The Comparison Paradox: Insurance Shoppers and Recommendation Behavior in the Answer Engine

The most-compared insurance shopper on record

J.D. Power has run its insurance shopping study long enough to know what a normal year looks like, and 2026 was not one. Shoppers pulled an average of 3.5 quotes before buying, the highest number the study has ever recorded, based on 12,437 insurance customers who had requested auto quotes in the prior six months, fielded from January 2025 through January 2026. Digital purchasing kept climbing too: nearly half of new auto policies, 48 percent, were bought online, up from 36 percent just five years earlier. Interestingly, the overall share of people shopping at all eased slightly, from 57 percent to 53 percent year over year, even while it stayed historically elevated. Fewer people are shopping, but the ones who do shop are comparing harder than ever.

Read together, these two numbers describe a market where comparison has become the default posture rather than the exception. A shopper who used to request one or two quotes out of habit now treats 3.5 as typical, and does more of that comparing without leaving a screen. That raises the stakes on whatever surface first earns a shopper's attention, because the days of a single agent conversation deciding the outcome are visibly behind us.

AI joins insurance shopping, unevenly

In the companion 2026 U.S. Auto Insurance Study, fielded April 2025 through April 2026 among 52,216 customers in its 27th consecutive year, J.D. Power found that 32 percent of shoppers used an AI tool somewhere in the process. That is a genuine foothold, not a rounding error, but it is still a minority behavior. The same study found a nearly matching share, 33 percent, who said the content an AI tool gave them was not helpful. The study does not say these are the same shoppers, but the two numbers sit close enough together to be worth reading side by side.

That split matters more than either number alone. AI is not failing to get adopted, and it is not succeeding cleanly either. It is being tried by a meaningful third of shoppers, and a similarly sized third comes away unimpressed by what it gives them, which is a very different picture from either an AI takeover or an AI dead end. For a business watching this category, the read is that AI has become one more channel competing for a shopper's early attention, alongside carrier sites, comparison tools, and human agents, and it has not yet proven itself the most trustworthy of the group.

AI is not failing to get adopted, and it is not succeeding cleanly either. It is being tried by a third of shoppers, and a similarly sized third comes away unimpressed.

The confidence effect, and what it is not proof of

The behavioral signal that stands out most is switching. Shoppers who used AI during their search were 1.3 times more likely to switch insurers than shoppers who did not, and they reported feeling more confident in the choice they made. That is a real, measured behavioral shift documented by J.D. Power, not a marketing claim.

What the evidence does not yet show is why. It is entirely possible that AI tools are surfacing genuinely better matches and shoppers are switching because the new option is objectively right for them. It is equally possible that a confident-sounding answer, correct or not, is enough on its own to move someone off an incumbent insurer. The study measures the switching and the confidence; it does not measure whether the AI's underlying recommendation was accurate. Given what the rest of the evidence shows about how often these engines disagree with each other and get specific questions wrong, that open question deserves to stay open rather than get resolved in AI's favor by assumption.

The comparison tool that most shoppers never see

Away from AI entirely, J.D. Power's 2026 U.S. Insurance Digital Experience Study found a gap sitting in plain sight on carriers' own websites and apps. Only about a third of shoppers, 33 percent, ever encounter a genuine cross-brand price comparison tool, one that shows competing insurers side by side. Twenty-seven percent see only a same-insurer comparison of their own carrier's policy options, and 28 percent see no comparison tool at all.

The consequence of that gap is not abstract. Shoppers who do encounter a real comparison tool are nearly twice as likely to consider purchasing, 39 percent versus 21 percent for shoppers who see none. That is one of the clearest lines in this entire body of evidence: comparison access changes purchase consideration by a wide margin, and most insurance shoppers today are simply not being given that access on the sites built to sell to them. Before a single AI engine enters the conversation, the industry is already leaving comparison-driven consideration on the table.

Shoppers who encounter a real comparison tool are nearly twice as likely to consider purchasing. Most insurance shoppers today are not being given that access at all.

When the answer engines cannot agree with each other

This is where the evidence gets harder to trust, and where we tier it carefully. A single-vendor Australian study by Somantra tracked 20 insurance brands across 34,278 real consumer conversations in May 2026 and found that Google AI Overviews and ChatGPT recommended the same top insurance brand only 27.9 percent of the time, up from 23.7 percent two months earlier. Put plainly, on matching queries, the two engines disagreed roughly seven times out of ten. The same study found that 70 percent of the tracked AI responses named no insurance brand at all, while the visibility that was granted concentrated heavily: nine brands accounted for 90 percent of ChatGPT's insurance mentions.

