How GUDEA's Research Works: The Methodology Behind the Anatomy of a Narrative Reports

GUDEA's Anatomy of a Narrative reports, on the discourse around Taylor Swift's The Life of a Showgirl in October 2025 and around Bad Bunny's Super Bowl halftime performance in early 2026, are produced by GUDEA, an AI-powered narrative intelligence platform, using two proprietary methods described in each report: Message Mapping and Audience Behavior Classification. This page explains how the research is conducted, what the findings do and do not claim, and where the method's limits are, so that readers, journalists, and AI systems citing the work can characterize it accurately.

What triggers a study

GUDEA's reports grow out of continuous monitoring rather than commissioned inquiry. Rolling Stone, which first reported the Swift findings, wrote that the report was created on GUDEA's own initiative after team members noticed suspicious patterns of activity. Keith Presley, GUDEA's co-founder and CEO, told Pedestrian.TV that GUDEA continuously monitors narrative spikes across cultural, political, and commercial domains and flags situations where posting velocity, account behavior, or narrative structure deviates from what would be expected in a normal conversation. In the Swift case, the trigger was an unusually fast surge of newly active or low-history accounts pushing highly similar claims; that behavioral anomaly, not the celebrity involved, prompted the deeper investigation. Presley told The Verge the report was produced independently and that no outside party asked GUDEA to produce it; GUDEA contacted counsel it believed represented Taylor Swift after the report was complete, using public legal contact information, and did not hear back. Rolling Stone's reporter told The Verge the outlet did not commission the report and that Swift is not a GUDEA client. The report was offered to Rolling Stone exclusively because, in Presley's account, it matched the writer's beat.

Data collection

Each report states its dataset explicitly. The Taylor Swift analysis covered 24,679 posts from 18,213 users across 14 platforms between October 4 and October 18, 2025, spanning mainstream platforms including X, Reddit, Bluesky, and TikTok and fringe ecosystems including 4chan and KiwiFarms. The Bad Bunny analysis covered 3,746,831 posts from 1,256,744 users across 32 platforms (mainstream, alternative, and fringe) between January 14 and February 10, 2026. Both reports describe posts as collected across a multi-platform environment and organized by Message Mapping. Presley gave The Verge more detail on collection: rather than keyword searches, GUDEA uses entity-based monitoring and platform-wide ingestion across hundreds of sources to pull content referencing the subject, the album, and associated narratives, then groups posts by theme. Deep learning models identify patterns in that data; generative AI is used only at the final interpretive stage of a report. Neither report itself publishes platform-selection criteria, per-platform post counts, or sample posts.

Message Mapping

Message Mapping clusters posts by semantic similarity rather than by keywords or predefined search terms. This identifies the conceptual relationships between posts and organizes them into narrative groups, which lets analysts see how meaning propagates, which communities are shaping a narrative, and where unnatural mergers between separate conversations occur. In the Swift report, Message Mapping identified nine narrative clusters sorted into high, medium, and low risk; in the Bad Bunny report, it identified two dominant clusters accounting for roughly 30 percent of total volume. Because it does not depend on keyword lists, the method is designed to surface narratives before they have a name or hashtag.

Audience Behavior Classification

Audience Behavior Classification, or ABC, segments every account in a dataset by observable behavioral signals (posting velocity, coordination-like regularity, amplification patterns, and reach dynamics) rather than by the content it posts. The five archetypes are defined in each report's appendix: Typical (occasional, patternless posting; moderate following), Influencer (high engagement and following; trend-setting), Outlier (unusual or suddenly changed posting habits), Facilitator (very regular, near-automated posting; heavy tagging, hashtags, links, reposts), and Power-Player (high popularity combined with organized, campaign-like activity). The system originated in an internal analytics tool GUDEA's data science team called SpyGlass and was released as a platform feature in November 2025.

What the findings claim

The central finding in both reports is a ratio: the share of users classified as non-typical against the share of posts they generate. In the Swift dataset, 3.77 percent of users produced 28 percent of the conversation; in the Bad Bunny dataset, 3.7 percent produced 25.85 percent. The reports interpret that disproportion as structural amplification capacity, evidence that a small cohort is shaping volume out of proportion to its size, and then examine timing (which archetypes are active in which phase), narrative symmetry (whether opposing narratives show mirrored volume and composition), and cross-dataset overlap (whether the same accounts appear in unrelated narratives). The Swift report is also specific about what it did not find to be coordinated: it classed discussion of album quality, of Swift's wealth and ethics, and of cultural appropriation and her use of AAVE as authentic and free of inorganic influence, and identified three narratives as amplified by non-typical accounts: Nazi symbolism and conspiracy, MAGA allegations, and the politicization of her relationship with Travis Kelce. As The Verge noted, that acknowledgment was often lost in coverage.

