Key takeaways
- Of 1,065 checkable claims AI engines made about insurers, 1,059 could be scored, and only 59.1% (626) were fully supported by the public record.
- Most of the rest were imprecise or out of date rather than invented; 84 claims (7.9%) directly contradicted the record.
- Market-share claims were the least reliable (47.7%), mostly because AI repeats rankings that have since changed.
- State Farm still leads every engine; Farmers Insurance (+38 ranks) and Erie Insurance (+35) posted the quarter's biggest climbs.
Brainpan.AI's second quarterly edition of the Insurance AI Visibility Index asked a simple question: when someone asks ChatGPT, Gemini, Copilot, Claude, Perplexity, or Meta AI about an insurance company, what does the AI actually say, and how often does it say something true?
The answer, from this quarter's data: not often enough. Brainpan's research team identified 1,065 specific, checkable claims that AI engines made about insurance brands: financial-strength ratings, market share figures, J.D. Power rankings, and state-by-state availability. Of those, 1,059 could be checked conclusively against authoritative sources (the rest were too ambiguous to verify either way). Only 59.1% of that scoreable set, 626 claims, held up as fully supported. The rest were partially supported, outdated, or flatly contradicted.
That gap is the headline of Brainpan's Q3 2026 Insurance AI Visibility Index, out now. It's the first quarter the study has measured accuracy at all, and it turns out to be the most consequential finding in the dataset.
What the study measures, and what's new this quarter
The Q3 2026 edition covers 226 insurance brands, up from 134 in Q2. The core visibility leaderboard, the rankings below, is scored across five AI engines (ChatGPT, Gemini, Perplexity, Copilot, and Claude) on 300 prompts and 1,500 responses. Meta AI runs as a sixth, supplementary surface this quarter. Two dimensions are new this quarter, layered onto the full six-engine corpus of 1,800 responses rather than collected separately: Sentiment, the tone of how AI engines describe a brand, and Accuracy, whether the specific factual claims AI makes about a brand actually check out.
On the visibility side, the leaderboard held mostly steady at the top, with real movement just beneath it. State Farm holds the #1 spot on every one of the five core engines, and on Meta AI too, a sweep it also held in Q2. Beneath it, USAA climbed into second place, GEICO moved up into third, and Progressive dropped two spots to fourth. The quarter's biggest movers were Farmers Insurance, which climbed 38 ranks, and Erie Insurance, which climbed 35, the two largest gains among brands with a reliable mention sample in both quarters. Both get their own in-depth analysis in the full research: Farmers Insurance and Erie Insurance.
On Sentiment, also new this quarter, the AI-generated portrayal of insurance brands leans positive in aggregate: 64.8% of scored mentions came across positive, 18.3% neutral, 14.8% mixed, and 2.1% negative, across every tracked brand. That's its own section in the full benchmark. This piece focuses on the other new dimension, Accuracy, which turned out to be the bigger story.
But visibility was never the whole question Brainpan set out to answer. Being mentioned by an AI assistant is one thing. Having that mention be true is another, and until this quarter, Brainpan hadn't measured the second half either.
The accuracy gap, in detail
Here's the full breakdown of the 1,059 scoreable claims: 626 (59.1%) were fully Supported by the public record. 185 (17.5%) were Partially supported, generally true but imprecise or missing context. 164 (15.5%) were Outdated, claims that were accurate once and no longer are. 84 (7.9%) were Contradicted outright, a distinct verdict from Outdated, reserved for claims that don't match the public record regardless of timing.
That distinction matters. It would be easy to round "59.1% fully supported" down to "AI is wrong 41% of the time," but that's not quite what the data shows. Most of the non-supported share isn't AI inventing facts from nothing; it's AI being imprecise (Partially supported) or behind the times (Outdated), and an outdated claim presented as current can mislead a shopper just as easily as a contradicted one. Claims that directly contradict the public record, with no timing excuse available, sit at 7.9%, a smaller and more specific number worth keeping separate from the rest.
Accuracy also wasn't uniform across claim types. Rating-agency score claims (AM Best, Moody's, and similar) held up best, with a 65.3% Verified Accuracy Rate. Market-share claims held up worst, at 47.7%, and the shortfall there skews heavily toward timing rather than invention: among market-share claims that didn't hold up, Outdated verdicts outnumbered Contradicted ones by more than 3 to 1. AI is mostly repeating who used to rank #1 or #2 long after the ranking changed hands, the same stale-facts pattern described below, rather than making up numbers from nothing.
| Claim type | Claims checked | Scoreable | Verified Accuracy Rate |
|---|---|---|---|
| Rating agency score | 357 | 354 | 65.3% |
| J.D. Power ranking | 185 | 184 | 62.5% |
| Binary state availability | 284 | 282 | 58.9% |
| Market share | 239 | 239 | 47.7% |
A small number of claims per category (6 of 1,065 total) couldn't be matched to a clear, checkable fact; Verified Accuracy Rate is calculated against each category's scoreable claims, not the raw number checked. The full breakdown, including results by AI engine, is on the AI accuracy research page.
