Supervised Search: the page asks what brought the visitor, and the search reaches its real goal.

Supervised Search is search in which the publisher supervises the last stretch. The page asks where a visitor comes from and what they have already experienced, clarifies the user intent in a short pre-loop, and routes the person to the answer that fits, on the same site or beyond it. A static document becomes a user-triggered dialogue.

Before the page explains, it asks back in the text itself: Have you experienced this before? Is this what brought you here, or was it something else? Each answer opens the passage that fits or points to the page where that need is answered. The question is part of the content, and every reader who does not answer can read on.

Give the visitor guidance on the page itself: keep them where the page fits, or send them on. When the answer to the pre-loop matches what the page covers, the page opens the passage for that situation and the visitor stays. When the answer shows a need that belongs somewhere else, the page names that need and links to the page where it is answered, on the same site or beyond it.

When the person's situation is one the page was written for, even if their words were different. The page then leads straight to the passage for that situation and skips what this person does not need. A visitor who finds their own case on the first screen has no reason to go back to the result page.

When the pre-loop reveals a need that belongs to a related subject, such as a symptom page reached by someone who needs emergency help, or a pricing page reached by someone who wants to cancel. Sending that person on in one step is the helpful answer. A page that tries to hold every visitor leaves the misplaced ones without help.

Pogo-sticking is the pattern of clicking a result, returning to the result page quickly and clicking another one. It shows that the page did not meet the need that brought the person. Supervised Search works on its cause: the page learns early that a visitor needs something else and hands them on before they leave through the back button.

Search systems never see an individual visitor's answer. They read what the page offers: the entry situations it names, the questions it asks back and the routes from each situation to an answer. That is a readable map from search intent to user intent, written by the party that knows the subject.

Google's internal API documentation that became public in 2024 names click signals such as good clicks, bad clicks and last longest clicks. In the United States antitrust case against Google, Pandu Nayak, a Google search executive, described Navboost as a system that uses about 13 months of click data. How much a quick return weighs in ranking has not been made public.

Plausibly, and indirectly. Every reformulated query and every return to the result page is another request the system has to answer, and in February 2023 Alphabet's chairman John Hennessy said an answer from a large language model could cost about ten times as much as a standard keyword search. A page that settles the need in one visit saves those follow-up requests, although the size of the saving has not been measured.

People In Reality Meant, PIRM, is a category introduced here next to People Also Ask. People Also Ask shows further questions that searchers type around a query, and PIRM records what people turned out to mean when a page asked them. It is the user intent behind the search intent, collected where the person arrives.

People Also Ask is built by the search engine from queries and shows what else is typed. PIRM is built by the publisher from answers to the pre-loop and shows what people meant when they typed. People Also Ask widens the search intent, and PIRM brings it closer to the user intent.

It lists the real situations behind a query as short entries in the visitor's voice: I searched for this, and what I needed was that. Each entry leads to the passage or page that answers it. Over time the list shows which meanings the result page never offered.

On this site, the section that defines Supervised Search sits inside an element named supervised-search, a valid custom HTML element. It makes the concept addressable in the page structure. A formal vocabulary for it would be a matter for schema.org.

The question sits in the content, and no dialog covers the page, since Google states that intrusive interstitials and dialogs can lead to poor search performance. The content stays readable for everyone who does not answer. Answers that are stored are personal data and need consent.

In the United States, supervised search refers to searches of people on probation or supervised release, and the result pages for the phrase are dominated by offender lookups and court documents. In information retrieval, it refers to supervised machine learning for ranking and diversifying results. Supervised Search as a publisher practice is a third meaning, and it needs to be named as such.

Search result diversification selects results that cover several possible needs behind an ambiguous or broad query. It accepts that one query can stand for many intents. A diversified list tries to give each intent a result near the top.

If a query meant the same to everyone who typed it, one list could serve six people as well as a list made for each of them. The best shared list reaches 0.85, the list Live Search showed 0.58.