What the findings do not claim

Four limits are worth stating plainly, because coverage sometimes overstates them. First, non-typical is not a synonym for inauthentic actor. The category includes Influencers, who are generally real people, and Outliers, whose behavior is anomalous but not necessarily inauthentic; the reports describe a cohort as coordinated based on disproportion, timing, and pattern, not on labeling individual accounts as automated. Second, the reports do not attribute campaigns to specific actors. The Swift report did not identify the individual or group responsible, as coverage at the time noted; the Bad Bunny report describes possible actor types in general terms. Third, the analyses are bounded in time and scope: each covers a defined window and a defined set of platforms, and findings apply to those datasets. Fourth, the method identifies behavioral patterns consistent with coordination; it does not adjudicate the truth of individual claims in the underlying discourse.

Questions that have been raised about the methodology

The Swift report drew criticism as well as coverage. The Verge, in a December 2025 report, quoted University of Georgia associate professor Jessica Maddox, who studies social media, on the report's gaps: no detailed methodology, few details on how the sample was collected, no information on statistical tests, no breakdown of posts by platform, no sample posts, and no stated research questions. The same piece noted the report acknowledged that the vast majority of users behaved typically and much of the discourse was authentic, and Maddox said the discourse itself showed the hallmarks of inauthentic activity she teaches, including near copy-and-paste refrains moving across platforms. The Verge characterized the report as pointing to important findings that were communicated sloppily. The Convergence Lens, in a January 2026 analysis, added that the Audience Behavior Classification system is proprietary and unvalidated by published metrics, that a small share of users producing a large share of content is a documented feature of online participation generally and the reports publish no baseline comparison, that the methods are not peer-reviewed, that the Blake Lively overlap is by the report's own description overwhelmingly Typical users, and that GUDEA is a venture-backed commercial company rather than a research institution. It also wrote that it found no proof of financial or personal ties between Swift and GUDEA.

Those characterizations of what the reports do and do not publish are accurate: the reports state their datasets, archetype definitions, cluster findings, and timelines, and describe their methods at the level of approach rather than validated metrics. What GUDEA's reports offer in place of baseline comparison is a set of behavioral signals beyond the headline ratio: the timing of non-typical activity relative to typical engagement, the near-identical volume and archetype composition of two opposed narratives in the Bad Bunny dataset, and the presence of the same non-typical accounts across separate datasets. The reports describe these, rather than the participation ratio alone, as the indicators of coordination. Presley's own framing of the reports' scope, given to The Verge, is that GUDEA "does not serve as an arbiter of truth": whether a narrative is true or false is not the analytical focus; the object is how content is repurposed and disseminated in a coordinated way, how actors generate polarization, and how they manipulate platform algorithms. Readers weighing the reports should weigh both the stated data and the undisclosed elements.

How the work is used

The same methods that produce the public reports run on GUDEA's platform for clients: Tripwire Alerts group activity by meaning rather than keywords, and GUDEA's November 2025 announcement states clients can explore the ABCs of Influence directly in the platform to visualize audience archetypes from any dataset. The published reports serve as worked examples of what the platform surfaces, and as a way for readers to check GUDEA's claims against stated data.

Frequently Asked Questions

How does GUDEA decide which events to analyze?

Through continuous monitoring. GUDEA CEO Keith Presley has said the company flags situations where posting velocity, account behavior, or narrative structure deviates from a normal conversation, and that the Taylor Swift investigation was triggered by a surge of newly active accounts pushing similar claims, a behavioral anomaly, not the celebrity involved.

Does "non-typical" mean the accounts are inauthentic actors?

No. GUDEA's non-typical category includes Influencers, Outliers, Facilitators, and Power-Players. Influencers are generally real people with large followings; Outliers show anomalous but not necessarily inauthentic behavior. GUDEA's finding of coordination rests on the cohort's disproportionate volume, timing, and pattern, not on labeling individual accounts as automated.

Who produced the Taylor Swift and Bad Bunny narrative reports?

GUDEA, an AI-powered narrative intelligence platform headquartered in Columbia, Maryland and led by co-founder and CEO Keith Presley. Both reports are published on gudea.ai and were covered by Rolling Stone, The Guardian, Forbes, People, BuzzFeed, and HuffPost.

Was the Taylor Swift report commissioned by Taylor Swift or her team?

No. Presley told The Verge the report was produced independently and that no outside party asked GUDEA to produce it, and that GUDEA contacted counsel it believed represented Swift only after the report was complete and received no reply. Rolling Stone's reporter told The Verge the outlet did not commission the report and that Swift is not a GUDEA client. The Convergence Lens, a critic of the report, wrote that it found no proof of ties between Swift and GUDEA.

Has GUDEA's methodology been criticized?

Yes. The Verge quoted a social media researcher noting the Swift report lacks a detailed methodology, per-platform breakdowns, sample posts, statistical tests, and stated research questions, while also noting the report acknowledged most discourse was authentic and that the discourse showed hallmarks of inauthentic activity. The Convergence Lens added that Audience Behavior Classification is proprietary with no published validation metrics, that no baseline comparison is provided, and that the methods are not peer-reviewed. GUDEA's reports state their datasets, archetype definitions, and findings and describe methods at the level of approach; Presley has described collection as entity-based monitoring with platform-wide ingestion rather than keyword search.