YMYL in AI Search: Why Health and Money Get a Stricter Test

By Muntazir · Published 2 November 2026

A study of more than 50,000 German-language health searches, published by SE Ranking in January 2026, found that Google’s AI Overviews cited YouTube more often than any single medical authority — more than any hospital, any government health body, any academic journal. Health questions are about as clearly YMYL, Your Money or Your Life, as a topic gets. If the systems answering them were applying a stricter filter, a smaller role for unvetted video over institutional sources is roughly where it should show up first. The data points the other way.

What does Google actually mean by YMYL?

YMYL is a classification in Google’s Search Quality Rater Guidelines, the manual human raters use to score search results by hand. Since a July 2022 rewrite, the definition is harm-based rather than a list of page types: a topic qualifies when bad information about it “could significantly affect the health, financial stability or safety of people.” Some topics are inherently risky — self-harm, violent extremism. Others only become dangerous if the information is wrong — heart attack symptoms, investing, who can vote. Google’s own illustrative spectrum puts a tsunami evacuation route at clearly YMYL, a weather forecast at possibly YMYL, and a list of music award winners at not YMYL at all. The guidelines were revised again on 11 September 2025, this time widening the “society” subcategory to name election and voting information explicitly.

That revision is the kind of thing that reads as evidence of tightening scrutiny. It’s a guideline for raters, though — a tool Google uses to calibrate what its own ranking systems are doing, not a public description of how any AI model decides what to say.

Are AI answers actually more careful with these topics?

This is the part worth checking before assuming the stricter framing translates into more cautious output. A 2025 study tracked medical disclaimers in generative AI responses to 500 medical questions, 500 mammograms, 500 chest X-rays and 500 dermatology images, across models released between 2022 and 2025. The expectation that models would get more careful over time isn’t unreasonable — more regulatory attention, more publicised cases of bad medical advice, more pressure on vendors to hedge. The study found the opposite. The share of text-model outputs carrying any disclaimer fell from 26.3% in 2022 to 0.97% in 2025. For image-reading models, it fell from 19.6% to 1.05% over a shorter span. By 2025, most models in the sample showed no disclaimer at all, on any of the four question types.

Nothing about this means the guidelines are failing on their own terms — rater guidelines constrain what human raters flag, not what a shipped model says in response to a user’s question, and the two systems aren’t contractually connected. But it does mean the plain assumption — that YMYL guidance getting stricter implies YMYL answers getting more careful — doesn’t hold up against the only direct measurement available.

Where do YMYL answers actually come from?

The SE Ranking study above gives one answer, for one topic and one market. Across 465,823 citations collected from a one-time December 2025 snapshot of German health searches, YouTube accounted for 4.43% on its own — the single largest source, despite ranking only 11th in ordinary organic results for the same queries. A broader group of what the study calls more reliable sources — hospitals, insurers, health foundations — made up about 34% combined. Government health institutions and academic journals together accounted for roughly 1%. More than 82% of the health searches in the study triggered an AI Overview at all.

Two limits are worth stating plainly. The study covers German-language health queries specifically, not financial or legal YMYL topics, and not English-language search. And it measures Google’s AI Overviews, which draw on a live web index, not a conversational assistant answering from memory — a different retrieval mechanism that may not behave the same way. Treat the figures as evidence about one corner of YMYL, not the whole category.

Does an AI answer treat a dentist differently from a plumber?

Here the honest answer is that there’s no aggregate measurement to report. This product’s own Deep Scan asks a model which providers it would recommend in a given market category — the category comes from an earlier read of the site itself, and the question is phrased identically regardless of what that category turns out to be. A probe asking “best accountant near [location]? List the top options” is built the same way as one asking “best plumber near [location]?” Nothing in that mechanism singles out a YMYL category for different handling, and no aggregate data exists yet comparing how often a model actually names a business, or declines to and recommends consulting a licensed professional instead, across YMYL and non-YMYL categories. That comparison would be the natural next thing to measure. It hasn’t been run.

What can a YMYL business actually make checkable?

Not trustworthiness itself — that’s not something a crawler or a model can verify from outside. What’s checkable is the language that would let a rater, or a model summarising a page, see that a claim is attached to something specific. One audit check looks for exactly that kind of language: named professional qualifications, words like “chartered,” “certified,” “accredited,” “licensed” or “registered with,” and recognised credential abbreviations rather than vague expertise claims. The premise isn’t that finding the phrase proves the expertise — it’s that an unevidenced claim in a regulated or YMYL field is worth very little to either a human rater or a model deciding whether to repeat it, and a checkable one is worth more by comparison. The rest of what makes a page’s authorship checkable at all — a named byline, a real address, a professional profile a reader could actually follow — is covered in more detail elsewhere, since none of it is specific to YMYL.

Google’s own rater guidance got more specific in September 2025. In the same stretch of time, the most direct available measurement of actual model output moved the other way — fewer disclaimers, not more, as the models themselves matured. A guideline can specify what a human rater is told to flag. It says nothing about what a shipped model actually generates for the person who asked, and so far nothing in the public evidence says that gap is closing on its own.