Research & Evidence
Numbers published. Methods open. Claims sourced.
“We publish what the models said - verbatim and dated - never our characterization.”
Study 01 · captured 2026-07-09
Ten brands every buyer knows, six engines answering for them, and every load-bearing claim checked against primary sources the same day. This is what the engines told customers on 2026-07-09.
Captured 2026-07-09 · every claim web-verified against primary sources as of the same day
| Engine | Claims checked | False or stale | Error rate |
|---|---|---|---|
| Claude (no web) | 19 | 17 | 89% |
| Grok | 20 | 15 | 75% |
| Perplexity | 20 | 14 | 70% |
| ChatGPT (search) | 32 | 4 | 13% |
| Gemini 2.5 Pro | 17 | 2 | 12% |
| Claude (web) | 18 | 2 | 11% |
“OpenAI ousted Sam Altman as CEO in October 2025, and Greg Brockman is now the CEO”
Perplexity on OpenAI · 2026-07-09
Altman is still CEO - he gave a live interview the same day this was captured. Oct 2025 was a PBC restructuring; Brockman is co-founder and President.
“two free checked bags, open seating (with priority boarding options)”
Grok on Southwest Airlines · 2026-07-09
Both cited as current defining traits, but both are gone: checked bags have cost money since May 28, 2025, and open seating ended January 27, 2026 in favor of assigned seats.
“CEO: John Donahoe (since 2020)”
Claude (no web) on Nike · 2026-07-09
John Donahoe was Nike CEO 2020 to 2024. Elliott Hill has run Nike since October 2024.
Three tags, kept distinct: WRONG is contradicted by the record, OUTDATED was true once and is not anymore, FABRICATED was invented outright. The 31 headline counts flat-wrong or invented claims; the 23 outdated claims are tallied separately.
These three are the shape of the problem; the record is longer. All 54 flagged claims, verbatim
We audited Liquid Death across the engines on 2026-07-09. Four returned something false or stale. Pick one and read it, word for word.
Fama audits 6 engines in every live audit. The example below is drawn from our public research snapshot, captured and dated, not a live audit.
See what they say about you - free, 30 seconds“founded in 2015 by Mike Cessario”
Claude (no web) · 2026-07-09
Founded 2017 (trademark filed July 2017, incorporated Dec 2018); first sales January 2019.
Six engines, two buyer-style questions per brand (a fact question and a buyer question), ten brands: 120 answers.
The six engines: ChatGPT (search) · Perplexity · Gemini 2.5 Pro · Claude (no web) · Claude (web) · Grok.
Answers were captured through Fama engine adapters, one sample per engine per question, on 2026-07-09. Every load-bearing claim was fact-checked against primary sources as of 2026-07-09.
The instrument
Your real buyer questions go to the AI engines your buyers use. A rules-based reader - not another AI’s opinion - scores how prominently you show up in each answer, 0 to 5. Those blend into one 0-100 Brand IQ.
The number cannot be bought, including from us. Confirming a correction can never move your score: the scorer is provably immune to the verified-facts ledger, and a pinned test asserts the same integer with and without your corrections. Free and paid tiers score the identical question set, so the free number is the real number. The engine weighting is labeled v1-beta in the code and treated as directional; we say so out loud.
Every surface behind the numberProminence, 0-5 per engine answer
Deterministic rules pass over where your name lands. No LLM judge.
The corpus
Not a form a marketer fills in: an extraction. Jinn reads what a brand has actually made and assembles the record in six groups, field by field, each one inspectable.
Every signal is stored as present or absent, never padded. A field the extraction cannot verify drops its block from every downstream surface; nothing is invented to fill a page. Partial beats fabricated. The contract is written down as the wear-field manifest (docs/brand/2026-07-wear-field-manifest.md), which governs exactly which fields may appear where.
Methods & principles
The mechanism these rules run on: how Jinn works
Research by product
The same evidence runs under every Jinn product. Here is which piece speaks to which, and the mechanism page that puts it to work.
The study shows AI guessing brand facts. Agents hand any AI a bounded, verified cut of your record so it reads your brand instead of inventing it.
The study catalogs what AI gets wrong about brands. Ghost drafts in your voice and fact-checks every source it cites before anything publishes.
Vermeer renders only claims you can back, on the same verified record this study is built on. The study itself measures AI search accuracy, not image output.
Silver grounds every script in the same verified record. The research program covers AI search accuracy; Silver applies the same sourcing discipline to video.
The study runs on Fama engines, and Brand IQ is the instrument behind it. This is the product the research sits closest to.
Chart writes strategy only from claims it can source and independently verify, dropping the ones it cannot. The research program measures AI-search accuracy; Chart applies the same source-and-verify discipline to strategy deliverables.
The live index
The study on this page is one fixed snapshot. The AI Brand Index is the running record: we ask the AI engines your buyers use the questions they ask, then publish, verbatim and dated, what each engine gets wrong about a brand. It is ordered worst first, and every verdict carries the source that settles it.
The AI Brand IndexA verbatim, dated, sourced record of what AI assistants get factually wrong about the brands buyers ask about.Open the AI Brand IndexThe shelf
One at launch. Each new study lands here with its capture date and method, and the old ones stay up: the record is the point.
Study 01 · captured 2026-07-09
Six engines, two buyer-style questions, ten brands: 120 answers, 126 load-bearing claims fact-checked against primary sources.
Jinn