---
title: "Do AI Engines Cite the Same Sources? For One Question, 86% of Cited Domains Appear on Only One of Three Engines"
description: "In 856 buying questions, 86.1% of the domains ChatGPT, Claude and Perplexity cited were cited by one engine only, and 1.8% by all three."
canonical: "https://machinerelations.ai/research/do-ai-engines-cite-same-sources-2026"
---

# Do AI Engines Cite the Same Sources? For One Question, 86% of Cited Domains Appear on Only One of Three Engines

In 856 buying questions, 86.1% of the domains ChatGPT, Claude and Perplexity cited were cited by one engine only, and 1.8% by all three.

Canonical URL: https://machinerelations.ai/research/do-ai-engines-cite-same-sources-2026
Published: 2026-11-08
Research type: Index Analysis
Tags: citation-behavior, engine-comparison, source-overlap, machine-relations-index

## Current Index data

Updated from the Index release through **2026-11-08** (mri_score_v2.1); observation window 2026-05-10–2026-11-08. This card updates with the published dataset. The article below remains a dated analysis; these figures do not retroactively validate its claims.

- [techradar.com](https://machinerelations.ai/index/domains/techradar.com): 2.78% citation rate; 1,044 of 37,536 observed runs, tier A.
- [arxiv.org](https://machinerelations.ai/index/domains/arxiv.org): 2.47% citation rate; 927 of 37,536 observed runs, tier A.
- [guideflow.com](https://machinerelations.ai/index/domains/guideflow.com): 1.50% citation rate; 563 of 37,536 observed runs, tier A.
- Engines named here and observed in this dataset: ChatGPT, Claude, Gemini, Google AI Mode, Google AI Overviews, Perplexity. Coverage varies by engine; see the [release manifest](https://machinerelations.ai/data/mri-release-manifest.json) for returned runs and availability.

Citation presence measures observed source selection, not endorsement or causation. [Explore the Index](https://machinerelations.ai/index) · [Read the latest weekly report](https://machinerelations.ai/research/machine-relations-index-report).

## Source Body

For the same buying question, ChatGPT, Claude and Perplexity mostly cite different websites. In the Machine Relations Index answer ledger, 856 buying questions drew cited sources from all three engines between 2026-10-08 and 2026-11-07. Pooling each engine's cited domains per question gave 29,632 domain-question pairs. 86.1 percent were cited by one of the three engines only, 12.0 percent by two, and 1.8 percent by all three.

The disagreement is much larger than an engine's disagreement with itself. Two ChatGPT answers to the same question on different days share 31.8 percent of their cited domains (Jaccard overlap, the shared domains divided by all domains either cited). A ChatGPT answer and a Claude answer share 4.4 percent, and a ChatGPT and a Perplexity answer 4.4 percent. Claude and Perplexity are the closest pair at 9.4 percent.

This page measures what was cited. It does not say why the engines differ, and it does not say that optimising for one engine does or does not help on another.

## What was counted

- **Population.** Successful runs in the Index's buying-question panel from 2026-10-08 to 2026-11-07, on its 1,008 buying questions. Questions flagged as naming a brand in the prompt are excluded by rule; none of the 1,008 questions fell out in this window. The ledger export is dated 2026-11-08 08:10 UTC. The window starts after Perplexity's last change of source count (2026-10-07), so its answers are comparable across the month.
- **Engines.** ChatGPT, Claude, Gemini, Google AI Mode and Perplexity, called through their APIs. Google AI Overviews are left out. These are API routes, not the consumer apps. The ChatGPT route sends the buyer's sentence as its search ([how the engines search](/research/when-claude-and-chatgpt-search-the-web-for-buying-questions)).
- **Who answered.** Runs that cited nothing carry no sources and are left out of every pair. In this window Perplexity cited in 3,995 of 3,995 runs, ChatGPT in 3,662 of 3,979, Claude in 3,037 of 3,922, Google AI Mode in 448 of 634 and Gemini in 303 of 2,851. Gemini therefore enters 151 questions and Google AI Mode 128, against 856 to 1,008 for the other three, and their intervals are wide.
- **Domain and address.** The domain is the root domain recorded for each cited address. An address is compared without its scheme, "www.", query string, fragment or trailing slash.
- **Run view.** For each question, one run of engine A is paired with one run of engine B, every pair in the window is scored, and the scores are averaged per question. For the same engine, a pair is two runs on different dates. The table reports the mean over questions.
- **Pooled view.** For each question, an engine's cited domains from all of its runs in the window are pooled into one set, and the sets of two engines are compared.
- **Uncertainty.** 95 percent intervals resample the questions with replacement, 2,000 draws, fixed seed.