Supervised methods learn a ranking function from labelled examples, such as human relevance judgements or subtopic labels. They generally outperform hand-tuned formulas. They also inherit whatever the labels get wrong.

Learning to rank applies machine learning to training data to order documents for a query. It replaced hand-set scoring in large search systems. The system learns what past judgements rewarded.

Maximal Marginal Relevance, introduced by Jaime Carbonell and Jade Goldstein in 1998, selects each next result by balancing its relevance against its novelty compared with results already chosen. It is the classic implicit diversification principle. It avoids ten results that say the same thing.

Explicit result diversification models the subtopics or intents of a query directly and picks documents to cover them, as xQuAD does. The intents are named. Coverage becomes measurable.

Subtopic attention is a neural method that focuses on the subtopics already covered by selected documents when choosing the next one. It lets a model track what the list still lacks. It is one of the supervised approaches to diversification.

Recurrent neural networks process the sequence of documents selected so far and carry information from step to step. In diversification they track coverage as the list grows. Transformer models have since taken over many of these tasks.

Distributed representations encode queries and documents as dense vectors so that similarity of meaning can be computed. They reach beyond exact word matches. Meaning is still learned from the words people wrote.

Query intent is the need behind a query, which can be modelled as a flat list or a hierarchy of subtopics. Systems estimate it from the query and from past behaviour. The query alone reveals only part of it.

If the query string carried the user intent, it would tell how differently people want results for that query. On its own it sorts 39.9 % of queries correctly. With the clicks of earlier searchers added, 81.3 %.

SVMDIV, introduced by Yisong Yue and Thorsten Joachims in 2008, is a supervised diversification method that learns to maximise the coverage of subtopics, using word coverage as a proxy. It was an early baseline for learned diversification. It shows how coverage became a training objective.

Alpha-NDCG, introduced by Charles Clarke and colleagues in 2008, rewards results that add new aspects of a query and penalises redundancy. It is the standard measure of diversification. It measures variety in a list, which says nothing yet about the person in front of it.

Intent coverage is the share of possible intents behind a query that the top results serve. High coverage spreads the risk of missing someone's need. Each individual person still receives mostly results meant for others.

Particle swarm optimisation is a heuristic method that searches for good parameter values by moving candidate solutions through a search space. It has been used to tune ranking parameters. It optimises the system, and leaves the question of what one visitor wanted untouched.

Query reformulation is the rewriting of a query, by a person or by a system, to express the need differently. Systems generate reformulations to discover subtopics. People reformulate when the first results missed their need.

Information foraging theory, developed by Peter Pirolli and Stuart Card, describes people searching for information like animals foraging for food, following cues and leaving a patch when returns fall. It explains when people stay and when they move on. A pre-loop changes the cues at the moment of arrival.

Information scent is the perceived value of a cue, such as a link text or a snippet, for reaching the needed information. Strong scent draws people on, weak scent makes them leave. A question that names the visitor's situation is strong scent.

Visual search is the scanning of a display for a target among distractors, guided by attention and by the similarity of items. Every page is scanned before it is read. A pre-loop placed where the eye lands first is found.

Behavioural supervision provides structured guidance and practice while someone carries out a task, and it has been used, for example, to support people looking for work. Supervision in this sense is guidance. That is the sense in which a page can supervise a search.

Exploratory search is searching to learn or investigate, as distinct from looking up a known fact. It unfolds over many steps and changes as the person learns. It benefits most from guidance.

The person who is searching, and the moment of arrival. The usual treatment in information retrieval optimises ranked lists from labelled data, and the usual treatment in psychology describes how people forage and scan. Both leave out interactive clarification before an answer, the biases in the click data that trains rankings, and the executive and social processes by which one person guides another's search. Supervised Search as a publisher practice starts exactly there.

Intent disambiguation resolves which of several possible needs a query expresses before results are ranked, for example by asking back. Diversification spreads bets across intents; disambiguation asks. Supervised Search does the asking on the page.

