ChatGPT Citation Change Log: It Skips 144 of the 500 Most-Cited Domains

ChatGPT cites 4,877 of the 22,179 domains observed in AI answers, the narrowest of all six engines — and one of the few rankings on this page that a collection gap cannot touch, because every other arm's count is either exact and larger or a floor already above it. The count is not the interesting part. Where it sits in the head of the index is: ChatGPT reaches 356 of the 500 most-cited domains in AI answers, the lowest of any of the six engines, while Perplexity reaches 489 and Google AI Overviews reaches 385. Head overlap is a presence count, so a collection gap can only lower it and ChatGPT's last place holds. The whole-universe width ranking does not hold, and it is withdrawn in the correction of September 25, 2026 below.

ChatGPT does not cite less than the other engines. It cites a different head.

How to read the exclusivity counts. An exclusive share computed across all cited domains counts each domain once whatever its citation frequency, so almost all of its weight sits on sources cited a single time. On the September 21 release, 46.2% of 22,377 cited domains were cited once in the whole window, and ChatGPT's exclusive pool of 2,741 domains falls to 12 once a source has to have been cited ten times — two engines have none left at all, and both of those are the Google surfaces whose collection was not continuous in the window, per the correction below. The exclusivity counts below are accurate and describe the once-cited tail; the finding about where ChatGPT gets its evidence is the head-overlap measurement, ranked by citation frequency, which this page leads on.

This page is a dated record of what ChatGPT cites and what moves. Every entry states the measurement, the release it came from, the primary source for any policy it references, and what a publisher can check. Entries are appended; nothing is rewritten. The measurements come from the Machine Relations Index, which observes six engines — ChatGPT, Claude, Gemini, Google AI Mode, Google AI Overviews and Perplexity — against a fixed basket of buying and research questions, and publishes a machine-readable release daily.

Correction, September 26, 2026 — Gemini is a third gapped arm

The September 25 correction below named two engines with dated collection gaps. There are three. Gemini has 22 blind days inside the scoring window — May 11 to 15, then scattered single days from August 11 to September 16 — carrying 442 eligible successful runs that produced no citations. The shape is why nine sweeps walked past it: intermittent single days rather than a continuous outage, so anyone hunting an outage found none. Three arms collected without a gap — ChatGPT, Claude and Perplexity — and three are gapped: Gemini, Google AI Mode and Google AI Overviews. (An earlier form of this note, posted the same day, named only ChatGPT and Perplexity. Claude's single flagged date carried one answer run and was an artifact of a zero-citation scan with no run floor, not an outage.) Every Gemini count on this page is therefore a floor, and every ranking that places another engine above Gemini is withdrawn on the same logic the September 25 correction applies to the two Google surfaces: a floor can rise past a fixed count, so it settles nothing. Rankings that place an engine below Gemini still hold, because that distance can only widen.

Correction, September 25, 2026

An earlier version of this log ranked all six engines against each other by how many distinct domains each cited, and described Google AI Overviews as citing "less than half as many domains overall". Every whole-universe width ranking on this page that places either Google surface against another engine is withdrawn. The Index's own collection recorded zero citations from Google AI Overviews on any monitored query between August 15 and September 24, 2026 — 41 of the last 132 days, covering every day of this entry's window from August 15 onward — and zero from Google AI Mode between September 14 and September 24. The cause was a defect in the Index's Google collection path, fixed September 25, 2026: the prior collector graded a response as a success whether or not it carried content. A distinct-domain pool only accumulates, so missing collection days can remove domains from an engine's set but never add them. Google AI Mode's 7,441 and Google AI Overviews' 2,279 are floors, not measurements.

What survives, and why. Every head-overlap and presence count on this page — top 25, 100, 250 and 500; community platforms cited of 27; segment leaders cited of 85 — is a presence count, and a collection gap can only lower one. ChatGPT is last on all three at figures the gap did not touch, and the two Google surfaces can only be higher than shown, so ChatGPT's last place is if anything understated. The finding this page leads on is unaffected. What is withdrawn is the size-of-universe comparison, which ran in the other direction.

