Which Sources AI Engines Share, and Which They Keep to Themselves

The finding

Three AI engines were asked the identical 30 buying questions, daily, over the same 16 days. Their answers cite 16,289 sources across 1,431 distinct domains - and the two groups divide up very unevenly.

81 domains - 5.7% of them - carry 46.8% of every citation. Meanwhile 1,132 domains (79%) were cited by exactly one engine, and together they account for only 24.5% of citations.

Domains and citations, by how many of the three engines cited them.

Cited by	Domains	% of domains	Citations	% of citations
1 engine	1,132	79.1%	3,998	24.5%
2 engines	218	15.2%	4,672	28.7%
All 3 engines	81	5.7%	7,619	46.8%
Total	1,431	100.0%	16,289	100.0%

So there is a small shared shortlist that every engine uses, and then a very large tail that each engine more or less keeps to itself. Past the shortlist, whether an engine cites a given site looks close to independent engine to engine.

How wide each engine goes

Engine	Source citations	Distinct domains	Cited by this engine only	Citations per domain
ChatGPT	3,917	776	597	5.0
Perplexity	7,693	617	338	12.5
Gemini	4,679	418	197	11.2

ChatGPT spreads widest and agrees least: 597 of the 776 domains it cited were cited by neither of the other two engines. Perplexity made nearly twice as many source citations as ChatGPT from fewer domains - about 12.5 per domain against ChatGPT at 5.0 - so it repeats its sources more. The three exclusive counts (597 + 338 + 197 = 1,132) are exactly the one-engine bucket above.

The shared shortlist: all 81 domains cited by every engine

These 81 domains carry 7,619 of the 16,289 citations. Note that this list includes brand-owned domains (hubspot.com, shopify.com, monday.com, asana.com, nike.com) and competitors in the tracked categories, not only third-party publishers.

Domain	Citations	Domain	Citations
reddit.com	949	marathonhandbook.com	43
tryprofound.com	546	visible.seranking.com	43
techradar.com	532	celoxis.com	40
runrepeat.com	375	siftly.ai	40
shopify.com	358	pipedrive.com	37
forbes.com	311	campaignmonitor.com	36
monday.com	269	keyword.com	35
clickup.com	253	nutshell.com	34
rtings.com	248	irunfar.com	34
zapier.com	231	atlassian.com	33
blog.hubspot.com	176	believeintherun.com	32
hubspot.com	176	adidas.com	32
salesforce.com	143	ahrefs.com	31
zoho.com	143	nicklafferty.com	28
thedigitalprojectmanager.com	140	scrunch.com	28
technologyadvice.com	122	doctorsofrunning.com	27
peec.ai	109	writesonic.com	27
asana.com	101	hibbett.com	26
wrike.com	99	creatio.com	26
teamwork.com	94	digitalapplied.com	25
bigcommerce.com	92	outdoorgearlab.com	25
learn.g2.com	89	crm.org	25
g2.com	86	stripe.com	21
llmpulse.ai	71	theretailexec.com	19
woocommerce.com	71	otterly.ai	19
wix.com	68	meltwater.com	18
business.com	67	searchable.com	18
techrepublic.com	67	ecommerceparadise.com	15
semrush.com	67	fastlancer.org	15
litextension.com	66	cloro.dev	13
airtable.com	66	autods.com	10
runtothefinish.com	66	atwix.com	10
nike.com	65	supwell.com	9
theruntesters.com	64	visiblie.com	9
capterra.com	59	wizcommerce.com	9
expertmarket.com	56	insightly.com	7
blog.salesflare.com	55	beomniscient.com	7
experte.com	52	searcheye.io	6
wise.com	48	linkeddit.com	3
ecommerce-platforms.com	48	decisioncircuit.com	3
-	-	metorik.com	3

Method

• Universe: 5 brands, one per category (CRM, project management, ecommerce, AI-visibility tools, running shoes), 6 buying questions each - 30 questions.
• Engines: ChatGPT, Gemini and Perplexity - the three that ran all 30 questions, so they are the only ones directly comparable to each other.
• Window: 30 September to 15 October 2026 (UTC) - 16 days.
• Sample: 1,502 answers (ChatGPT 505, Perplexity 513, Gemini 484) - 15 to 20 answers per question, because a few were re-run by hand during setup.
• Measure: citations the engines made to a publication (classified as info_site) - 16,289 in total across 1,431 distinct domains.
• Domain: the host, with www. stripped, so blog.hubspot.com and hubspot.com count separately.