## One run against one run

Mean Jaccard overlap of cited domains between one run of each engine on the same question (95% interval over questions):

| Pair | Questions | Run pairs | Domain overlap |
|---|---:|---:|---:|
| ChatGPT with itself, different day | 961 | 5,300 | **31.8%** (30.8 to 33.0) |
| Claude with itself, different day | 812 | 4,233 | 25.0% (24.1 to 26.1) |
| Perplexity with itself, different day | 1,008 | 6,093 | **36.2%** (35.5 to 36.9) |
| Google AI Mode with itself | 113 | 695 | 23.9% (20.7 to 27.5) |
| Gemini with itself | 94 | 218 | 25.7% (22.9 to 28.6) |
| Claude with Perplexity | 874 | 12,308 | 9.4% (9.0 to 9.9) |
| Claude with Gemini | 150 | 1,149 | 7.9% (6.7 to 9.3) |
| Claude with Google AI Mode | 120 | 1,842 | 6.8% (5.6 to 8.1) |
| Google AI Mode with Perplexity | 128 | 2,135 | 6.0% (5.1 to 6.9) |
| Gemini with Perplexity | 151 | 1,209 | 7.3% (6.3 to 8.4) |
| ChatGPT with Claude | 856 | 11,257 | 4.4% (4.1 to 4.8) |
| ChatGPT with Perplexity | 989 | 14,809 | 4.4% (4.0 to 4.8) |
| ChatGPT with Google AI Mode | 128 | 1,891 | 3.8% (2.7 to 5.1) |
| ChatGPT with Gemini | 141 | 1,032 | 2.9% (2.2 to 3.5) |

(Gemini with Google AI Mode, 14 questions and 111 run pairs, is 9.8% with an interval of 6.8 to 13.6 and is too thin to read.)

Every engine overlaps with itself far more than with any other engine. ChatGPT's own overlap is 7.2 times its overlap with Claude. The same ordering holds for exact addresses: an engine's two runs on different days share 17.5 to 30.6 percent of their cited addresses, against 0.5 to 5.7 percent across engines (leaving out the 14-question Gemini and Google AI Mode pair). ChatGPT with Gemini is lowest at 0.5 percent (0.3 to 0.7).

Read the same-engine rows as the yardstick. Even one engine does not return the same set twice: Perplexity, the most repeatable, shares about a third of its domains between two days.

## Pooled over the month

Pooling a month of runs per question raises every overlap, because each engine's set grows. The ordering stays. Mean Jaccard overlap of cited domains (95% interval), and the share of one engine's pooled domains that the other also cited:

| Pair | Questions | Domain overlap | Share of A also cited by B | Share of B also cited by A |
|---|---:|---:|---:|---:|
| Claude (A) with Perplexity (B) | 874 | 13.5% (13.0 to 14.1) | 29.9% (28.7 to 31.1) | 19.9% (19.1 to 20.7) |
| ChatGPT (A) with Perplexity (B) | 989 | 6.2% (5.8 to 6.6) | 18.0% (16.9 to 19.1) | 9.1% (8.5 to 9.6) |
| ChatGPT (A) with Claude (B) | 856 | 6.1% (5.6 to 6.5) | 13.6% (12.6 to 14.7) | 10.6% (9.9 to 11.4) |
| Claude (A) with Gemini (B) | 150 | 10.6% (9.2 to 12.1) | 21.4% (18.6 to 24.4) | 18.2% (15.8 to 20.6) |
| ChatGPT (A) with Gemini (B) | 141 | 4.3% (3.4 to 5.1) | 10.0% (7.9 to 12.2) | 7.9% (6.4 to 9.4) |

A pooled set holds 10.1 domains for ChatGPT, 11.8 for Claude and 18.0 for Perplexity, on average. The larger Perplexity set explains why the share of ChatGPT's domains that Perplexity also cited (18.0 percent) is twice the share running the other way (9.1 percent): Perplexity has more domains to match. Overlap scores built on set size should be read with this in mind.