Entry 1 — September 18, 2026: the baseline

Release mri_score_v2.0, contract machine_relations_index_public_view_v2.0, generated September 18, 2026. Window May 10 to September 18, 125 days, 15,782 observed answer runs, 124,397 citation events, 22,179 cited domains. A subject category paired with a question type forms a stratum, and a stratum publishes a rate only after clearing an evidence floor of 10 observed runs across 7 distinct dates; 85 of the 151 measurable strata are published today.

The Index records, for each domain, which engines cited it at least once in the window. It does not publish a per-engine citation rate, so everything below is presence and counts, not per-engine rates.

ChatGPT's citation set is narrow at the top and wide at the tail

The exclusivity column below is a count of domains, not of citations, so read it with the note above: it describes the composition of each engine's tail, where most domains are cited once, and not the evidence behind what any engine actually leans on.

Engine	Domains cited	Cited by that engine only	Exclusive share
Perplexity	9,497	4,246	44.7%
Gemini	7,079	2,837	40.1%
Claude	5,138	1,431	27.9%
ChatGPT	4,877	2,708	55.5%
Google AI Mode †	7,441	3,382	45.5%
Google AI Overviews †	2,279	365	16.0%

† Floors, not measurements, and listed below the four rather than ranked among them. Both Google surfaces have collection gaps inside this window and a domain pool only accumulates, so each is at least this wide and no more can be said. See the correction of September 25, 2026 above.

Now hold that against the head of the index — the most-cited domains, the ones every visibility programme targets.

Engine	Top 25	Top 100	Top 250	Top 500
Perplexity	25	96	243	489
Claude	23	94	236	462
Gemini	23	89	227	455
Google AI Mode	24	93	227	448
Google AI Overviews	24	90	205	385
ChatGPT	22	88	203	356

ChatGPT has the lowest overlap with the index head of any of the six engines, at every band; a collection gap can only lower a presence count, so that last place holds and is if anything understated. Google AI Overviews reaches more of the top 500 than ChatGPT does. The comparison this sentence originally drew — that it does so on a domain pool less than half the size — is withdrawn: that pool is a floor, so its size cannot be compared.

That combination is the finding. An engine with a small, head-concentrated citation set is a narrow engine. An engine with a large, head-concentrated set is a broad one. ChatGPT is neither: it skips 144 of the 500 most-cited domains in the market while still reaching 4,877 domains in total, so its breadth sits below the head rather than inside it.

The community gap

The Index classifies cited domains by the role they play as sources. Only 3,556 of the 22,179 carry a classification; the rest are uncategorised, so shares within the classified set are the honest comparison, not shares of everything.

Twenty-seven classified domains are community and social platforms. That small set is where the engines diverge most sharply.

Engine	Community platforms cited (of 27)	Editorial share of its classified set
Google AI Mode	20	30.0%
Perplexity	18	32.5%
Gemini	15	31.4%
Claude	10	35.5%
Google AI Overviews	10	31.9%
ChatGPT	8	37.9%

ChatGPT cites the fewest community platforms of any engine and carries the highest editorial-publication share of its classified set. Whatever else is true of it, this engine leans to published media and away from forums harder than the other five.

Where that shows up: the leading source in 24 of 85 segments

For each of the 85 published strata, take the domain with the highest citation rate in that stratum — the source the engines cite most when buyers ask that category's question. Then ask which engines cite that domain anywhere in the window.

Engine	Cites the leading source, of 85 segments
Perplexity	85
Gemini	78
Google AI Mode	76
Google AI Overviews	74
Claude	64
ChatGPT	61

Perplexity cites the leading source in every published segment. ChatGPT misses 24 of them — and 21 of those 24 are the same domain: reddit.com, leading segments including is_x_worth in AI infrastructure (32.1%), consumer finance (31.8%) and consumer products (27.9%), and best_x in education and training (34.0%).

