{"id":1509,"date":"2026-07-21T07:32:45","date_gmt":"2026-07-21T07:32:45","guid":{"rendered":"https:\/\/maxaeo.ai\/blog\/sources-cited-across-ai-engines\/"},"modified":"2026-07-21T07:32:45","modified_gmt":"2026-07-21T07:32:45","slug":"sources-cited-across-ai-engines","status":"publish","type":"post","link":"https:\/\/maxaeo.ai\/blog\/sources-cited-across-ai-engines\/","title":{"rendered":"Sources Cited Across AI Engines: An Eight-Platform Overlap Study"},"content":{"rendered":"

Sources cited across AI engines overlap far less than most reports suggest.<\/strong> In our twelve-week study of 607,318 citation instances from eight AI platforms, 76.4% of cited URLs appeared on exactly one engine, and only 1.9% were cited by six or more. That thin universal tier absorbed 29.3% of all citation volume<\/strong>.<\/p>\n

Most published research stops at "the engines barely agree." That is the least useful half of the finding. We ran the inverse study: isolate the URLs that six, seven or all eight engines independently chose, then work out what those pages share.<\/p>\n

This article reports the tier distribution, the engine-pair overlap matrix at URL level rather than domain level, the five page traits that separated universal URLs from single-engine ones, the four widely-recommended factors that predicted nothing, and one worked case of moving a page from two engines to seven.<\/p>\n

\"Overlap<\/figure>\n

What "citation overlap" means, and why URL-level differs from domain-level<\/h2>\n

Citation overlap is the share of cited sources that more than one AI engine independently selects for the same question.<\/strong> Measured at domain level, it counts two engines citing g2.com<\/em> as agreement. Measured at URL level, it only counts agreement when both engines cite the same page.<\/p>\n

That distinction is not academic. In our data, domain-level overlap ran 4.1\u00d7 higher than URL-level overlap on the same prompt set<\/strong>. Two engines can both "cite G2" while pulling entirely different category pages \u2014 so a domain-level number tells a marketer almost nothing about whether their specific page is working.<\/p>\n

Published studies overwhelmingly report the domain figure. Pairwise domain overlaps in the 10\u201317% range are consistent with our own domain-level results, but they imply far more consensus than the underlying URLs support. When you read any overlap statistic, the first question is which unit it was counted in.<\/p>\n

How the study was run<\/h2>\n

We tracked 1,140 prompts across eight AI platforms, re-running the full set weekly for twelve weeks between 3 February and 24 April 2026. That produced 109,440 answer captures and 607,318 raw citation instances, resolving to 81,542 unique canonical URLs<\/strong> after normalisation.<\/p>\n

Engines tracked:<\/strong> ChatGPT (web search mode), Google AI Overviews, Google AI Mode, Gemini, Perplexity, Claude (web search), Microsoft Copilot, Grok. AI Overviews and AI Mode were treated as separate surfaces because they returned different source sets for identical queries in 61% of paired captures.<\/p>\n

Prompt set:<\/strong> 1,140 buyer-style prompts across nine B2B software categories (CRM, data observability, HR tech, payments, cybersecurity, martech, dev tooling, analytics, ITSM), split between category discovery ("best X tools for Y"), comparison ("X vs Y for Z"), and validation ("is X any good for Y").<\/p>\n

Normalisation rules:<\/strong> URLs lowercased, stripped of tracking parameters and fragments, resolved through redirects, and collapsed to the canonical tag where one was declared. Homepage citations retained. Repeat citations of the same URL within a single answer counted once.<\/p>\n

Known limits:<\/strong> B2B software, US-English, desktop, unpersonalised, no logged-in accounts. Consumer, local and YMYL categories behave differently and these numbers should not be extended to them. Twelve weeks also predates any engine's next retrieval change \u2014 the cluster structure below is more durable than the exact percentages.<\/p>\n

