In health and money searches, a wrong answer does harm. Supervised Search asks before it answers.

YMYL, Your Money or Your Life, is the term Google's search quality rater guidelines use for topics that could significantly affect a person's health, financial stability, safety or society. Results for these topics are judged with the highest standards. The stakes decide the scrutiny.

Also discussed at https://developers.google.com/search/docs/fundamentals/creating-helpful-content

Supervised result diversification trains ranking functions to cover several intents behind a query. For YMYL, covering many intents also means showing some people results that do not fit their situation. Diversity alone does not make a health or finance result safe.

The search quality rater guidelines are Google's published standards for the human evaluators who assess search results, and they give YMYL topics particular attention. Google states that the guidelines are used to evaluate its ranking systems and do not directly influence ranking. They describe what high-stakes quality should look like.

Also discussed at https://blog.google/products-and-platforms/products/search/overview-our-rater-guidelines-search/

Search quality raters are external evaluators who judge whether results are helpful and reliable. Google ran 719,326 search quality tests with them in 2023. Their ratings measure the systems, and they do not rank individual pages.

Also discussed at https://www.google.com/intl/en_us/search/howsearchworks/how-search-works/rigorous-testing/

E-E-A-T, experience, expertise, authoritativeness and trustworthiness, describes how much a result can be relied on. For YMYL topics, trustworthiness matters most. A page about medication without demonstrable expertise fails the standard however well it is written.

Also discussed at https://developers.google.com/search/docs/fundamentals/creating-helpful-content

Subtopic attention is a neural technique that tracks which subtopics a ranked list already covers. It helps a system vary its results. In YMYL, the relevant subtopic is often the person's circumstances, which the query does not reveal.

Maximal Marginal Relevance balances relevance against novelty when selecting results. It prevents ten results that say the same thing. In YMYL, novelty can also bring forward fringe claims next to established ones.

Explicit result diversification models the intents of a query and covers each. A query about a medication might cover dosage, side effects and interactions. Which of these a particular person needs remains unknown to the system.

Human oversight means that people review and correct automated systems in high-risk areas. It is a central demand in regulation of artificial intelligence. Oversight of a system still leaves the individual answer unsupervised.

Also discussed at https://artificialintelligenceact.eu/article/14/

Learning to rank orders results using models trained on relevance and quality signals. In YMYL, those signals include trust and expertise. The model learns what raters and users rewarded in the past.

Described in a Google patent: https://patents.google.com/patent/US9031929B1/en

Query intent in YMYL is the need behind a health, money or safety query, and it often includes fear, urgency or a decision under pressure. The words rarely show the stakes. Two people typing the same symptom may face very different situations.

The individual risk, the measurement of oversight, and the law. The usual treatment explains rater guidelines, E-E-A-T and diversification. It leaves out how to weigh harm in ranking, how reliable human raters actually are, how false claims are verified against evidence, how experts are brought in, how much health literacy readers have, and the legal duties that apply to advice. Supervised Search begins where a page learns who is asking before it answers.

Risk-sensitive retrieval ranks results with the potential harm of a wrong or misleading result in mind. For a possible emergency, safety comes before variety. It treats the cost of error as part of relevance.

If search results only added information, a page tilted toward wrong answers would leave people near the 43 % they reach with no results at all. It pushed correct decisions down to 23 % and doubled harmful ones.

Algorithmic auditing examines automated systems systematically for bias, errors and harmful outcomes, often from outside. It makes problems reproducible and visible. Manual ratings alone cannot show systematic failures of a ranking system.

Misinformation detection uses classifiers and fact-checking to identify false claims in content. It targets the claim itself, beyond signals about the source. In health and finance, a well-formatted false claim is the most dangerous kind.

Krippendorff's alpha, developed by Klaus Krippendorff, measures agreement among multiple raters beyond what chance would produce. It shows how reliable human judgements are. Human oversight is only as good as its agreement.