Two patterns show up again and again
Two specific failure modes recurred across multiple AI engines independently, not as isolated glitches in one system but as a pattern consistent enough to name.
Invented ratings. In one notable case from this quarter's data, different AI engines confidently assigned a company a financial-strength rating it doesn't actually hold, each engine naming a different grade, none of them correct. The company's own public filings disclose that it isn't rated at all for the category AI kept citing. This wasn't one model making one mistake; it was multiple, different AI engines converging on the same kind of fabrication.
Stale facts. A second pattern involves AI repeating a rating, a ranking, or a market position that was accurate at some point and has since changed, without the update ever catching up. In multiple instances, trade-press and regulatory sources confirmed a rating downgrade, a lost ranking, or a shift in market position that several AI engines still describe as if it never happened.
Brainpan doesn't name the specific brands behind either pattern publicly; that level of detail is reserved for the private brand diagnostic Brainpan provides directly to the companies involved. But the patterns themselves, invented ratings and stale facts, are real, repeatable, and worth any insurance marketer's attention regardless of which brand they run.
Why this matters beyond the research itself
It would be fair to ask whether any of this matters if most people aren't actually making decisions based on what an AI assistant tells them about an insurance company. J.D. Power's own 2026 U.S. AI Insurance Experience Study, an independent study with no connection to Brainpan, suggests they already are.
That study, based on 8,352 customer evaluations across 24 insurance brands and fielded in June and July 2026, found that roughly 3 in 10 (29%) auto and home insurance customers already use AI for insurance-related activities, researching coverage, comparing quotes, and servicing existing policies among them. Among those who used AI specifically to research coverage, about 4 in 10 (37%) changed their policy based on what the AI told them; among those who used it to shop for a policy, about 4 in 10 (42%) went on to purchase one as a result.
AI-mediated insurance shopping isn't a hypothetical anymore. It's already shaping what a meaningful share of customers buy. That's exactly why the accuracy gap in Brainpan's own data carries real weight: a 59.1% Verified Accuracy Rate isn't just a research curiosity; it's a description of the information environment a growing share of insurance customers are already shopping inside.
What this means for insurance marketers
"Visibility tells you whether AI is talking about your brand. Accuracy tells you whether it's telling the truth," said Kevin Walsh, founder of Brainpan.AI. "This quarter is the first time we've been able to measure both on the same data, and the gap between them is bigger than most marketing teams probably assume."
If a customer is already using AI to decide whether to change their coverage or switch carriers entirely, and J.D. Power's own numbers say a meaningful share already are, then what AI says about your brand, accurately or not, is doing real work on your pipeline whether anyone signed off on it or not.
For a marketing team wondering what to do with any of this, three starting points:
- Check what AI engines are saying about your own financial-strength rating, market position, and state availability, across more than one engine. This quarter's data shows real disagreement between engines is common, not rare.
- Document anything wrong against a dated, citable, authoritative source: your own current AM Best or Moody's listing, your state filing, your own investor materials, not just an internal note.
- Check back. Several of the stale facts in this quarter's data were years old by the time they were caught. A single correction request doesn't guarantee every AI engine updates, or updates at the same pace.
Brainpan's Insurance AI Visibility Index tracks how 226 insurance brands are mentioned, recommended, and described across six major AI engines, and now, for the first time, whether what's said about them holds up.
Read the Q3 2026 Insurance AI Visibility Index
The complete leaderboard, the Q2-to-Q3 index movement, and the full Sentiment and Accuracy findings are live now.
Read the Q3 benchmark → Download the full report (PDF)Methodology note: Brainpan's Q3 2026 Insurance AI Visibility Index covers 226 insurance brands that resolved to a canonical entity with scoreable Q3 data (229 distinct brand candidates were identified before resolution). The core visibility leaderboard is built from 300 prompts and 1,500 AI responses across five engines: ChatGPT, Gemini, Perplexity, Copilot, and Claude. Meta AI runs as a sixth, supplementary surface. Sentiment and Accuracy are scored across the full six-engine corpus of 1,800 responses. Accuracy verdicts are based on 1,065 specific factual claims checked against primary and authoritative secondary sources; the research team reached a definitive verdict on 1,059 of them, 99.6% of the 1,063 claims specific enough to check, a measure of how resolvable the verification process is, not of how often AI hedges or declines to answer. Full methodology: AIVI™ Methodology v1.1.