How many sources are actually shared across engines<\/h2>\n

Agreement decays fast. The fully universal tier \u2014 pages cited by all eight engines \u2014 is 327 URLs out of 81,542.<\/p>\n\n\n\n\n\n\n\n\n\n
Engines citing the URL<\/th>\nUnique URLs<\/th>\nShare of URLs<\/th>\nShare of citation volume<\/th>\n<\/tr>\n<\/thead>\n
1 engine only<\/td>\n62,298<\/td>\n76.4%<\/td>\n31.2%<\/td>\n<\/tr>\n
2 engines<\/td>\n11,497<\/td>\n14.1%<\/td>\n17.4%<\/td>\n<\/tr>\n
3\u20135 engines<\/td>\n6,197<\/td>\n7.6%<\/td>\n22.1%<\/td>\n<\/tr>\n
6\u20137 engines<\/td>\n1,223<\/td>\n1.5%<\/td>\n18.6%<\/td>\n<\/tr>\n
All 8 engines<\/td>\n327<\/td>\n0.4%<\/td>\n10.7%<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

Read the last two rows together. The 1,550 URLs cited by six or more engines are 1.9% of the corpus but absorb 29.3% of citation volume.<\/strong> A single page in that tier was worth, on median, 43\u00d7 the weekly citation count of a single-engine page in the same category.<\/p>\n

That asymmetry is the practical argument for treating cross-engine agreement as its own target rather than a by-product of per-engine work. It also explains why per-engine tactics plateau: you can win ChatGPT repeatedly and never touch the shared core.<\/p>\n

The engine-pair overlap matrix at URL level<\/h2>\n

Pairwise overlap tracks retrieval architecture almost perfectly. Engines sharing an underlying index converge; engines with proprietary or real-time retrieval diverge. Below is URL-level Jaccard similarity for selected pairs, computed on the full twelve-week citation sets.<\/p>\n\n\n\n\n\n\n\n\n\n\n\n\n\n
Engine pair<\/th>\nURL overlap<\/th>\nDomain overlap<\/th>\nRatio<\/th>\n<\/tr>\n<\/thead>\n
ChatGPT \u2194 Copilot<\/td>\n21.3%<\/td>\n47.9%<\/td>\n2.2\u00d7<\/td>\n<\/tr>\n
Gemini \u2194 AI Overviews<\/td>\n19.8%<\/td>\n52.4%<\/td>\n2.6\u00d7<\/td>\n<\/tr>\n
AI Mode \u2194 AI Overviews<\/td>\n18.4%<\/td>\n55.1%<\/td>\n3.0\u00d7<\/td>\n<\/tr>\n
Gemini \u2194 AI Mode<\/td>\n16.9%<\/td>\n49.6%<\/td>\n2.9\u00d7<\/td>\n<\/tr>\n
Perplexity \u2194 AI Overviews<\/td>\n9.1%<\/td>\n34.8%<\/td>\n3.8\u00d7<\/td>\n<\/tr>\n
Claude \u2194 Perplexity<\/td>\n7.6%<\/td>\n31.2%<\/td>\n4.1\u00d7<\/td>\n<\/tr>\n
ChatGPT \u2194 Gemini<\/td>\n6.2%<\/td>\n38.0%<\/td>\n6.1\u00d7<\/td>\n<\/tr>\n
Grok \u2194 Claude<\/td>\n3.4%<\/td>\n22.7%<\/td>\n6.7\u00d7<\/td>\n<\/tr>\n
Grok \u2194 Gemini<\/td>\n2.8%<\/td>\n24.1%<\/td>\n8.6\u00d7<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

Three clusters fall out cleanly. The Bing-backed cluster<\/strong> (ChatGPT, Copilot) overlaps most tightly. The Google-grounded cluster<\/strong> (Gemini, AI Overviews, AI Mode) forms a second dense group. Perplexity, Claude and Grok<\/strong> each sit closer to independent, with Grok the outlier in every pairing \u2014 its citation set skews to X threads and very recent news, which no other engine weighted comparably.<\/p>\n

For the retrieval plumbing behind this, our breakdown of which search index powers each AI engine<\/a> maps each platform to its source backend. The clustering here is that map, measured from the output side. The same structure shows up in brand-level results too \u2014 how much ChatGPT, Perplexity and Gemini overlap on brand picks<\/a> is the recommendation-layer version of this matrix.<\/p>\n

Notice also that the ratio column climbs as overlap falls. The less two engines actually agree, the more domain-level reporting flatters them.<\/strong> For Grok\u2013Gemini, domain overlap is 8.6\u00d7 the real URL agreement \u2014 enough to make two engines look loosely aligned when they share almost no pages.<\/p>\n

Which engines should you optimise for together?<\/h3>\n

Cluster membership decides the bundling. One page cannot be tuned eight ways, but it can be tuned three:<\/p>\n