The other three are teramind.co on AI security and privacy how_choose (26.2%), unmannedsystemstechnology.com on deep-tech hardware top_list (28.3%), and erpfocus.com on enterprise software x_vs_y (20.2%).

The 47 head domains absent from ChatGPT, and what their robots files say

Forty-seven of the 250 most-cited domains do not appear in a single ChatGPT run in this window. Reading each publisher's own robots.txt on the same day is a cheap check: it establishes the publisher's stated policy. It never establishes cause, and several of these domains contradict it.

Rank	Domain	Class	Rate	robots.txt, read 2026-09-18
1	reddit.com	Community	12.75%	User-agent: * / Disallow: / — everything blocked
18	substack.com	Community	1.58%	No OpenAI agent named; no blanket block
23	facebook.com	Community	1.29%	GPTBot restricted to path rules, not a full block
42	nytimes.com	Editorial	0.89%	GPTBot, ChatGPT-User and OAI-SearchBot each Disallow: /
54	pcmag.com	Editorial	0.74%	All three OpenAI agents in a blocked group
107	cnet.com	Editorial	0.47%	All three OpenAI agents in a blocked group
108	wsj.com	Editorial	0.47%	No OpenAI agent rule
174	investopedia.com	Unclassified	0.36%	OpenAI agents allowed but for /thmb/

The New York Times blocks all three OpenAI crawlers by name and is absent, which is consistent — and it is also in active litigation with OpenAI and Microsoft, with filings through early September 2026. CNET and PCMag name the same three agents in blocked groups. OpenAI's crawler documentation distinguishes GPTBot (training), OAI-SearchBot (search index) and ChatGPT-User (user-triggered fetch), so a publisher can block one and permit another; these three block all of them.

Then the contradictions. Investopedia permits every OpenAI agent except an image path and is absent anyway. The Wall Street Journal names no OpenAI agent at all — its parent has a reported $250 million five-year agreement with OpenAI, the largest publisher deal on record — and is absent anyway. Substack and Facebook impose no relevant block and are absent.

Permission is necessary. It is plainly not sufficient.

The Reddit inversion

Reddit is the most-cited domain in AI answers at 12.75%, cited in 2,012 of 15,782 runs. Its robots.txt is eight lines and blocks every crawler: User-agent: *, Disallow: /, pointing at Reddit's Public Content Policy for access terms.

So the four engines that cite Reddit are not reaching it by crawling. Google reached it by contract — a licensing agreement reported at roughly $60 million a year — and OpenAI signed a comparable agreement, reported at around $70 million a year, in May 2024. Reddit has since sued Anthropic and pursued other platforms over unlicensed use.

The engine holding the licence is the one that does not cite the source. That is the inversion this log will keep watching.

Where the measurements disagree, and that stays on the record

Third-party panels do record ChatGPT citing Reddit. Treg's citation panel for the query "what sources does chatgpt cite", sampled September 18, 2026, returns 429 cited domains with reddit.com second at 204 mentions. And ZeroClick Labs reported in August 2026 that ChatGPT's visible Reddit citations fell roughly 86% between August 8 and August 17 while retrieval held steady — a display-layer change, on their reading, rather than a block.

Those are different instruments measuring different things over different prompt baskets, and the disagreement is the honest state of the record. The claim that survives all of them is narrow and worth stating precisely: across this basket of buying and research questions over 125 days, the Index observed no ChatGPT run citing reddit.com. Whether that is access, ranking or display is not settled by any public measurement, and this log will not pretend otherwise.

