the Machine Relations window, 10 May to 23 September 2026. That is a share of domains. It is not a share of citations.
- The top 10 domains inside a category hold 7.0% to 24.2% of that category. Industrial is the low end. Family Software is the high end. 86.6% of domains appear in exactly one category.
- The top 1% of pages earn 30% of citations in Resonance's B2B technology panel. Most pages are cited once. The exact once-share of pages, and the page count behind the 1%, were not published.
- 15 sites do not own 68%. The page that prints 68% mixes ranges from different denominators. The wide indexes put the biggest domains in the single digits of all citations.
- Analyze.ai's top-10 shares are from a post that was not fully fetched. They are included as a conflicting tracker, graded partial, and they are not in the chart. Empact's 46.5% is a citation of Qvery, not a re-opened Qvery release.
How large the once-only tail is
Machine Relations Index release mri_score_v2.0, generated 23 September 2026, covers 130 days and six engines: ChatGPT, Claude, Gemini, Google AI Mode, Google AI Overviews, and Perplexity. The file has 23,280 cited domains, 16,475 answer runs, and 129,264 source observations. Partial It is a vendor index. The once-only rate is the cleanest domain-tail figure opened for this page.
The second bar is arithmetic. If 46.9% of domains were seen once, the other 53.1% were seen more often. The release does not print "53.1%" as its own bucket, so the chart says so. A domain observed once can still live on a site that is cited every day. The unit here is the domain, not the company and not the page.
Two other cuts from the same release keep the tail from being one number. 86.6% of cited domains appear in exactly one category. Distinct domains per category run from 236 in Local Services to 4,025 in Legacy News, with a median of 1,299. The top 10 domains' share of a category runs from 7.0% in Industrial to 24.2% in Family Software. Reaching half of a category takes 49 domains in Family Software and 304 in Legacy News. The 228 domains with 50 or more observations hold 25.0% of 115,066 citation observations in that cut. Raise the evidence floor to 50 observations and the single-category share falls from 86.6% to 14.0%. Most names are tourists. The names you see twice a week are not.
Empact, citing a March 2026 Qvery release, says 46.5% of cited domains appear only once. The Qvery page was not opened. The agreement with 46.9% is encouraging and it is still secondhand. Partial
How small the head is
Promptwatch, ChatGPT Search, 1 to 27 June 2026: Reddit was first at 3.71% of all citations. Wikipedia was second at 0.56%. No domain passed 4%. Reddit had been 6.11% in May. That is a share of every citation, which is why the percentages look nothing like a top-50 mention share. The month after this window is a different story, told on how the list changes. The June level is enough to kill the idea that one domain is the market.
| Panel | What the head holds | Denominator |
|---|---|---|
| Promptwatch, ChatGPT, June 2026 | Largest domain 3.71%. None over 4%. | Share of all citations |
| Profound, all tracked engines, through late Oct 2025 | Reddit 3.11%, YouTube 2.13%, Wikipedia 1.35%, Forbes 0.80% | Share of citations. Post is with Reddit. The 4 billion sample size is in secondary coverage, not in the extract of the opened page. |
| Trakkr, through 18 Aug 2026 | YouTube, English Wikipedia, and Reddit together 7.04% | 48,422,016 appearances. 529,165 domains. 86.75% of appearances sit outside the named source map. |
| Analyze.ai, indexed 4 Aug 2026 | Top 10 hold 12.2% on ChatGPT, 11.1% on Perplexity, 13.1% on AI Mode | Not fully fetched. Top 50 shares in the same extract: 29.2%, 30.1%, 34.0%. |
Inside a topic, the head gets bigger, which is what you would expect. Search Engine Journal's 24 March 2026 summary of a Kevin Indig study: 21,482 ChatGPT citation rows, 670 domains, 2,344 URLs, 127 prompts. The top 10 domains take 46% of citations in a topic, and the top 30 take 67%. Education's top 10% of domains capture 59.5%. CRM and SaaS has the highest one-hit page rate in that set, 84.7%. A Search Engine Land write-up the same week, of a Gauge study on product-comparison topics, prints the same 46% and 67% from about 98,000 citation rows. The samples as described are not identical, so they are not averaged. Both say the same kind of thing. Global rank and within-topic rank are different tools. Neither names a universal top 15.