## How many engines cite a given source on one question

For the 856 questions where ChatGPT, Claude and Perplexity all cited, each domain is counted once per question, by the number of the three engines that cited it in the window:

| Engines citing the domain on that question | Domain-question pairs | Share |
|---|---:|---:|
| One | 25,524 | **86.1%** (85.7 to 86.6) |
| Two | 3,561 | 12.0% (11.6 to 12.4) |
| All three | 547 | 1.8% (1.7 to 2.0) |

At the level of exact addresses it is sparser: of 37,748 address-question pairs, 34,944 (92.6 percent) were cited by one engine, 2,629 (7.0 percent) by two and 175 (0.5 percent) by all three.

The domains cited by all three on the same question come from many places: arxiv.org and erpresearch.com lead with 11 questions each, then cdp.com (9), techradar.com (8) and guideflow.com (6). Nothing in this ledger says that a domain cited by all three is better evidence than one cited by one.

## This is not the same as "most cited sources overlap"

An earlier Index page found that among domains cited ten or more times across the whole index, 98.5 percent were cited by two or more engines ([multi-engine citation overlap](/research/multi-engine-ai-citation-overlap-data-2026), with its own collection-gap corrections). Both findings can hold. That page asks whether a domain cited often anywhere is cited by more than one engine: the big, general sources are. This page asks, for a single buying question, how much of the answer's source list the engines share, and the answer is not much. A domain cited on many questions by several engines is a different thing from a domain several engines cite on the same question.

## What this changes

- **Check each engine on its own.** One engine citing a source is weak evidence that another will. For a given question the chance that a ChatGPT-cited domain is also cited by Perplexity was 18.0 percent in the pooled view, and by Claude 13.6 percent.
- **Expect two answers from the same engine to differ.** A same-engine rerun on another day kept about a quarter to a third of the domains. A single check on a single day is a sample, not a result ([why repeat counts matter](/research/ai-citation-decay-how-brands-lose-visibility-over-time)).
- **Agreement across all three is a short list.** Fewer than 2 percent of domain-question pairs were cited by all three engines, so a source on every engine's list for one question is the exception.
- **Do not turn one engine's source list into a plan for another.** The ledger gives no test of whether a page that one engine cites is easier to get cited by another.

## What this page does not show

- **Why the engines differ.** Different search backends, models and prompts can all produce this. The ledger records the cited addresses, not the retrieval path.
- **Consumer apps.** These are API routes with search enabled, as configured by the Index. The apps may behave differently.
- **Google AI Overviews.** Not included.
- **Gemini and Google AI Mode.** Gemini cited in 10.6 percent of its runs in the window and Google AI Mode answered 128 questions, so their pairs are few and their intervals wide.
- **Whether a cited source was relied on.** The Index finds engines name only 17.6 percent of the sources they cite ([engines cite sources they do not name](/research/ai-engines-cite-sources-they-do-not-name-2026)).
- **Buyer demand.** The 1,008 questions are the Index's own panel, written for coverage of buying decisions, not a sample of what buyers type.
- **Trend.** One month. The page does not test whether overlap is rising or falling.