Entry 2 — September 22, 2026: reddit.com still absent, confirmed on a second, independent panel

Not a Machine Relations Index release — the next one is due September 25, 2026, and this entry does not wait for it. It reads the same central claim through a second instrument: AuthorityTech's own daily AI-visibility capture, 89 logged runs across 85 calendar days from June 24 to September 22, 2026 (four days were captured twice; duplicate-day reruns are collapsed to one observation per query per date), a tracked panel that held 30 to 35 buying and research queries on any given day and 77 distinct query ids across the full window as the registry evolved. 2,625 query-date observations returned a usable ChatGPT answer; the remainder errored and are excluded, not counted as absence.

Reddit: still zero. ChatGPT cited reddit.com in 0 of 2,625 observations. That confirms entry one's finding independently — a different instrument, a narrower and evolving query basket, a window that starts eight weeks before the Index baseline and now runs four days past it. Same result on both.

Two more watch-list items, unchanged. wsj.com and investopedia.com — entry one's "permitted and absent" cases — were also cited zero times each across the same 2,625 observations. No state change yet.

One first appearance. facebook.com was cited once, on September 5, 2026 — the panel's only ChatGPT citation of a platform named in entry one's community table. One occurrence in 85 days is noise until it repeats; it had not repeated by September 22.

New York Times litigation: still no ruling, more on the record. No ruling has been issued and court reporting does not expect one before 2027. Both sides filed summary-judgment motions September 4; a less-redacted version of the news plaintiffs' filing went public September 17 to 18, showing internal statements from OpenAI and Microsoft executives that the plaintiffs argue undercut the fair-use defense, and the U.S. Department of Justice filed a brief in early September supporting OpenAI and Microsoft. None of it changes access posture: nytimes.com's robots.txt still disallows all three OpenAI agents by name, and ChatGPT still cited it zero times in this panel.

What the same instrument shows about movement, which a single release cannot. An Index release records only whether a domain was ever cited across its whole window; it cannot show how often the answer to one question changes day to day. This panel can, on the same tracked questions asked daily. Comparing each query's ChatGPT-cited domain set to the same query the previous day, across every consecutive pair where both days returned a usable answer (2,464 pairs, drawn from the 77 query ids the panel carried at some point in its life): the cited domain set changes by a mean 53.5% from one day to the next. But when a domain is cited on both days, the exact URL ChatGPT cites for it is the same page 82.0% of the time (4,783 domain-retention pairs). ChatGPT swaps which sources it reaches for constantly; when it keeps a source, it keeps the specific page more often than not.

What a publisher can check: the robots.txt reads use the same links as entry one — reddit.com, wsj.com, investopedia.com, nytimes.com. The litigation record is public: Reuters, September 17 and RTÉ, September 18 both report the unsealed filings and confirm no ruling has issued. The 89-run panel itself is AuthorityTech's internal daily capture, not a public release; its method — a tracked query set answered once per engine per query per day, ChatGPT reached through OpenAI's web_search tool — is stated here in full because there is nowhere else to read it.

How to read this log

Each future entry compares a release against the one before it and reports what moved. The figures above are the fixed baseline for the September 18, 2026 release; the next entry reads the following weekly release against them. Four numbers are the ones to watch:

• Head overlap. 203 of the top 250, 356 of the top 500.
• Exclusive pool at an evidence floor. 12 domains cited only by ChatGPT and cited at least ten times, on the September 21 release. The unconditioned 55.5% share is tail composition and is not tracked as a finding; see the note above.
• Community platforms. 8 of 27.
• Segment leaders cited. 61 of 85.

A change in any of them is an entry. A release with none is recorded as no change, because a log that only reports movement teaches readers to mistake quiet periods for missing data.

What a publisher should take from entry one

Being in the index head is not the same as being in ChatGPT. One hundred and forty-four of the 500 most-cited domains in AI answers do not appear in a ChatGPT run in this window, and the reasons are not uniform: some blocked the crawlers, some signed agreements and are absent anyway, some did neither.

