AI search and YMYL content: what actually differs for health, finance and legal pages

What changes by vertical: health, finance, legal, safety

"YMYL content" is usually treated as one undifferentiated thing. It is not. The four categories differ in what a wrong answer costs, in what counts as a primary source, and in how fast the underlying guidance moves — and those three differences drive every practical decision on the page.

Three implications follow. Health and safety content carries the shortest tolerance for an out-of-date page, so a review schedule matters more there than a citation tactic ever will. Finance content is the easiest of the four to keep honest, because the rules generally arrive with a date attached — put that date on the page. And legal content is the one where a nationally correct answer is confidently wrong one border away, which is the single most common way a well-intentioned YMYL page misleads someone.

What YMYL means, and why it is not an AI ranking factor

The term comes from Google's Search Quality Rater Guidelines, where it tells human raters to look harder at content that could seriously affect a person's health, money, safety or well-being if it is wrong. It is a human-evaluation concept. Applying it to AI-citation behaviour is an extension, not a documented AI-company policy.

That distinction survives contact with the documentation. Google's published guidance on AI features and its optimisation advice for site owners says nothing about a separate track for high-stakes topics. Neither do the retrieval descriptions in Anthropic's web search documentation or Perplexity's developer docs. The absence is not evidence that nothing happens. It is the reason this page is a hypothesis rather than a finding. What GEO is covers the wider discipline this sits inside.

What is known, and what is only assumed

What does exist is descriptive work on how engines pick sources in general — how ChatGPT chooses its sources, how each platform cites differently, and a 2026 cross-platform sourcing comparison. All three are useful. None segments by stakes.

Nor do the quantitative studies. Semrush's AI Overviews analysis, Ahrefs' work on which sites AI Overviews cite most and Profound's platform citation patterns all report behaviour averaged across a mixed query set, as do the roundups that quote them — including GEO statistics and the master AI SEO statistics index here. The same blind spot runs through the domain-level work: the most-cited domains analysis and Perplexity's Reddit dependency describe general sourcing habits, and whether a forum thread is an acceptable source shifts enormously depending on whether the question is about a router or a drug interaction.

Does the 2018 "Medic Update" tell us anything?

It tells us the question is reasonable, and nothing more. A widely discussed 2018 Google core algorithm update, nicknamed the "Medic Update" by the SEO community, had a large visible impact on health and wellness sites and is broadly understood to have raised the quality bar for YMYL content in classic search ranking. That is a major search company treating YMYL as its own category, at least once, in a publicly observed way.

No comparable AI-citation event has been documented. The closest current analogue is the observation that AI Overview citations from top-ranking pages have dropped sharply, which shows the source mix moving without telling us whether it moves differently by topic. Meanwhile the delta between AI Mode and AI Overviews shows two surfaces from the same company already disagreeing about sources, which is a reminder that "AI search handles health carefully" is not a claim that can be true or false as stated.

Three ways differential handling could arise

It helps to separate the mechanisms rather than treating "engines are careful with health" as one idea. They leave different fingerprints, and only one of them is visible to a citation study.

Hypothesis That asymmetry is a serious limitation for any study built on citation counting, including anything this site runs. It is also why the tactic evidence scoreboard grades several confidently marketed YMYL tactics as untested rather than ineffective — the two words mean very different things and get used interchangeably.

What would fake a YMYL effect?

Suppose someone runs this study and finds that YMYL queries produce more institutional citations. Before calling that a safety policy, rule out the alternatives. Several are quite likely.

None of these are exotic, and two of them — supply-side composition and link structure — are strong enough that they could plausibly account for the entire expected effect. The link-structure confound is particularly awkward, because whether links or mentions are associated with citation is itself unresolved, and work on the citation-to-mention visibility gap suggests the two signals come apart more often than expected.

How would you test this yourself?

This is not a study only a well-funded lab can run. Someone with patience and a spreadsheet could produce a more credible answer than currently exists.

