Study: Google AI Overviews lean on few sources, often YouTube
Google's AI Overviews favour their own platforms and a tight cluster of sources, leaving most brands and institutions effectively uncited.
Key takeaways
- Google AI Overviews draw from a narrow source pool, with YouTube the most-cited property in a 4,480-query study.
- The same query does not reliably trigger an AI Overview, making brand citation intermittent across millions of searches.
- YouTube's dominance signals a retrieval preference for Google-owned platforms, disadvantaging text-heavy publishers.
- Multilaterals and policy institutions publishing primary research may go uncited despite source quality.
- Brands without structured video content on YouTube face a growing and largely unmeasured citation gap.
AlgorithmWatch ran 4,480 election-related queries through Google Search and found something that should unsettle any brand hoping to appear in AI Overviews: the system draws from a strikingly narrow pool of sources, and the dominant one is YouTube. The Decoder reports that the German advocacy group, using access granted under the EU's Digital Services Act, discovered Google's AI Overviews appeared inconsistently across identical queries and repeatedly favoured its own platform over independent publishers.
The mechanism matters. AI Overviews do not sample the web; they sample a hierarchy. When Google's system decides which sources to synthesise, it appears to weight properties it already controls or trusts deeply. YouTube, a Google subsidiary, topped the citation list. That single finding reframes the conventional advice to "create good content and earn links." In an AI Overview world, the question is not whether your content ranks but whether the model treats your domain as a citation-worthy node at all. For most organisations, the answer is quietly no.
Inconsistency as a structural problem
The inconsistency finding is at least as significant as the source concentration. AlgorithmWatch found that the same query, run multiple times, did not reliably trigger an AI Overview, and when it did, the sourcing varied. That is not a bug being fixed; it is how probabilistic systems behave under conditions of thin training signal or contested topics. Election queries sit precisely in that territory: high sensitivity, Google's stated caution about political content, and yet overviews appeared anyway, sometimes taking discernible positions.
For a brand or institution, inconsistency is a citation problem dressed as a reliability problem. If a query about your organisation's work in climate finance or labour standards triggers an AI Overview on some occasions and not others, your presence in that answer space is intermittent at best. Intermittent presence compounds over millions of queries into effective invisibility.
The implication for multilateral institutions and policy bodies is specific. Organisations such as UNDRR or CGAP publish credible, primary-source material on exactly the kinds of contested policy questions AI Overviews are now attempting to answer. If the model's source pool is narrow and self-reinforcing, that material goes uncited regardless of its quality. The DSA access that enabled AlgorithmWatch's study exists only within the EU; outside that jurisdiction, organisations have no comparable right to audit what Google's system is actually reading. That asymmetry will persist.
What wins citations in a narrow pool
YouTube's dominance points to a structural advantage: video content hosted on a Google property, with Google's own engagement signals feeding the retrieval layer. That is not a coincidence. It is a design preference made visible. For brands in financial services or industrial groups that have invested in written thought leadership but not in structured video or multimodal content, the citation gap is real and growing.