Limitations

• It includes brand-owned domains. hubspot.com, shopify.com, monday.com, asana.com and nike.com all rank high in this data. "Source" here includes the brands themselves, not only third-party publishers. If you want publishers only, filter the tracked and competitor domains out.
• Hosts, not pages. Subdomains count separately, and counts are raw citation rows rather than deduped by page. One domain can be cited hundreds of times.
• Three engines only. This is a three-engine comparison. It says nothing about the other engines available in the platform, which did not run this question set.
• A sample, not a census. AI answers vary run to run. This is 16 days of a small brand set; a re-run moves these numbers.
• Our own data, not a neutral study. No paid placements were involved, and we make no claim that any of this causes citations.

Check it yourself

Every number on this page can be recomputed from the file that ships with it. /data/ai-citation-source-overlap-2026-10/dataset.json contains domainsAll: all 1,431 domains, how many of the three engines cited each one, the citations each engine made to it, and the total. Sum the citations column and you get 16,289. Count the rows where engines is 3 and you get the shortlist of 81. Sum the chatgpt, gemini and perplexity columns and you get 3,917, 4,679 and 7,693. That is the entire dataset, and being able to check the arithmetic without taking our word for it is the point of publishing it.

What you cannot do is re-run our monitoring. The query below reads our internal measurement database, which readers do not have access to, so treat it as the audit trail for how these numbers were produced rather than something you can execute yourself. It is here because a method you cannot inspect is just an assertion.

WITH pairs AS (
                  SELECT DISTINCT Platform,
                         LOWER(REPLACE(SUBSTRING_INDEX(SUBSTRING_INDEX(Url,'/',3),'/',-1),'www.','')) AS host
                  FROM tracker_citations
                  WHERE TenantId = 9001 AND CitationSourceType = 'info_site'
                    AND Platform IN ('ChatGPT','Gemini','Perplexity')
                    AND Date >= '2026-09-30' AND Date < '2026-10-16'
                ),
                per_host AS (SELECT host, COUNT(*) AS engines_citing FROM pairs GROUP BY host),
                cites AS (
                  SELECT LOWER(REPLACE(SUBSTRING_INDEX(SUBSTRING_INDEX(Url,'/',3),'/',-1),'www.','')) AS host,
                         COUNT(*) AS cites
                  FROM tracker_citations
                  WHERE TenantId = 9001 AND CitationSourceType = 'info_site'
                    AND Platform IN ('ChatGPT','Gemini','Perplexity')
                    AND Date >= '2026-09-30' AND Date < '2026-10-16'
                  GROUP BY host
                )
                SELECT p.engines_citing, COUNT(*) AS domains, SUM(c.cites) AS citations
                FROM per_host p JOIN cites c ON c.host = p.host
                GROUP BY p.engines_citing ORDER BY p.engines_citing;
                

If you do compute something similar, one trap is worth knowing: reduce to DISTINCT engine/host pairs before counting how many engines cite a host, otherwise a domain cited 400 times by one engine looks like 400 engines agreeing. Also note that the Date column in this schema holds UTC even though some MySQL clients render it shifted by the session time zone, which can make the last row look like 16 October when it is really 15 October 23:41 UTC.

Per-engine totals, the exclusivity split, the robustness check at higher citation thresholds, the complete 1,431-row table and the full query set are all in the machine-readable version at /data/ai-citation-source-overlap-2026-10/dataset.json.

Cite this

AEO Spotlight, Which Sources AI Engines Share, and Which They Keep to Themselves (30 September to 15 October 2026). Retrieved from https://aeospotlight.com/data/ai-citation-source-overlap-2026-10