The practical reading is that per-engine presence is a measurement, not an inference. A brand tracking a single aggregate "AI visibility" number cannot see any of the above. Where the engines genuinely diverge is in the head, on sources with enough citations behind them to be a target: ChatGPT reaches 133 fewer of the top 500 than Perplexity does, and the engine most people mean when they say AI search is the one whose head overlaps the market's least.

Watch list for the next entry

• Whether ChatGPT's head overlap moves in either direction at the next release.
• Whether reddit.com appears in any ChatGPT run, and whether the OpenAI agreement is renewed, expanded or lapses.
• Whether the Wall Street Journal and Investopedia — permitted and absent — change state.
• Whether the New York Times litigation produces a ruling that changes either party's access posture.
• Whether TechRadar, absent from both Google Search surfaces with no relevant robots rule, resolves or persists.

Watch list for entry 3

• Whether ChatGPT's head overlap moves in either direction at the September 25, 2026 Index release — still open.
• Whether reddit.com appears in any ChatGPT run on either instrument, and whether the OpenAI-Reddit agreement is renewed, expanded or lapses — confirmed absent through entry 2, still open.
• Whether wsj.com or investopedia.com change state — confirmed unchanged through entry 2, still open.
• Whether the New York Times litigation produces a ruling — none issued as of entry 2; summary-judgment briefing and newly unsealed filings are on the record, no ruling expected before 2027.
• Whether facebook.com's single September 5 citation repeats — it had not by entry 2.

Frequently asked questions

Does this mean ChatGPT cites fewer sources than other engines? No. It cites 4,877 distinct domains, the narrowest of the six. That ranking survives every collection gap in the window: Claude's and Perplexity's counts are exact and both are larger, and Gemini's, Google AI Mode's and Google AI Overviews' are floors that can only rise. Google AI Overviews' recorded 2,279 sits below ChatGPT's count, but it is a floor rather than a measurement and the two are not compared here. The distinctive thing is the composition: the lowest overlap with the most-cited domains in the market, reaching 356 of the top 500 against Perplexity's 489.

Is a blocked robots.txt the reason a domain is absent from ChatGPT? Not reliably. It matches in several cases — the New York Times, CNET, PCMag — and fails in others, including the Wall Street Journal and Investopedia, both of which permit OpenAI's agents and are still absent. A robots.txt file records a publisher's stated policy, and nothing more.

Why does a third-party tool show ChatGPT citing Reddit when this says it does not? Different prompt baskets, different sampling windows, and a distinction between retrieval and displayed citation. Both readings are on this page, unreconciled, because no public measurement settles it.

What does "leading source in a segment" mean? A segment is one subject category paired with one question type. Its leading source is the domain with the highest citation rate in that segment across all six engines. ChatGPT cites that domain, anywhere in the window, in 61 of the 85 published segments.

How often is this updated? On each weekly Index release, and on any dated engine or policy change with a primary source.

Change log

Date	Entry	Release
2026-09-18	Baseline established: 4,877 domains cited, 55.5% exclusive, 203 of top 250, 8 of 27 community platforms, 61 of 85 segment leaders	mri_score_v2.0, generated 2026-09-18
2026-09-22	Method: exclusivity read conditioned on citation frequency; page leads on head overlap (356 of the top 500). Counts unchanged	mri_score_v2.0, generated 2026-09-21
2026-09-22	Entry 2: reddit.com, wsj.com, investopedia.com confirmed absent from ChatGPT across an independent 89-run daily panel (2,625 observations); day-over-day cited-domain churn 53.5%, URL retention when a domain is retained 82.0%; NYT litigation still without a ruling	AuthorityTech daily AI-visibility capture, 2026-06-24 to 2026-09-22 (internal panel, not an Index release)

Related: the per-engine source map of the top 25 domains, which covers all six engines for the same release, and the Machine Relations Index Report for September 18, 2026. Reddit's full per-segment record is on its Index domain profile.

Updated 2026-09-23: figures reflect the current Machine Relations Index methodology; see the MRI methodology and update log.