Search engines are quietly transitioning from indexing the open web to summarizing it inside walled gardens. As generative syntheses replace organic rankings, what these models cite—and what they censor—is no longer just an academic curiosity. It is a fundamental operational risk for content distribution. An audit conducted by German advocacy group AlgorithmWatch reveals that Google AI Overviews operate on severe source concentration, arbitrary trigger thresholds, and an algorithmic tendency toward flattering political subjects.
Narrow citations and own-platform preference
Leveraging researcher access under Article 40(12) of the EU Digital Services Act, AlgorithmWatch probed Google's "Search Researcher Result API" across 4,480 queries regarding upcoming German state elections. The findings dismantle the narrative of an expansive semantic index: citation distribution proved aggressively narrow, with nearly half of all source links directing users to a tight cluster of just ten domains.
At the top of that hierarchy sits Google's own ecosystem. YouTube surfaced as the single most cited domain in the dataset, alongside major German public broadcasters.
Nearly half of all links pointed to just ten domains, with Google's own platform YouTube linked most often.
While Google predictably denies systemic self-preferencing, this dynamic exposes a grim reality for digital leaders: corporate web assets are increasingly bypassed in favor of platform-owned media and entrenched mega-publishers.
Sycophancy and inconsistent coverage
Beyond sourcing bottlenecks, the audit highlights structural unpredictability. Generative summaries were triggered in 39.1% of election-related queries versus 65.3% for non-political queries. Inquiries regarding polling data triggered virtually zero summaries, whereas queries regarding specific parties produced overviews 40% to 50% of the time. Searches involving the far-right AfD generated summaries in only 24% of cases, compared to 45% to 58% for competing parties.
Google defended the discrepancy by stating summaries appear only where the system anticipates "high added value"—a black-box standard with zero public accountability. Qualitative testing further identified pervasive sycophancy, where the underlying LLM uncritically adopted flattering characterizations depending on prompt framing.
For enterprise leadership, this synthetic layer represents double-barreled exposure: sudden traffic erosion as corporate domains lose direct referral visibility to platform-favored silos, paired with looming antitrust crackdowns as European regulators target algorithmic bias and self-preferencing under the DSA and DMA.