AI search is changing the economics of visibility on the web.
The strongest recent research points to a two-part shift. First, answer-first search interfaces reduce visits to source pages. Second, the citation layer inside those interfaces allocates visible credit through a narrower set of sources than the open web would suggest. Those two shifts matter together. One cuts downstream traffic. The other concentrates authority.
This is the machine relations problem in its cleanest form. A brand no longer wins just by publishing a page that can rank. It wins by becoming a source the machine selects when it compresses a broad retrieval field into a short answer and a few citations.
The traffic side of the shift is now causal, not anecdotal #
A 2026 University of Washington study measured the effect of Google AI Overviews on Wikipedia traffic using a difference-in-differences design across 161,382 matched article-language pairs and 46.5 million observations (Khosravi and Yoganarasimhan, 2026). The authors found that exposure to AI Overviews reduced daily traffic to English Wikipedia articles by about 15% (Khosravi and Yoganarasimhan, 2026).
That finding matters because Wikipedia is not weak inventory. It is a trusted, heavily cited informational source. If a source with that level of authority loses traffic when AI summaries appear, then citation is no longer the same thing as demand capture.
The paper also found heterogeneous effects. Relative declines were larger for Culture content and smaller for STEM content (Khosravi and Yoganarasimhan, 2026). The result is directionally consistent with substitution for informational queries.
AI search exposure is expanding fast enough to change the baseline #
A 2026 MIT-led study executed 24,000 search queries in 243 countries and generated 2.8 million AI and traditional search results across 2024 and 2025 (Aral, Li, and Zuo, 2026). The authors found that 67% of tested queries in the United States returned AI answers in 2025, compared with 42% in 2024 (Aral, Li, and Zuo, 2026).
The same paper reports that AI search surfaces significantly fewer long-tail information sources and lower response variety than traditional search (Aral, Li, and Zuo, 2026). As AI exposure expands, this narrower answer layer becomes a larger part of how users encounter information (Aral, Li, and Zuo, 2026). That is the second part of the shift. The answer layer is not just larger. It is structurally narrower.
Traditional search exposes the user to a ranked field of pages. AI search compresses that field into a synthetic answer and a limited citation set. As exposure expands, being omitted from that visible citation layer becomes more consequential than being buried on page two used to be.
Citation concentration persists even when engines disagree #
A 2025 study using AI Search Arena data examined more than 24,000 conversations, 65,000 responses, and more than 366,000 citations across systems from OpenAI, Perplexity, and Google (Yang et al., 2025). The authors found that providers cite different news sources, but they still share a concentration pattern in which a small number of outlets capture a disproportionate share of citations (Yang et al., 2025).
That detail matters. Citation concentration is not just one vendor’s quirk. The selection logic differs across providers, but the visible outcome is still compressed (Yang et al., 2025). The open web may contain a broad source set. The answer layer still privileges a thinner credited set.
Attribution gaps make the concentration problem even tighter #
A 2025 paper using approximately 14,000 real-world LMArena conversation logs documented major attribution gaps in web-enabled LLM systems (Strauss et al., 2025). The authors found that 34% of Google Gemini responses and 24% of OpenAI GPT-4o responses were generated without explicitly fetching online content (Strauss et al., 2025). They also found that Gemini provided no clickable citation source in 92% of answers (Strauss et al., 2025).
The same paper reports that Perplexity Sonar visited roughly 10 relevant pages per query but cited only three to four, and that citation efficiency ranged from 0.19 to 0.45 additional citations per extra relevant page visited depending on system design (Strauss et al., 2025).
This is the quiet part of the market shift. The retrieval layer may be wider than the citation layer the user sees (Strauss et al., 2025). In commercial terms, that means the machine can learn from more sources than it publicly rewards.
Data points that define the shift #
The minimum set of numbers worth carrying forward is simple. AI Overviews reduced Wikipedia traffic by about 15% in the causal study sample (Khosravi and Yoganarasimhan, 2026). AI answer exposure in the United States rose from 42% of tested queries in 2024 to 67% in 2025 in the MIT study (Aral, Li, and Zuo, 2026). Cross-provider citation analysis covered more than 366,000 citations across more than 24,000 conversations (Yang et al., 2025). Attribution-gap analysis found 34% no-search behavior for Gemini, 24% for GPT-4o, and 92% no-clickable-citation behavior for Gemini (Strauss et al., 2025). Those are not edge-case numbers. They define the current shape of the answer layer.
The mechanism is compression #
Put those papers together and a cleaner model appears.
- More queries are being answered inside AI interfaces.
- Those interfaces reduce downstream visits for informational sources.
- The answer layer exposes fewer long-tail sources than traditional search.
- The citation layer is more concentrated than the retrieval layer beneath it.
- Systems often consume more sources than they visibly credit.
The right word for that sequence is compression. Search is compressing a wide source field into a narrow credited answer surface.
That is why conventional SEO logic is no longer enough on its own. Ranking still matters. But ranking is no longer the final battlefield. Selection into synthesis is.
What this means for Machine Relations #
The strategic implication is blunt. Brands cannot treat AI visibility as a simple extension of blue-link SEO.
If clicks are falling while citations are concentrating, then the scarce asset is not just page-one real estate. It is trusted source status inside the machine’s answer construction process. That is a different game. It rewards proprietary evidence, clean entity signals, answer-first formatting, and earned authority that gives a model a reason to trust and cite the source.
This is why the category needs Machine Relations. The question is no longer just how to rank. It is why the machine would choose your source when it only has room to show a few. That same logic also explains why founder authority and editorial authority increasingly work together. See the operating layer at JaxonParrott.com, the execution layer at ChristianLehman.com, and the diagnostic layer in this AI visibility audit framework.
FAQ #
Is this just a Google problem? #
No. The traffic evidence in this piece centers on Google AI Overviews, but the concentration and attribution evidence comes from cross-provider studies covering Google, OpenAI, and Perplexity (Yang et al., 2025; Strauss et al., 2025).
Does citation still matter if clicks fall? #
Yes. Citation still matters because visible credit shapes authority, trust, and downstream discovery. The point is that citation should no longer be confused with guaranteed traffic capture (Khosravi and Yoganarasimhan, 2026).
What should brands optimize for now? #
Brands should optimize for being selected into answer construction. That means becoming a source with evidence, clarity, and trust signals that an answer engine can use and defend. More of that work lives at Machine Relations.