Query clarification is a system-initiated question to the user when the need is unclear, such as asking which of several meanings is intended. Search systems use it sparingly. On a page it becomes a pre-loop.

Subtopic mining discovers the subtopics of a query from query logs, documents or suggestions before any ranking. Evaluation campaigns such as the NTCIR intent tasks made it a research problem. Knowing the subtopics is the precondition for asking the right question back.

ERR-IA, the intent-aware version of expected reciprocal rank, weights each intent by its probability and models a user who stops at the first satisfying result. It measures diversification from the side of a scanning person. It brings user behaviour into the metric.

Expected reciprocal rank, introduced by Olivier Chapelle and colleagues in 2009, estimates how soon a user is likely to find a satisfying result, assuming they scan from the top and stop when satisfied. It rewards the first good answer. It fits people better than measures that treat every position alike.

Subtopic recall is the share of all subtopics of a query that the retrieved list covers. It is the plainest measure of coverage. It shows breadth, and depth for any one intent stays unmeasured.

Pairwise learning to rank trains a model on pairs of documents, learning which of the two should rank higher. RankNet is an early example. It learns preferences between results.

Listwise learning to rank optimises a loss over the entire ranked list at once. It targets what users see, a whole list. It is closer to the evaluation measures than pairwise training.

Pointwise ranking scores each document independently and sorts by the score. It is simple and fast. It ignores how documents in a list relate to each other.

LambdaMART, developed by Christopher Burges and colleagues at Microsoft, combines gradient-boosted decision trees with a ranking objective. It won major learning-to-rank competitions and became a standard in industry. It remains a strong baseline.

The relevance-diversity trade-off is the balance between showing the most relevant results and covering many intents. More diversity can mean less relevance for each person. Asking the person resolves the trade-off for that person.

Intent-aware ranking scales document scores by the estimated probability of each intent. It turns intent estimates into ranking decisions. The estimates come from populations, and one visitor may be the exception.

Counterfactual learning to rank trains ranking models from logged clicks while correcting for the bias in how results were shown, as Thorsten Joachims, Adith Swaminathan and Tobias Schnabel showed in 2017. Clicks reflect position as well as relevance. Without correction, a system learns its own past rankings.

Inverse propensity scoring weights each observed click by the inverse of the probability that the result was seen. It corrects for results that were rarely examined. It is the standard tool of counterfactual learning to rank.

Position bias correction removes the effect that higher-ranked results receive more clicks regardless of their value. Eye-tracking studies showed that people click the first result even when the second is better. Uncorrected, the bias confirms whatever already ranks.

The cascade model, formalised by Nick Craswell and colleagues in 2008, assumes that users scan results from top to bottom and click the first one that satisfies them. It explains position bias. It is the basis of many click models.

A click-through log records queries, shown results and clicks. It is the raw material of learned ranking. It records what people did with what they were shown, and never what they would have done with something else.

Off-policy evaluation estimates how a new ranking policy would perform using data collected under an old one. It avoids exposing users to untested systems. It depends on how much the old policy explored.

The exploration-exploitation trade-off is the tension between showing what is known to work and trying alternatives to learn. Systems that only exploit never discover unmet intents. A page that asks explores directly.

A contextual bandit algorithm chooses an action for each situation and learns from the reward it receives, balancing exploration and exploitation online. Lihong Li and colleagues applied it to personalised news recommendation in 2010. It learns from live behaviour.

Dense retrieval finds documents by comparing learned vector representations of queries and passages, as dense passage retrieval showed in 2020. It works in the first stage of retrieval. It retrieves by meaning learned from words.

A cross-encoder reads a query and a document together through one transformer model to score their relevance. It is accurate and expensive, so it is used to re-rank short candidate lists. It captures fine interactions between query and text.

A transformer ranking model uses self-attention over queries and documents to judge relevance, as BERT-based rankers do. It set new accuracy standards in retrieval. It still learns from the judgements and clicks it is given.