## Method note

Source: the Index answer ledger, daily export of 2026-11-08 08:10 UTC, read with DuckDB. Tables used: the run table (successful runs, the buying-question panel, run dates 2026-10-08 to 2026-11-07), the question table (buying-question panel) and the citation-events table (a non-empty cited URL and root domain). Questions with any event flagged as brand-prompted are excluded. Engines: ChatGPT, Claude, Gemini, Google AI Mode, Perplexity. A run's set is its distinct root domains, or its distinct normalised addresses (lower-cased, scheme and a leading "www." removed, query string and fragment removed, trailing slashes removed). Run view: for each question, every pair (run of engine A, run of engine B) with A before B alphabetically is scored with Jaccard overlap, and for the same engine every pair of runs on different dates with the earlier run id first; scores are averaged per question and then over questions. Pooled view: per question and engine, the union of the sets of all its runs in the window; Jaccard overlap and containment (the share of one set found in the other) per question, averaged over questions. Consensus: for the 856 questions with a pooled set from ChatGPT, Claude and Perplexity, each domain (or address) is counted once per question by the number of those three engines that cited it. Intervals resample questions with replacement, 2,000 draws for the overlap tables and 1,000 for the consensus shares, fixed seed; consensus shares are pooled counts, not averages of per-question shares.

## Frequently asked questions

### Do ChatGPT and Gemini use the same sources?

Mostly not, for the same question. In the Index's API runs, a ChatGPT answer and a Gemini answer to the same question shared 2.9 percent of their cited domains (2.2 to 3.5), over 141 questions where both cited. Pooled over a month the figure was 4.3 percent (3.4 to 5.1). Gemini cited in only 10.6 percent of its runs in this window, so the Gemini figures are the thinnest on the page.

### Is a source that ChatGPT cites likely to be cited by Perplexity too?

On the same question, 18.0 percent of the domains ChatGPT cited over the month were also cited by Perplexity (16.9 to 19.1). Perplexity cites more domains per question than ChatGPT, which raises that share.

### Which two engines agree most?

Claude and Perplexity. They share 9.4 percent of cited domains run to run (9.0 to 9.9) and 13.5 percent when a month of runs is pooled (13.0 to 14.1).

### How often do all engines cite the same source for one question?

Rarely. For ChatGPT, Claude and Perplexity on the same question, 1.8 percent of domain-question pairs were cited by all three, and 0.5 percent of address-question pairs.

### Does an engine give the same sources if you ask twice?

Not exactly. Two runs of the same engine on the same question on different days shared 25 to 36 percent of cited domains, from 25.0 percent for Claude to 36.2 percent for Perplexity.

## Attribution

This research is published by Machine Relations Research, the research program of machinerelations.ai — the public research initiative that publishes the glossary, research, evidence, and measurements for the Machine Relations discipline. Provenance and editorial standards: https://machinerelations.ai/about

## Attribution

This research is published by Machine Relations Research, the research program of machinerelations.ai — the public research initiative that publishes the glossary, research, evidence, and measurements for the Machine Relations discipline. Provenance and editorial standards: https://machinerelations.ai/about

## Machine-readable related links

### Related concepts

- [Machine Relations Index (MRI)](https://machinerelations.ai/glossary/machine-relations-index)
- [Machine Relations (MR)](https://machinerelations.ai/glossary/machine-relations)
- [Share of Citation](https://machinerelations.ai/glossary/share-of-citation)
- [Citation Decay](https://machinerelations.ai/glossary/citation-decay)

### Supporting research

- [ChatGPT's API Tags Every Link It Cites With utm_source=openai: 35,357 of 35,357. Claude, Gemini and Perplexity Almost Never Tag Theirs](https://machinerelations.ai/research/chatgpt-api-citations-utm-source-openai-2026)
- [Domains Cited on Only One AI Engine Are Mostly a Thin Sample](https://machinerelations.ai/research/how-to-measure-ai-search-visibility-brand-share-of-voice)
- [Fortune Business Insights Answer-Engine Citation Authority: 1.58% Citation Rate Across 16,898 Domains — Market Sizing Infrastructure Holds](https://machinerelations.ai/research/fortunebusinessinsights-answer-engine-citation-authority-mri)
- [A Rerun Shares 46% of AI Cited Sources, a Reworded Prompt 39%, a Broader or Narrower One 23%](https://machinerelations.ai/research/does-rewording-a-prompt-change-ai-cited-sources)

### Framework context

- [Machine Relations Index](https://machinerelations.ai/index)
- [Machine Relations Stack](https://machinerelations.ai/stack)
- [Evidence Base](https://machinerelations.ai/evidence)