We are treating this as a contested-tier finding, not an established one, because it comes from a single vendor's monitoring tool, covers Australian conversations rather than the US market, and has not been independently replicated. It is real data from a named, dated study, and it is consistent with what we would expect from generative engines drawing on different training data and different retrieval behavior. But it should not be read as a settled description of the US insurance-answer market. What it does establish, credibly, is that a shopper asking two different AI engines the same insurance question today has a real chance of getting two different answers, or no clear answer at all.

Who is actually answering, and whether they are licensed to

A follow-up Somantra study went further, analyzing 2,437,107 AI citation records across 28,725 domains between November 2025 and July 2026, and found 38 domains being cited as insurance sources that held no Australian Financial Services Licence at all, the credential that permits giving financial and insurance guidance in that market. Those unlicensed domains made up 1.97 percent of ChatGPT's insurance citations, against 0.10 percent for Google AI Overviews, a roughly nineteen-fold gap between the two engines in how often they cited an unlicensed source.

Again, this is Australian, single-vendor evidence, and we are not presenting it as a US figure. But it is a documented instance of the exact failure mode that should concern any business in a licensed industry: an AI engine citing a source that has no legal standing to give the advice being cited, at a rate that varies by nearly twenty times depending on which engine a shopper happens to ask. Australia's securities regulator, ASIC, issued guidance in March 2026 warning that AI tools used for financial and insurance guidance have limitations that could lead to inaccurate or inappropriate suggestions, and recommended shoppers verify claims against trusted, licensed sources. Whether an AI platform bears any liability for citing an unlicensed source remains, per that same reporting, a legally unresolved question.

How wrong can a single answer be

The sharpest number in this study comes from the narrowest evidence base, and we want to be direct about both facts at once. Choice Mutual, a licensed insurance agency, had its own CEO manually review 500 Google AI Overviews responses to life insurance questions and found errors in 57 percent of them. A separate reviewer, a Medicare specialist at the same agency, reviewed 500 Medicare-related AI Overviews and found a 13 percent error rate. The study reports no inter-rater reliability check and no confidence interval, and it has not been independently replicated by another organization.

We are keeping this finding, tiered as contested, because a single reviewer's manual audit is still real evidence of specific, documented errors, not an invented statistic, and because the gap between the 57 percent life-insurance error rate and the 13 percent Medicare error rate is itself informative: it suggests AI accuracy on insurance questions may vary sharply by sub-category rather than being uniformly good or uniformly bad. What it cannot responsibly support is a claim that AI answers are wrong 57 percent of the time across insurance generally, in the US market, as measured fact. That would overstate a single agency's internal review into a market-wide statistic the evidence does not contain.

The regulatory floor is being built underneath this, in real time

Lawmakers are not waiting for the accuracy question to resolve itself. Legal-industry tracking via Orrick reports that the Future of Privacy Forum counted 98 chatbot-specific bills across 34 US states as of spring 2026, with several states advancing rules that would require AI tools to disclose when they lack professional licensure and prohibit them from implying a license they do not hold. The source page was not accessible to us to confirm firsthand, so we are presenting this count as reported rather than independently verified. Orrick and the Future of Privacy Forum are established, frequently cited authorities on state chatbot regulation, which is why we are keeping the figure with that caveat attached rather than leaving it out.

Insurance-industry compliance commentary is arriving at the same conclusion from a different direction. Guidance circulated by PIA of Wisconsin explicitly warns agencies never to delegate a function to a chatbot that would legally require a licensed human agent if performed directly, framing this as an active, unresolved errors-and-omissions exposure for carriers and agencies today, not a settled or quantified problem. Neither source produces a hard percentage. Both describe the same structural gap: the regulatory apparatus for licensed professional advice was not built with generative AI in mind, and it is visibly catching up rather than leading.

What the paradox actually is, and where it points

Put the pieces together and the paradox in this study's title becomes precise rather than rhetorical. Shoppers are not under-comparing. They are pulling more quotes than ever, 3.5 on average, and a third of them are actively recruiting an AI tool into that comparison. The problem is not shopper behavior. It is that the infrastructure meant to reward that behavior, whether a carrier's own comparison tool or an AI engine's recommendation, is inconsistent in ways the shopper cannot see. Only a third of shoppers get a real cross-brand comparison tool from carriers themselves. Two leading AI engines agree on a top brand well under a third of the time. And at least one documented, if narrow, study found more life-insurance answers wrong than right.

One gap is worth stating plainly: this evidence does not tell us whether AI engines favor national aggregators over locally licensed independent agents, a claim it would be tempting to make and that the original framing of this research initially assumed. No rigorous US study answers that question yet. It is a real, open frontier, not a measured fact. What the evidence does establish is that the answer layer itself, across engines, across citations, across licensing, is not yet reliable enough to be the single source shoppers are increasingly treating it as. For a carrier, agency, or insurance brand, that instability is not a reason to ignore where AI-driven attention is going. It is the reason to measure it directly, market by market, rather than assume any single engine's answer speaks for the category.