Inter-annotator agreement measures how consistently different human judges label the same items. In relevance judgement it is often far from complete. Supervised systems learn from labels whose disagreement is rarely reported.

Label noise is error in training labels, from human mistakes or from treating clicks as relevance. Models can overfit to it. The supervision itself can be wrong.

A negative sampling strategy decides which non-relevant or unclicked documents a model sees as negative examples. The choice shapes what the model learns to reject. Unclicked is not the same as irrelevant.

Query intent drift is the change over time in what people mean by the same query. A model trained on last year's behaviour serves last year's intent. Asking the visitor captures today's.

Session-based recommendation predicts what a person needs next from the sequence of actions in the current session, as models such as GRU4Rec do. It treats a search as a path. The intent becomes visible in the sequence.

User session history is the record of queries and interactions within a person's session or over time. It carries context a single query lacks. It is also personal data and needs care.

Click-through rate prediction estimates the probability that a user clicks a result or an ad. It drives commercial ranking and ad placement. It optimises for clicks, which are only a proxy for having helped.

Search intent classification assigns queries to classes such as informational, navigational or transactional. It structures intent for systems. It works with categories, and the person is larger than the category.

Data snooping bias arises when models are tuned repeatedly on the same benchmark until the gains fit the benchmark and fail in real use. Halbert White described the problem for statistical testing in 2000. Benchmark success can mislead about real search.

Cognitive search behaviour is how people actually perceive, judge and decide while searching. Ranking models built on statistical patterns rarely model it. Supervised Search builds on it.

Epistemic diversity is the variety of perspectives and kinds of knowledge represented in results, beyond the variety of subtopics. A list can cover every subtopic from a single viewpoint. It concerns what people get to know.

Epistemic pluralism is the view that several ways of knowing and several reasonable positions deserve representation. In search it raises the question of balance in public debate. It adds a normative dimension to diversification.

The supervisory attentional system, proposed by Donald Norman and Tim Shallice in 1986, is the control system that intervenes when routine behaviour is not enough. It interrupts automatic action to plan and decide. A pre-loop does the same for a search: it interrupts the automatic click and lets the person choose.

Executive control is the set of processes that plan, direct and regulate behaviour, associated with the prefrontal cortex. Searching a complex subject draws on it heavily. When it is overloaded, people give up and it looks like lack of interest.

Inhibitory control is the ability to suppress irrelevant responses and distractions. On a crowded page, ignoring what does not matter takes effort. A page that removes distraction spares that effort.

Metacognitive monitoring is the ongoing judgement of one's own understanding and confidence, described by Thomas Nelson and Louis Narens. Searchers stop when they feel they know enough, whether they do or not. A question back can recalibrate that feeling.

Co-regulation is the regulation of attention, emotion or action between two people, as when a helper and a learner adjust to each other. Search is usually treated as a solitary task. Supervised Search makes it a shared one.

Joint attention is two people attending to the same thing and knowing that they do. In guided search, the guide and the searcher look at the same options together. A pre-loop creates a written form of it.

Scaffolding, described by David Wood, Jerome Bruner and Gail Ross in 1976, is support that lets a learner do what they could not do alone, withdrawn as competence grows. Good guidance fades. A pre-loop helps at the start and gets out of the way.

The zone of proximal development, a concept from Lev Vygotsky, is the range between what a person can do alone and what they can do with guidance. Learning happens in that range. Supervised Search serves visitors exactly there.

A feedback intervention gives people information about their performance to change it. Avraham Kluger and Angelo DeNisi found in a 1996 meta-analysis that such interventions improved performance on average and made it worse in about a third of cases. How and when feedback is given decides its effect.

External regulation is behaviour steered by outside demands or guidance, in self-determination theory the least autonomous form of motivation. Guidance that is felt as control can backfire. A pre-loop works when it feels like help and fails when it feels like a gate.

