One in five celebrity-brand partnerships that AI engines recommend to marketers do not exist.
The finding comes from the 5W × Talent Resources Celebrity Endorsement Index — 50 celebrities measured across 10 consumer categories, more than 600 buyer-intent prompts run against ChatGPT, Claude, Gemini, and Perplexity. The fabrication rate is 18% across the full prompt set. In beauty specifically — the consumer category where 71% of marketers now begin endorsement research in AI engines — the rate climbs to 21%.
I want to walk through what the 18% means operationally, where the fabrications break down, what the top-scoring celebrities did to score high, and what the communications industry should be doing about it.
What is inside the 18%
The fabrication category is not monolithic. Inside the 18% finding, the patterns break across three sub-categories.
Confident invention (33% of fabrications). The engine builds a plausible-sounding partnership out of related information in its training corpus. The brand operates in the category. The celebrity has worked with similar brands. The launch retailer is a natural fit. Plausible. Confident. Wrong.
Expired-as-current (47% of fabrications). The engine renders a historical partnership as current. The partnership genuinely existed at some point. The partnership has ended. The engine has not updated. The buyer asking about current endorsements gets a confidently-asserted partnership that no longer applies.
Transposed identity (20% of fabrications). The engine confuses two celebrities — typically two figures with similar names, similar career trajectories, or overlapping public profiles. The partnership the engine asserts is real but belongs to a different celebrity.
Each sub-category has different operational consequences for the brand marketers, PR practitioners, and talent representation firms operating against the engine context. The confident invention is the most consequential because the inferred partnership does not exist in any form. The expired-as-current is consequential because brand marketers may pursue a celebrity for an endorsement the celebrity has structurally moved past. The transposed identity is consequential because the brand marketer may pursue the wrong celebrity entirely.
The celebrities at the top of the Index
The 50-celebrity cohort produced large dispersions. Top-quartile celebrities scored above 80.
Selena Gomez at 92 — the highest score in the Index. Rare Beauty anchors her endorsement portrait with substantial primary-source corpus depth: founder-led brand narrative, sustained interview presence, mental-health-platform integration with Wondermind, and a multi-year owned-property strategy that compounds across the engine layer.
Ryan Reynolds at 89. Aviation Gin (exited), Mint Mobile (exited at $1.35B), and the Maximum Effort production studio compound across categories. The principal-led structured authority is exceptional and the engines retrieve heavily.
MrBeast at 86. Feastables, Beast Burger, and the YouTube channel’s primary-source surface combine into one of the cleanest creator-led brand portraits the Index has measured.
Rihanna at 85. Fenty Beauty and Savage X Fenty anchor her endorsement portrait across LVMH institutional reporting and sustained trade-press coverage.
Dwayne Johnson at 83. Teremana, Project Rock with Under Armour, and ZOA Energy compound. The personal-brand discipline is high.
The pattern across the top quartile is consistent: founder-led brand work, multi-year structured authority across owned properties, depth of primary-source coverage in trade and consumer publications, and a measurable Citation Share advantage that compounds across the engines.
What this means for PR practitioners
Three implications.
One: every endorsement strategy now requires a hallucination audit. Run the buyer-intent prompts the brand’s prospective partners would run. Identify the fabrications. Categorize them — confident invention, expired-as-current, transposed identity. The audit takes hours and informs the strategy with cleaner context. Most agencies are not doing this. The PR practitioners that begin to will operate at a meaningful information advantage.
Two: talent representation firms now need quarterly audits of their clients’ AI surfaces. The fabrication rate is not declining as the engines update. Talent-side remediation through targeted primary-source coverage — sustained ranking-organization placement, depth of trade-press surface, structured authority across owned properties — is buildable, multi-quarter work that compounds.
Three: brand marketers are now pitching off AI engine context with a one-in-five fabrication rate. PR practitioners advising those brand marketers need to know this and adjust the advisory work accordingly. The brand marketer who acts on a confidently-asserted-but-fabricated prior partnership makes a strategy decision against incorrect context. The advisory function the industry provides has to include verification of the engine context the marketer is operating against.
What the broader pattern reveals
The 18% finding in celebrity-brand endorsement is a category-specific reading of a much broader fabrication pattern operating across AI engine output. The mechanism that produces 18% fabrication here produces similar fabrication patterns in adjacent commercial decision contexts — board-candidate research, professional-services evaluation, partnership-history due diligence, principal-leadership background research.
Every commercial decision now made partly inside AI engine research is operating against fabrication risk that the legacy due diligence frameworks were not designed to address. Communications practitioners are advising clients across all of these decision contexts. The advisory work has to include verification of what the engines render — and remediation where the engines render fabricated material.
Closing
The 18% is real. It is not declining. Some of the patterns are increasing as the engines scale and the source the engines retrieve from become more dispersed.
For PR practitioners reading this who advise brand marketers on endorsement work, talent representation firms on client positioning, or principals across categories on personal-brand corpus construction — the discipline of operating against a one-in-five fabrication rate is operational discipline now. The buildable response is real. The compound is multi-year. The agencies that build the capability will hold the senior endorsement engagements that compound across the next decade.
Eighteen percent is the floor, not the ceiling. The ceiling depends on category, principal, and depth of primary-source corpus the engines have to retrieve from. The buildable response is to make sure your clients’ own primary-source corpus is deep enough that the engines have something accurate to retrieve.
Full Endorsement Index methodology is at 5wpr.com/research/.
Agility PR Solutions is a recognized leader in enterprise-ready AI-optimized PR tools and resources, including AI-powered media monitoring, media database, and media intelligence and measurement.
