Perplexity Citation Change Log: 490 of the Top 500 Domains, Recorded Six Sources Deep

A dated record of what Perplexity cites and what changes, read from the Machine Relations Index and AuthorityTech's own daily panel. Entry one: Perplexity reaches 490 of the 500 most-cited domains in AI answers — the widest head overlap of the three engines the Index collected without a gap, and above the recorded counts of the other five, though three of those — Gemini and the two Google surfaces — are floors that could in principle rise past it. It cites the leading source in all 93 published segments. The panel that tracks it day to day was recording a mean of 6 sources per answer against the roughly 20.8 the engine actually returns — disclosed here before any churn number built on it.

Published: 2026-09-22
Author: Paralax Editorial
Canonical: https://paralax.ai/blog/perplexity-citation-change-log-2026
Tags: ai-search, citations, measurement

Perplexity cites 9,986 of the 22,828 domains observed in AI answers — more distinct domains than any other engine measures against. That is the wide version of the story. The narrow version, the one that matters to a publisher deciding whether to chase this engine, is where those citations land: Perplexity reaches 490 of the 500 most-cited domains in the market, 244 of the top 250, and cites the single most-cited source in every one of the 93 published question segments. No other engine clears all three.

This log is a dated record of what Perplexity cites and what changes. Every entry states the measurement, the release or window it came from, the primary source, and what a publisher can check. Entries are appended; nothing is rewritten. Entry one draws on the Machine Relations Index, which observes six engines — Perplexity, ChatGPT, Claude, Gemini, Google AI Mode and Google AI Overviews — against a fixed basket of buying and research questions and publishes a machine-readable release daily. Entry one also covers a methodological finding about AuthorityTech's own daily panel that has to be on the record before this log uses that panel for anything: the panel's citation log keeps only the first six sources per answer, and it caught Perplexity worse than any other engine it tracks.

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. ChatGPT, Claude and Perplexity are the arms that collected without a gap. One Claude date was flagged by the same scan and carries exactly one answer run against it, which is a scan artifact rather than a collection gap; Claude counts on this page are exact. 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

Every ranking on this page that places an engine above Google AI Mode or Google AI Overviews is withdrawn and restated over ChatGPT, Claude, Gemini and Perplexity, of which ChatGPT, Claude and Perplexity collected without a gap and Gemini did not. 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 — 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.

The direction is the whole of this correction. A presence count and a distinct-domain count only accumulate, so missing collection days can remove entries from a Google surface's count but never add them. Every Google AI Mode and Google AI Overviews figure on this page is therefore a floor: the true count is at least that and may be higher. So a claim that another engine sits below a Google surface still holds, because the gap can only widen that distance — while a claim that another engine sits above one does not, because the floor can rise past it. (This sentence, as published on September 25, read "Rankings among ChatGPT, Claude, Gemini and Perplexity are unaffected; all four collected continuously through this window." It is false — Gemini was not one of the continuously collected arms — and it is superseded by the September 26 note at the top of this page.)

Entry 1 — September 22, 2026: the baseline

Release mri_score_v2.0, contract machine_relations_index_public_view_v2.0, generated September 22, 2026. Window May 10 to September 22, 129 days, 16,307 observed answer runs, 34,720 cited-domain observations, 22,828 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; 93 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 — the same boundary this log's companion ChatGPT, Gemini and Claude entries run under.

490 of the top 500: the widest recorded head overlap, exactly measured

Engine	Top 25	Top 100	Top 250	Top 500
Perplexity	25	96	244	490
Claude	23	94	237	463
Gemini	23	89	227	452
Google AI Mode	24	91	227	444
Google AI Overviews	24	89	203	378
ChatGPT	22	88	203	361

Perplexity misses 10 of the top 500 most-cited domains in the market. Every other engine misses at least 37 on the counts the Index recorded, and ChatGPT, the narrowest, misses 139 — a count no collection gap touched. Perplexity's lead is stated over Claude and ChatGPT, the only other exactly measured arms. Gemini's and the two Google surfaces' misses are ceilings rather than measurements, so Perplexity is not ranked above any of those three. This is the head-overlap comparison this log's ChatGPT entry now leads on, because an unconditioned exclusive-domain share is mostly tail composition — sources cited once — and does not describe where an engine's evidence actually sits. Head overlap does, because it is ranked by citation frequency, not counted flat.