Vicarious learning is learning by observing someone else perform a task. Watching an expert search teaches strategies that trial and error teaches slowly. A page can show how an expert would proceed.

Signal detection theory separates the ability to detect a target from the willingness to say it is there. Clicks mix both. A searcher under guidance may change the decision criterion without seeing better.

Guided visual search is Jeremy Wolfe's model in which attention is guided by a priority map built from salience and goals. What the person is looking for shapes what they see. A clarified intent sharpens that guidance.

Target detection rate is the share of targets a searcher actually finds. A click does not prove the target was found. Measuring detection separates finding from clicking.

Search efficiency describes how search time grows with the number of items, the search slope. Cluttered result pages raise it. A page that narrows the options first makes every later search faster.

Distractor interference is the slowing of search by irrelevant items that compete for attention. Diversified lists and busy pages add distractors for any single person. What is relevant to someone else distracts this visitor.

Bottom-up salience is the power of physical features such as contrast, size and motion to attract attention regardless of goals. It can pull the eye to what does not matter. Salience is not relevance.

Cognitive load is the demand a task places on working memory. Strong scent that overloads memory still fails. Supervised Search reduces load by narrowing the path.

Working memory capacity is the limited amount of information a person can hold and use at once, and it varies between people. Multi-step searches strain it. Differences in success are often differences in capacity.

Fixation duration is how long the eyes rest on one point. Longer fixations indicate harder processing or hesitation. They show where a page makes people stop and think.

Saccade amplitude is the distance the eyes jump between fixations. It reveals how efficiently people scan a layout. Long jumps across a page often mean searching for something that should have been obvious.

Dual-task performance is how well people do two things at once. Real searches happen while doing other things. Guidance that asks little helps most under divided attention.

The task-switching cost is the loss of time and accuracy when switching between tasks, reviewed by Stephen Monsell in 2003. Jumping between subtopics carries this cost. A clear route through a subject avoids needless switches.

The attentional blink, described by Jane Raymond, Kimron Shapiro and Karen Arnell in 1992, is the brief period after detecting one target when a second one is often missed. Rapidly presented information hides second targets. Not everything shown in quick succession is seen.

The vigilance decrement is the decline in detection over long periods of monitoring, first studied by Norman Mackworth with radar operators. Attention fades with time on task. Long searches lose accuracy toward the end.

Pupil dilation increases with mental effort and arousal and can be measured with eye trackers. It reveals difficulty people do not report. It makes the effort of searching visible.

Mental workload is the subjective and objective demand of a task, often measured with the NASA Task Load Index. Two interfaces with the same completion time can demand very different effort. Workload shows what time hides.

The Wisconsin Card Sorting Test measures cognitive flexibility, the ability to shift rules when feedback changes. Perseverating on a failing strategy shows in search as repeating the same query. Flexibility decides whether people find another way.

Cognitive ageing brings slower processing and changes in attention and memory, as Timothy Salthouse has documented. Older searchers need clearer routes and less clutter. Guidance helps them more.

Attention deficit hyperactivity disorder involves difficulties with sustained attention and inhibitory control. Distracting pages and long paths cost people with ADHD disproportionately. Clear guidance makes search more accessible.

Parent-child interaction in searching together is a natural form of supervised search, with the adult guiding and the child learning. It combines joint attention, scaffolding and feedback. It is the everyday model of what a page can do.

Operant conditioning, described by B. F. Skinner, shapes behaviour through rewards and consequences. Search interfaces condition users by what they reward with quick answers. Pages teach people how to search them.

Rankings, diversification and answer engines all serve a population of intents. The individual visitor still arrives with one need, often one that no list anticipated. The page that asks and routes that visitor wins the person that the average missed.

Name the situations people arrive from, ask one short question back in the content, keep the visitor where the page fits and send them on in one step where it does not, and keep everything readable without answering. Record what people in reality meant and show it as PIRM. That is the practice: the publisher supervises the last stretch of the search.