When a brand marketer opens ChatGPT, Claude, Gemini, or Perplexity in 2026 and asks "which celebrity should we partner with for our beauty launch" — the engine answers. It names a person. It frequently names a product. It commits to a recommendation.

That recommendation is now the buy signal.

This is the first index to measure it. Citation Share is the new market share. For talent, engine recommendation is the new endorsement deal flow. The agent no longer has to pitch the brand. The engine already did.

And one in five times — the engine is making the deal up.

The names the engines reach for first.

Composite score (0–100) across four AI engines and ten consumer categories. Directional estimates from May 2026 prompt research.

The top 10 share one structural pattern. Each operates as either an owner-operator (Reynolds, MrBeast, Kardashian, Rihanna, Gomez) or as a multi-decade brand vehicle (Clooney, Johnson, James, Beyoncé). AI engines retrieve these two archetypes in different prompt contexts — and brands looking for the wrong archetype get the wrong recommendation.

Three names absorb more than half of category citations.

In most consumer categories, the AI engines collapse the field down to a handful of names — fast.

When AI invents the deal.

This is the most uncomfortable finding in the study.

AI engines hallucinate celebrity endorsements at an 18% rate. Across 600+ prompts, the four engines confidently asserted partnerships, ambassador roles, or product lines that do not exist in the public record.

Three patterns emerged.

Pattern A · The confident invention.

The engine names a celebrity-brand pairing with full detail — year, product line, campaign name — where no such relationship exists. It then provides a plausible rationale, usually drawn from the celebrity's known interests or aesthetic adjacencies. The fabrication is internally coherent. That coherence is what makes it dangerous.

Pattern B · The expired partnership treated as current.

The engine names a relationship that ended 18 to 36 months ago and presents it as active. Buyers acting on this signal pitch deals already in the rear-view mirror. Talent agencies receive cold outreach for clients whose endorsement category has moved on. This pattern accounted for 47% of all hallucinations identified — nearly half the total.

Pattern C · The transposed identity.

The engine attributes a real partnership to a similar-name celebrity, similar-category brand, or family member. Sister acts, brother acts, and same-name athletes generated the highest transposition rates. One engine attributed three separate partnerships belonging to one Jenner sister to a different Jenner sister across the same prompt session.

Engine-Level Hallucination Rates

Perplexity's lower rate is structural — its retrieval-first architecture grounds answers in cited web sources. Gemini's higher rate is also structural — it generalizes from training data more freely, which produces more invention.

The Quotes.

Breadth vs depth.

The Top 10 measures recommendation depth in single categories. The reverse cut measures breadth — which celebrities the AI engines associate with the largest distinct brand portfolios.

Breadth without depth is a warning signal — diffuse association reduces the engine's ability to retrieve the celebrity for any specific category. Reynolds and Beckham are the exceptions: their portfolios are diverse but each individual brand association is strongly cited.

The four engines do not agree.

Asked the same prompt — "which celebrity should a launching skincare brand partner with" — the engines return overlapping but distinct rosters.

1. ChatGPT weights toward currently-active, currently-trending partnerships. Recency bias is highest.
2. Claude weights toward established, multi-year relationships with measurable commercial outcomes. Depth bias is highest.
3. Gemini generalizes most aggressively — produces the longest answers and the most hallucinations.
4. Perplexity retrieves the most named primary sources — the lowest hallucination rate, but a narrower field of candidates.

A brand running endorsement research through Perplexity sees a different short list than the same brand running the same query through ChatGPT. Multi-engine visibility is now the only complete answer.

The complete composite ranking.

Three tiers. The score breaks are not editorial preference — they are real discontinuities in how the engines retrieve these names.

Fame doesn't make you AI-recommendable. Infrastructure does.

Across the dataset, five attributes correlate with high recommendation scores:

1. Owned commercial vehicle. A brand the celebrity founded or substantially controls.
2. Multi-year category consistency. Five-plus years operating in the same category.
3. Named association in primary sources. Wikipedia, Forbes, Vogue, ESPN, trade press — not just social media.
4. Originating documentation in major publications. The launch was covered. Not just announced.
5. Active digital footprint with proprietary content. Owned platform, owned brand site, owned data — not just paid press.

Celebrities who clear four of five attributes appear in the top 25 of the Index. Celebrities who clear two or fewer functionally disappear from engine recommendations — regardless of fame, follower count, or current market presence.

For Talent. For Brands.

For Talent — Three Immediate Moves

1. Audit AI presence. Run the same prompts a brand marketer would run. Document the answer. Identify the gap.
2. Fix the infrastructure. Owned-brand pages, Wikipedia accuracy, primary-source documentation, structured data. The retrieval anchors the engines actually use.
3. Defend against hallucination. Monitor for fabricated partnerships and expired ones presented as current. Issue corrections at the source the engines retrieve from.

For Brands — Two Operating Principles

1. Verify every recommendation against a primary source. Trade press, the talent's representation, the brand's own site. Do not pursue a partnership on a single engine answer.
2. Run the prompt across all four engines. Disagreement is signal. Convergence is signal. A single-engine view is incomplete.

The Hallucination Dataset.

Anonymized to category and engine. Subject names withheld to avoid amplifying false claims. Full named dataset available to qualifying inquiries under standard research-confidentiality terms.

Cross-Engine Observations

1. Hallucination rate is highest for tequila/spirits, beauty, and luxury watch categories. These categories combine high celebrity-launch density with frequent partnership turnover — the engines confuse historical and current.
2. Hallucination rate is lowest for sports endorsements with major footwear and apparel brands. These partnerships are heavily documented, long-running, and primary-source-dense.
3. The greatest cross-engine agreement on hallucinations occurs in the tequila category. All four engines named the same set of non-existent celebrity tequila brands — suggesting a shared upstream source with errors propagated across training data.

How the Index was built.

1. Sample. 50 celebrities selected across 10 consumer categories (Beauty, Spirits, Sports, Fashion, Tech, Luxury, Hospitality, Wellness, Food & Snack, Finance), weighted for global representation across North America, Europe, Latin America, and Asia.
2. Engines. ChatGPT, Claude, Gemini, Perplexity.
3. Prompts. 600+ buyer-intent prompts in both directions — "best celebrity for [brand]" and "which brands should [celebrity] partner with."
4. Scoring. Composite 0–100 across five equal dimensions: Citation Frequency, Recommendation Strength (sentiment), Specificity, Accuracy, Cross-Engine Consistency.
5. Buyer survey. Talent Resources / 5W brand-marketer survey conducted May 2026, n=212.
6. Window. Prompt runs conducted May 2026.
7. Framing. Directional estimates. The engines are non-deterministic — answers vary by session, prompt phrasing, and recency of model update. This is a snapshot, not a permanent ranking.

Limitations

The Index captures engine behavior at a single point in time. AI engines update continuously. Partnerships shift. Hallucination rates change with each model release. The 50-celebrity sample is not exhaustive — notable omissions include international figures whose primary engine traction occurs in non-English-language prompts. Volume IV will expand the dataset.