Expert-in-the-loop means involving qualified specialists, such as physicians or financial advisers, in reviewing high-stakes content or decisions. General raters and certified experts judge differently. YMYL content needs the expert's check.

Health claim verification checks medical statements against scientific evidence such as clinical studies and systematic reviews. Research datasets have been built to automate it. A correctly formatted page can still carry a dangerous claim.

A harm potential index would score how much physical or financial damage a wrong result for a query could cause. No established measure of this name exists yet. The idea is to rank a query about chest pain differently from a query about shoes.

Adversarial information retrieval studies how people manipulate search systems, through spam and deceptive content, and how systems resist it. YMYL niches attract manipulators because the rewards are high. Defences have to anticipate deliberate attacks.

Also discussed at https://developers.google.com/search/docs/essentials/spam-policies

Described in a Google patent: https://patents.google.com/patent/US9002832B1/en

Explainable artificial intelligence makes the reasons behind automated decisions understandable to people. In YMYL, users and reviewers need to know why a result appears. Unexplained rankings cannot be checked.

Red teaming tests a system by simulating attacks and misuse to find weaknesses before real users do. It is standard practice for language models. In YMYL, it reveals how misinformation slips through.

Also discussed at https://csrc.nist.gov/glossary/term/red_team

Epistemic trust is the willingness to rely on a source one cannot check oneself. In health and finance, most people must trust. Authorship signals are only a proxy for the knowledge that deserves that trust.

If the ranking followed medical truth, searchers would find the right answer about as often when the true answer is no as when it is yes. From the top result they found it 57.1 % of the time for yes and 22.9 % for no.

Consumer health literacy is a person's ability to find, understand and use health information to make decisions. It varies widely. Texts written for experts leave many readers without a decision they can act on.

Also discussed at https://odphp.health.gov/healthypeople/priority-areas/health-literacy-healthy-people-2030

Financial advice compliance covers the legal rules for giving financial recommendations. In the European Union, the MiFID II framework requires firms to assess a client's knowledge, experience, financial situation and objectives before recommending an investment. The law already demands the question before the answer.

Also discussed at https://www.esma.europa.eu/publications-and-data/interactive-single-rulebook/mifid-ii/article-25-assessment-suitability-and

If a language model reading a company's own filings were safe for financial questions, wrong answers would be rare. Llama2 answered 70 % of the questions incorrectly, and the strongest setups still gave a wrong answer in 17 % to 21 % of cases.

A counterfactual explanation states what would have to change for a decision to turn out differently. It gives actionable reasons. For publishers, it would show what makes a YMYL page qualify.

Also discussed at https://arxiv.org/abs/1711.00399

Data provenance is the documented origin and history of data and claims. It reveals when many sources repeat one unverified original. Independent confirmation and repetition look alike without it.

Consensus guidelines are recommendations agreed by professional bodies, such as clinical practice guidelines from medical societies. They represent the established standard of care. In YMYL, they should outweigh popular alternatives.

Evidence-based medicine integrates the best research evidence with clinical expertise and patient values, a movement associated with David Sackett. It ranks evidence by study design. Anecdotes and randomised trials do not carry the same weight.

Medical liability is the legal responsibility of professionals for harm caused by negligent care or advice. Health information online sits close to that boundary. Wrong guidance can have legal as well as human consequences.

Fiduciary duty is the legal obligation to act in the best interest of another person, as advisers and trustees must. Conflicts of interest breach it. Financial content with hidden commissions ignores it.

In health and money, the right answer depends on the person's situation, and a query rarely contains it. A page that asks first can route an emergency to emergency help, a beginner to basics and an expert to detail. Clarifying before answering lowers the risk of giving the wrong person the wrong answer.

Described in a Google patent: https://patents.google.com/patent/US8484190B1/en

Ask the one question that separates safe from unsafe situations, point urgent cases to professional help at once, state evidence and its limits, and name who checked the content. Keep the text readable without answering. That is Supervised Search applied where mistakes cost most.