It also cites the leading source in every segment

For each of the 93 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 93 segments
Perplexity	93
Gemini	86
Google AI Mode	81
Google AI Overviews	79
Claude	72
ChatGPT	66

Perplexity has not missed a single segment leader across the measured window. Combined with the head-overlap table, the two published facts about Perplexity's composition point the same direction: whatever else it does, this engine reliably reaches the source the market already concentrates on.

Perplexity cites 9,986 domains, ahead of Claude's 5,326 and ChatGPT's 5,054 — the two other arms with gap-free collection, so both of those comparisons are exact against exact. The Index recorded 7,118 for Gemini, 7,441 for Google AI Mode and 2,279 for Google AI Overviews; all three are floors and none is ranked against Perplexity in either direction. Of Perplexity's 9,986, 4,560 — 45.7% — are cited by no other engine in the window. Read that number conditioned on citation frequency, the way this log reads every exclusivity figure, not the way it is usually reported: at an evidence floor of 10 citations, Perplexity's exclusive pool falls from 4,560 domains to 18. The other 4,542 are sources cited once or a handful of times, which is tail composition, not a claim about where Perplexity's evidence concentrates. This page states the floor-adjusted number in the same entry it reports the raw one.

Community platforms: 18 of 27, above both other exactly measured engines

Twenty-seven classified domains in the Index are community and social platforms — a small set where engines diverge most.

Engine	Community platforms cited (of 27)
Google AI Mode	20
Perplexity	18
Gemini	15
Claude	10
Google AI Overviews	10
ChatGPT	8

Of Perplexity's 1,618 classified citations (domains carrying a source-role tag; the other 8,368 domains it cites are uncategorised, so this is a share of the classified set, not of everything), 32.9% are editorial publications and 1.1% are community platforms. Perplexity sits below Google AI Mode's community reach, which the correction above leaves standing, and well above ChatGPT's.

The ten domains in the top 500 it does not cite

Perplexity is absent from 10 of the top 500 most-cited domains. Two are worth a robots.txt check because every other engine in the panel cites them.

Rank	Domain	Class	Cited by	robots.txt, read 2026-09-22
166	quora.com	Community	4 of the other 5 engines	`User-agent: PerplexityBot` / `Disallow: /` — Perplexity named and blocked
90	researchgate.net	Academic/government	All other 5 engines	`User-agent: *` / `Allow: /` — no block on any agent

Quora names PerplexityBot specifically and blocks it; that is a clean case of stated policy matching observed absence. ResearchGate blocks nothing — its robots.txt allows every user agent, including Perplexity's — and is absent from every Perplexity run in the window anyway. Perplexity's own crawler documentation distinguishes PerplexityBot, used to surface results in Perplexity's own search product, from Perplexity-User, triggered by a live user request; a publisher can permit one and block the other. Neither distinction explains ResearchGate. The remaining eight absent domains — itpro.com, tracxn.com, 6sense.com, pymnts.com, americanbar.org, pitchbook.com, infosecurity-magazine.com, prezly.com — carry no comparable contradiction and are not itemised here; this log will if any of them becomes load-bearing to a later entry.

What the daily panel was actually recording

AuthorityTech runs its own tracked panel alongside the Index: 35 fixed queries, five retrieval surfaces, once a day, the same instrument behind this log's ChatGPT entry two and the site's 89-day per-engine churn measurement. That page documents that the daily log stores the first six sources returned by each engine, a cap that affects engines unequally. It affects Perplexity most of all five.

Measured against the stored per-run answer artifacts, which preserve each provider's complete returned source list — available from September 1, 2026 onward, not for the full panel history — Perplexity returns a mean of 20.8 source URLs resolving to 15.3 distinct domains per answer. The daily log recorded a maximum of six. Perplexity's answers exceeded six sources on 350 of 350 runs in that window: every single one was truncated, against 91% of Gemini's runs, 16% of Claude's, 12% of ChatGPT's, and none of Google AI Mode's. Perplexity is the deepest answer of the five surfaces this panel tracks and the one the six-source cap distorted worst.

