When an engine cites the certifying body's own register, does it get the answer right?
HighEucerin was named as an EcoBeautyScore member in 24 answer cells across all five engines and all three markets. Twenty of those answers cited ecobeautyscore.com. Schauma was named in 22 cells, 19 of them citing the same source.
The EcoBeautyScore Association publishes its members and, under a separate adjacent heading, the brands that have published scores, printing each brand's parent company in parentheses. The engines retrieved that page and reported the second list as the first.
Implication: Retrieval worked and the answer was still wrong. A citation evidences that a source was consulted, not that it was understood, so citation share is not a measure of how accurately AI represents a brand.
Are the engines inventing certifications?
HighMostly not. Of 46 false claims, 18 report a credential after it lapsed, 12 attach a parent's credential to a consumer brand, 7 get the degree or the scheme wrong, 2 name a company outside the category, and 7 are inventions.
An early pass classified four CDP claims as fabrication. Checking each against the company's own disclosure history showed three had genuinely held the credential or held a partial version. Beiersdorf, Kao and Symrise all really were CDP Triple A, years ago.
Implication: The engines are mostly telling you something that was true, and nothing in the answer tells you when it stopped being true. Staleness, conflation and overstatement have different causes and different fixes, and none is solved by more citations.
Whose credential is it?
HighEight false claims attach a parent group's credential to a consumer brand that does not hold it. Six are Science Based Targets claims, where Chantecaille, Eucerin, Guerlain, Yves Rocher, Aveda and the Estée Lauder brand were each credited with their parent's validated targets.
Eucerin is conflated twice, on two different credentials, by two different routes. On EcoBeautyScore the engines misread a page that separates members from brands; on SBTi they simply attached the parent's targets to the parent's brand.
Implication: There is no rule of thumb. Aveda holds B Corp certification in its own right while its parent does not, and does not hold SBTi targets while its parent does. Whether a brand can hold a credential independently is a fact about the credential, not about the corporate tree.
Are engines reporting current certification status, or an older one?
HighEvery engine answered the CDP Triple A question using the 2022 A List, reached through a trade article published in January 2023. CDP's 2025 A List had been public since 8 January 2026, 68 days before the prompts ran.
Beiersdorf, Kao and Firmenich were each named as Triple A holders. None appears among the 23 Triple A companies of 2025. Beiersdorf and Kao genuinely held the rating in earlier cycles; Firmenich has not existed under that name since its 2023 merger into dsm-firmenich.
Implication: A lapsed credential can outlive its own expiry in the answer layer. Only 3 of Beiersdorf's 28 mentions cited cdp.net at all, so here the register was not consulted rather than misread.
Does the same engine give the same answer twice?
HighChatGPT stated 'Coty Inc.: Upgraded to a Triple A rating in 2025' in one run, and in another run of the same prompt correctly described an A rating in CDP's 2025 Climate Change disclosure. Both runs cited coty.com.
Coty's own announcement of 12 December 2025 records Climate A, Water B and Forests B-. An A on one of three questionnaires is not Triple A. The source containing the correct answer was cited in the run that got it wrong.
Implication: A brand checking this once had roughly even odds of seeing a flattering error or the truth, with no way to tell which it had received.
Do the engines fail in the same way across different credentials?
HighEcoBeautyScore answers cited the register in 70% of assertions and failed by conflation and staleness. EWG answers cited it in 64% and confused two different EWG programmes. CDP answers cited it in 13% and failed by staleness.
Two structurally different roads to a wrong answer: the primary source retrieved and misread, or the primary source never retrieved and trade press used instead.
Implication: Neither pattern is visible to a brand monitoring citation share, because in both cases the engines cited something plausible.
Were the engines wrong about everything?
HighNo. The two largest claims under a credential we did not fully verify were spot-checked and both were correct. Aveda was named as a certified B Corp in 18 cells and has held the certification since January 2023. Clarins was named in 16 cells and Groupe Clarins was certified in 2025, with the certificate naming the Clarins and myBlend brands.
Aveda is a wholly owned subsidiary of The Estee Lauder Companies, which is not itself a B Corp. That is structurally identical to Eucerin being credited with Beiersdorf's membership. One is true and one is false, and nothing in either answer distinguishes them.
Implication: The finding is not that AI gets certifications wrong. It is that you cannot tell from the outside which answers are the right ones. Only the register can settle it.
Is a Skin Deep rating the same as EWG Verified?
HighNo, and the engines conflated them. Era Organics was named as EWG Verified three times, all three citing ewg.org. It has 27 products rated in EWG's Skin Deep database and none carrying the EWG Verified mark.
Skin Deep is a ratings database any product can enter. EWG Verified is a paid certification requiring application, review and ongoing standards. Clean-beauty publishing uses the two interchangeably and the engines reproduced that.
Implication: A brand marketing itself as 'EWG rated' and a brand holding EWG Verified are making different claims, and the answer layer does not distinguish them.
What was the one invention?
HighA brand. Two engines named 'Natural Organic Skincare' as an EWG Verified brand in four cells, all four citing ewg.org. No brand of that name exists in the EWG Verified directory or Skin Deep; the phrase is a product category.
The only claim in the study with nothing real behind it invented the subject rather than the credential, assembling a plausible brand name out of the category language in the question.
Implication: No brand is harmed, since none exists, but a reader acting on it would be searching for a product that cannot be bought.
Can one credential be substituted for another?
HighSeven times, across four different schemes. Urban Decay and Wet n Wild hold PETA certification rather than Leaping Bunny. Saie signed The Climate Pledge rather than validating targets with SBTi. Era Organics is rated in EWG's Skin Deep database rather than EWG Verified. Coty, Shiseido and Unilever hold a single or double CDP A rather than Triple A.
Each substitutes something real, adjacent and weaker for the specific credential asked about. The engines treat 'cruelty-free', 'climate committed' and 'EWG' as single categories when each contains separate schemes with different requirements.
Implication: That distinction is the entire reason a brand pays for the stronger scheme.
How out of date can an answer be?
HighDr. Hauschka was named as a certified B Corp in eight cells across three engines and all three markets. Its US entity was certified in September 2015 and lapsed around 2018; its German parent has never been certified.
Beautycounter is the more instructive case. Counter Brands LLC filed for Chapter 7 in April 2024, and the relaunched brand still trades, still appears on Leaping Bunny's recommitment list, and has years of sustainability press behind it. It has not recertified.
Implication: A lapse leaves almost no public trace, while the original certification leaves a great deal. The answer layer weights what is abundant over what is current.