Pages concentrate harder than domains
Resonance, on the same B2B technology panel as the SaaS split, says the top 1% of cited pages take 30% of citations, and that most pages are cited only once. They audited about 190,000 cited pages. They do not say how many distinct pages sit under the 1.3 million citations, so "top 1%" cannot be turned into a count. They do not publish the exact share of pages with one citation. "Most" is the word in the summary. Partial
That 30% can sit beside Promptwatch's "no domain above 4%." The heavy pages do not have to share a host. A long tail of domains and a short head of URLs are compatible. The page-type classification, which is still not a URL list, is on most-cited pages.
Green Flag's finance extract points the same way at page level. Among domains at DR 90 to 100 in that vertical slice, the median citations per cited page was 1.00. A median of one means most cited pages, even on the strongest domains in the slice, were cited once. The page was read from a long extract. Partial
Small files agree on the shape and disagree on the size, which is why they stay out of the chart. Foglift's benchmark, 2,583 domain-citations and 1,119 domains, puts the top 10 at 8.1%. DataDab's teardown of 100 prompts and 328 citations found 229 domains once each, and a top 10 at 17.4%. Both are too small to rank a market. Both are another sighting of "most names appear once."
The 68% claim
Everything-PR, 16 August 2026, says the top 15 domains capture roughly 68% of consolidated citation share, and that Reddit is about 40% on a "multi-engine aggregate." The table mixes ranges that are not the same measurement. Wikipedia is given as 26% to 48% of ChatGPT's top-10 citations. YouTube is given as about 19% of AI Overviews. Those are shares of a head, or of one engine's class, restated as if they were shares of everything. The page does not show a sample that would support 68% of all citations.
Authority Tech's answer, using the 23,280-domain index, is that the 15-site pitch list describes a pooled platform head and does not survive a category cut. This page agrees with that reading. It does not need a new crawl to do so. The contradiction is already in the published denominators.
Digital Applied's April 2026 study of 1,000 AI Overviews is the other concentration figure that gets quoted next to these. The top 1% of cited domains capture 47% of citations inside that file, the next 9% capture 31%, and the long-tail 90% share 22%. Wikipedia appeared in 24.3% of those overviews and Reddit in 21.6%. That is domain concentration inside 1,000 overviews. It is not the top 1% of pages, and it conflicts with panels in which the top three domains are about 7% of appearances. A thousand overviews can be real and still be the wrong denominator for "the web."
Which engines share a domain at all is a third question, already separated on domains cited by every engine. A domain that appears once, on one engine, in one category, is the ordinary row. The concordance protocol on this site has not measured it yet.
What this does not prove
Open question A long tail of domains does not mean your site has a long tail of citable pages, or that it does not. It does not say the once-only domains are low quality. It does not say the top three are unimportant. Seven percent of 48 million appearances is a large absolute number and a small share, and both descriptions are accurate. It does not give you the share of pages cited exactly once. Resonance said "most" and stopped.
How to use it
Do not build a pitch list of 15 sites and call it the citation market. Do not hear "46.9% appear once" as "citations are evenly spread." The citations bunch up on a small set of pages, scattered across more domains than a league table suggests, and inside a topic the top 10 can hold close to half. If a dashboard shows a domain at 40%, ask whether that is a share of all citations, a share of the top 50, or a share of answers. The three produce 40% for different reasons.
The series hub, which is the argument against pooling those reasons, is most-cited domains in AI search.
46.9% of cited domains in one six-engine panel appeared once. The top domain in a June 2026 ChatGPT window was 3.71% of citations. Fifteen sites do not own 68%.
Share on XVerification status
Machine Relations via Authority Tech, Promptwatch, Trakkr, Profound, Resonance, and the two Indig write-ups were opened or read from the notes' fetched extracts. Analyze.ai was not fully fetched. The 68% claim is quoted in order to show the denominator problem, not as a finding.
Namdev, R. (2026). The long tail of AI citations (v1). Retrieved from https://ritiknamdev.com/blog/ai-citation-long-tail Published under CC BY 4.0 — reuse freely with attribution.
Pages, as opposed to domains, are on most-cited pages. A fixed-prompt concordance study is registered, and not yet run, at cross-platform citation concordance.