New research across 2,100 Medicare plans reveals major differences in how Google AI and Microsoft Copilot present and attribute plan facts—and finds attribution concentrated among a small group of publishers. Read the report: https://www.medicareplans.com/research/medicare-ai-visibility-study/
When you ask an AI system about a Medicare plan, where does the answer actually come from?
AI systems increasingly do something different from traditional search. Instead of simply returning webpages, they assemble answers for you.
Trust Publishing Institute measured that process across more than 2,100 Medicare plans and found something striking.
A relatively small group of publishers accounted for most of the source attribution we observed. Google AI and Microsoft Copilot also constructed substantially different information profiles for the same class of Medicare plans.
The findings come from the Medicare Visibility Report: September 2026 Baseline,.
The study examined 39 categories of Medicare plan information across approximately 2,100 CMS Plan I.D. entities.
The researchers don't simply ask which websites ranked or which publishers received citations.
For a specific Medicare plan and fact, did the answer engine state that fact? And if it did, which source did the system associate with it?
That's important because a citation alone doesn't tell us what role a publisher played.
One publisher might be associated with a premium, another with eligibility, and another with supplemental benefits. Attribution also doesn't establish that information is accurate, current or original to the publisher.
As the report put it: “Search engines discover documents; answer engines assemble assertions.”
So, what did the report find?
First, source attribution was much more concentrated than the number of available sources might suggest.
Google AI exposed attribution from 188 source domains. Microsoft Copilot exposed attribution from 74.
But on Google AI, just three sources accounted for approximately 70 percent of observed attribution. The top 10 accounted for approximately 91 percent.
On Copilot, the top three accounted for approximately 89 percent, and the top 10 accounted for approximately 99 percent.
That doesn't mean those publishers control the AI systems. It means a relatively small group accounted for a disproportionate share of the source-attribution relationships.
The second major finding involved the answer engines themselves.
Out of the 39 Medicare plan facts we measured, Google AI asserted approximately 54 percent. Copilot asserted approximately 68 percent.
But when researchers measured how often those assertions included identifiable sources, the pattern reversed.
Google AI exposed source attribution for approximately 94 percent of the facts it asserted. Copilot did so for approximately 50 percent.
So Copilot generally provided a more complete representation of a plan, while Google exposed sources for a much larger share of what it stated.
The differences became even more pronounced at the individual-fact level.
For emergency-room copays, Copilot asserted the information for nearly 91 percent of measured plans. Google AI did so for about 25 percent.
For Medicare eligibility, it was nearly 97 percent for Copilot and about 38 percent for Google.
For residency requirements, it was roughly 90 percent for Copilot and 37 percent for Google.
But Copilot didn't provide more of everything. Google AI asserted some information, including service area and insulin cost, more frequently.
So there isn't one simple measure of whether an AI system “knows” a Medicare plan. Its observable representation is fact-specific.
The researchers then looked at the publishers associated with those facts.
Medicare dot org was the largest measured source on both systems.
Google AI associated Medicare dot org with more than 34,000 assertion-level source relationships. Q1Medicare followed with more than 20,000, and MedicareAdvantage dot com with more than 14,000.
On Copilot, Medicare dot org had more than 22,000 attributed assertions, followed by MedicareAdvantage dot com and Medicare Plans dot com.
Those aren't quality scores or trust rankings. They answer a narrower question:
Which publisher does an answer engine associate with a particular fact about a particular Medicare plan?
That brings us to the larger issue.
CMS, the Centers for Medicare and Medicaid Services, is a primary official source of Medicare plan data. Medicare carriers also publish detailed information about their products.
Yet the baseline found that AI systems frequently exposed attribution to private publishers.
That raises an infrastructure question.
How should authoritative Medicare information be structured and distributed so machines can reliably identify the correct plan, year, geography, eligibility requirements, costs and benefits?
The study doesn't answer that question.
What it gives us is a baseline for observing what an answer engine asserts, what it omits, which sources it exposes, and how those relationships differ by engine, plan, fact and publisher.
There's also an immediate implication for consumers.
AI-generated Medicare answers can be useful starting points, but they shouldn't be treated as the final authority for an enrollment decision.
Plan availability, eligibility, costs, provider networks, prescription-drug coverage and benefits can depend on plan year, service area and individual circumstances.
Before enrolling, important information should be verified through official Medicare resources and information supplied directly by the plan.
The Street dot com recently examined the consumer side of this issue. Its reporting tested how small differences in the words consumers use can produce materially different AI responses and examined earlier Medicare Visibility Monitor research.
The Street did not independently reproduce every quantitative finding in this study, but its reporting provides independent scrutiny of the broader consumer-information problem motivating the research.
The Medicare Visibility Report is an independent measurement project of Trust Publishing Institute and Medicare Plans dot com.
They conducted the research, methodology and analysis behind the report.
Medicare dot org is owned and operated by Health Network Group, an Allstate company. Neither Medicare dot org nor Allstate commissioned the study or determined its methodology, findings or conclusions.
This baseline does not establish causality, factual accuracy, consumer exposure, referral traffic, commercial influence or cross-engine semantic agreement.
It measures observed answer-engine behavior under a defined protocol.
Because knowing that a source was cited isn't enough.
We need to know what was asserted, about which plan, by which system and which source was associated with the fact.
That's what the Medicare Visibility Report was designed to measure.
The complete September 2026 report and study methodology are available at Medicare Plans dot com. Trust Publishing Institute City: Bullhead City Address: 1800 Club House Drive #93 Website: https://trustpublishing.org/