The practical consequence, recomputed rather than assumed small: day-over-day domain churn on Perplexity's complete source lists for the September 1–22 window (577 day pairs) is 20.0%, against 21.8% on the panel's six-source cap — a difference of 1.9 percentage points, in the direction the six-slot depth predicts, with the cross-engine ranking unchanged. The full recomputation table lives on the churn page linked above; this log will not restate the panel's older 89-day retention, drop-out, and core-domain figures as if they described the complete list, because the complete-list artifacts do not reach back past September 1 and the churn page says so explicitly. Any figure in this log drawn from the panel is labelled by which list produced it.

How to read this log

Each future entry compares a release or a panel window against the one before it and reports what moved. Four numbers are the ones to watch:

Head overlap. 244 of the top 250, 490 of the top 500.
Segment leaders cited. 93 of 93.
Exclusive pool at an evidence floor. 18 domains cited only by Perplexity and cited at least ten times, on the September 22 release. The unconditioned 45.7% share is tail composition and is not tracked as a finding.
Panel-recorded source depth. 15.3 distinct domains per answer on the complete list, against a six-item cap in the daily log; watch for the cap narrowing as the answer-evidence store's history lengthens past September 1.

A change in any of them is an entry. A release with none is recorded as no change.

What a publisher should take from entry one

Reaching the head of the index is not automatic just because an engine cites a lot of domains. ChatGPT cites 4,877 domains, well below Perplexity's 9,986, and still trails Perplexity by 129 domains in the top 500 (361 against 490). Perplexity's combination — the widest raw domain count and head overlap of the four continuously collected engines, and a perfect segment-leader record — means presence in Perplexity is close to a proxy for presence in the market's actual citation head. Among the top 250 most-cited domains, only one, researchgate.net, is absent from Perplexity despite carrying no stated crawler block; that is the case worth watching, not the nine other, lower-ranked absences.

The second finding is about measurement, not the engine. A dashboard, a competitor tracker, or an internal panel that caps a citation array anywhere below roughly 20 entries per answer will report a fraction of what Perplexity actually cites, understating both its reach and its churn. AuthorityTech's panel quantifies the effect by checking the complete artifact against the six-item log: the six-item log reaches its limit on 100% of Perplexity's runs. A publisher relying on any vendor's Perplexity citation count should ask what the cap is before trusting the number.

Watch list for entry 2

Whether Perplexity's head-overlap or segment-leader numbers move at the next Index release — the release updates daily, so entry 2 will report on genuine movement, not a fixed weekly cadence.
Whether quora.com's PerplexityBot block lifts, or whether researchgate.net's continued absence despite an open robots.txt develops an explanation.
Whether the complete per-run artifact history extends further back than September 1, letting the panel restate churn over a longer window than 22 days.
Whether any additional top-500 domain drops out of or into Perplexity's cited set.

Frequently asked questions

Does citing more domains than any other engine mean Perplexity is a worse filter? Not on this measurement. Domain count and head concentration are not in tension here: among the engines the Index collected without a gap, Perplexity cites the most domains and still reaches the deepest into the head of the index. A wide base and a covered head are not mutually exclusive. Gemini and the two Google surfaces are excluded from the comparison, per the corrections above; their counts are floors, and a floor cannot be ranked below a fixed number.

Why does the exclusive-share number in this entry not match a simpler "45.7% exclusive" headline? Because an unconditioned exclusive count is weighted almost entirely by domains cited once, as the ChatGPT entry of this log sets out. This entry states the floor-adjusted number, 18 domains, in the same breath as the raw one.

Is the six-source cap a property of Perplexity or of the measurement? A property of AuthorityTech's daily panel, not of the engine and not of the Machine Relations Index — the Index used for entry one's composition figures is not affected by this cap. It is stated here, before this log uses the daily panel for anything, so that every churn or stability number in a future entry is read at the right depth.