Step one carries most of the difficulty, because classification is where the judgement lives. "Maximum daily dose of paracetamol" is obviously health; "how do I file for bankruptcy" is obviously legal. The trouble is the middle. Is "best running shoes for flat feet" a product query or a health query? "Is this mushroom edible" is arguably the highest-stakes query on this page and looks nothing like a classic YMYL topic. The honest handling is to publish the rule and the borderline cases together, say who classified the queries, whether anyone did it independently, and how often they disagreed.

Step five is the one most often skipped and the one that makes the rest checkable. Whatever the conclusion, the query list, timestamps and captured citations should be downloadable — which would put such a study ahead of nearly everything circulating in this field. If this site runs the YMYL slice of the Citation Index and finds nothing, that null lands in the null results registry with the same prominence a positive finding would get.

What should a YMYL publisher do today?

Everything worth doing here is defensible without resolving the citation question, which is a comfortable position and a rare one in this field.

Run that list before worrying about engines at all: a page that fails it has a reader problem, and a reader problem is the more urgent one. Two adjacent tactics are worth the effort for independent reasons. Structured markup makes an author, a review date and a medical reviewer machine-readable rather than merely present — the published evidence on schema and AI citations is mixed, but the accessibility gain is not in dispute. And keeping pages current matters disproportionately when the underlying guidance is legally binding.

On the mechanical side, an assistant cannot cite what it cannot fetch. A technical GEO audit covers whether retrieval bots are allowed by your robots.txt — the blocking census found this is a widespread accidental failure — and whether your content survives without JavaScript.

The costs are real and specific to this category. The most serious is the temptation to write more assertively than the evidence supports, because hedged language feels less quotable; in these verticals the caveat is often the part carrying the safety, and a reader who acts on the trimmed version may be worse off than if they had read nothing.

The second is maintenance — if you cannot commit to reviewing a YMYL page on a schedule, it is fair to ask whether you should publish it. The third is credential inflation: adding an expert reviewer's name to a page they did not meaningfully review is a form of fabrication, with professional and legal exposure far beyond anything an SEO tactic should carry.

Common misreadings

"Engines prefer official sources for health, so my health blog cannot get cited." That overstates a hypothesis into a rule. Even if institutional sources dominate, they do not answer every question, and specific well-sourced content often covers ground official guidance does not.

"YMYL is an AI ranking factor." It is not documented as one anywhere. It is a concept from a human rater handbook for classic search, and an analogy is not a factor.

"If the engine cites a government source, the answer is safe." A good citation does not guarantee the answer summarised it correctly. Accuracy of the generated text is a separate measurement from the quality of the sources beside it.

"AI referral traffic will tell me whether it worked." Not at this volume and with this attribution. AI referral traffic is small and undercounted, and on YMYL topics the answer-without-click outcome that zero-click research describes is arguably the normal one.

Open questions

Open question Do engines decline to answer YMYL queries at a measurably higher rate, and does that rate differ between engines? Refusals are rarely reported in citation studies.

Open question Does citation quality on YMYL topics vary by country and language? Health guidance, tax law, tenancy law and drug approvals are all national, and the institutional corpus a retrieval system can draw on is unevenly distributed across languages. This is the equity question nobody appears to be asking.

Open question When an engine cites an authoritative source, how often does the generated answer preserve that source's caveats? Citation accuracy and answer accuracy are different measurements.

Open question Does the presence of a named, credentialed reviewer correlate with citation on YMYL queries, controlling for domain authority? Widely assumed, not publicly measured.

Open question How quickly does updated official guidance propagate into answers? When a regulator revises a dose or a deadline, the lag before assistants stop repeating the old figure is a safety property, not a marketing one. The crawl-to-citation latency study measures the general case; the YMYL case matters more.

Limitations

• No measured YMYL citation effect is reported here — we located no disclosed-method study comparing YMYL and non-YMYL citation behaviour, and absence of a located study is not the same as absence of an effect.
• YMYL classification itself involves judgment calls at the query level — the classification method used for any future study will need to be published alongside results.
• The per-vertical table is editorial, not empirical. It describes what changes about publishing duty of care in each category; it is not a measurement of engine behaviour by vertical.