Glossary

Position Bias

Web search behavior research; Guo et al., 2009; Pan et al., 2007

A finding from web search behavior research, established across eye-tracking and click-log studies rather than introduced by a single paper. The tendency of users to click higher-ranked search results regardless of their actual quality. Ranking position itself influences click behavior, causing higher-ranked results to receive disproportionate attention independent of content merit. It is the rank-location member of the broader Presentation Bias family; correcting for it does not correct for the others.

Presentation Bias

Yue et al., 2010; Bar-Ilan et al., 2009; US Patent 8,938,463

A term from information retrieval research on clickthrough data. The family of effects in which how a result is presented — its rank position, its title and snippet, its visual prominence, the source it appears to come from — influences attention, clicks, and quality judgments independently of the content behind it. Position Bias is the rank-location member of this family; presentation bias is the broader set, which also covers result attractiveness and other display effects. The distinction matters for measurement rather than vocabulary: a click model that corrects for rank alone leaves the other members uncorrected, so residual presentation effects survive the correction and are easily mistaken for quality. Documented in retrieval research and addressed directly in Google's US Patent 8,938,463, which describes a rank modifier engine correcting for presentation bias before behavioral signals are allowed to influence rankings.

Site Reputation Bias

Information science; Bar-Ilan et al., 2009

A finding from information-science research on how people choose among search results. The tendency of users to favor results from well-known or trusted sources regardless of their position in the results. Observed alongside presentation bias, site reputation bias shows that prior brand associations can override the positional heuristic — well-known sites may be favored even when ranked mid-page. Not to be confused with "site reputation abuse," an unrelated search-engine policy term concerning third-party content published on an established domain; this entry describes a documented user-selection effect.

Anchoring Effect

Tversky & Kahneman, 1974

A cognitive bias identified by Amos Tversky and Daniel Kahneman in 1974, in the judgment-under-uncertainty research program that founded behavioral economics. The first piece of information encountered sets a reference point for all subsequent judgments. In search, early results may establish expectations that influence how users evaluate everything that follows.

System 1 / System 2

Terminology: Stanovich & West, 2000. Popularized: Kahneman, 2011.

A distinction from dual-process theory in cognitive psychology. The System 1 / System 2 terminology was introduced by Keith Stanovich and Richard West in 2000 and later popularized by Daniel Kahneman in Thinking, Fast and Slow (2011) — Kahneman made it famous; he did not name it. System 1 is fast, automatic, and emotional. System 2 is slow, deliberate, and logical. Most search behavior runs on System 1 — users scan and click before engaging critical evaluation.

Cognitive Miser

Fiske & Taylor, 1984

A concept from social psychology, introduced by Susan Fiske and Shelley Taylor in 1984. It describes the human tendency to minimize mental effort by relying on shortcuts and heuristics rather than extensive analysis. In search contexts, users default to fast judgments rather than systematically evaluating all available results.

Latent Construct

Psychometrics and measurement theory (latent variable modeling)

A term from psychometrics and measurement theory, applied here to search. A theoretical variable that cannot be observed or measured directly, and must be inferred from observable indicators. In search research, user satisfaction is a latent construct: no single behavioral measurement is equivalent to it. Retrieval systems therefore approximate it from several indicators at once — which result was selected, what happened after the click, how long attention was held, whether the query was reformulated, whether the person came back, and whether the task was completed. No individual indicator should be read as a direct measure of satisfaction, because each is ambiguous alone: a long visit can mean absorption or confusion, and a short one can mean failure or a question cleanly answered. Which indicators are available depends on who is looking. A retrieval system observes the whole session, including reformulations and the queries that follow. A site owner does not — reformulation happens on the results page, outside their analytics, and can only be inferred indirectly, for instance from a page that earns impressions and clicks while logging consistently short sessions. The framework treats that asymmetry as part of the definition: a proxy you cannot observe is not a proxy you can act on. It is also why 'Google rewards user satisfaction' describes an outcome while providing no operational guidance about mechanism.

Revealed Preference

Samuelson, 1938

A principle from economics, introduced by Paul Samuelson in 1938 as a foundation for consumer choice theory. Preferences are better inferred from observed actions than stated opinions. In search, behavioral patterns such as return visits, query reformulations, and dwell time may indicate a different picture of satisfaction from the one users would report directly.

Post-Hoc Rationalization

Nisbett & Wilson, 1977

A finding from social and cognitive psychology, demonstrated most influentially by Richard Nisbett and Timothy Wilson in 1977, whose subjects confidently explained choices they could not have introspected. The tendency to construct explanations after decisions are made, often making intuitive choices appear more deliberate than they originally were. In search, users who click a result based on position may subsequently justify that choice as quality-based.

Measurement Gap

Finding: Parry et al., 2021; Scharkow, 2016. Label: Behavioral Responsivity Framework.

A finding from communication and digital-media research. The name used here is the framework's shorthand; the result is not. It names a discrepancy between self-reported behavior and a corresponding behavioral record collected over the same population and observation window. Evidence from digital-media research indicates that self-reported duration and frequency are only moderately associated with device-logged use, and that the two rarely agree on the actual amount. A moderate correlation with poor agreement is the characteristic pattern: people who report more use do tend to log more use, but they do not report how much. This does not mean self-report is an inappropriate instrument for subjective experience.

Explicit Signals

Term: information retrieval (explicit vs implicit feedback). Limits: Nisbett & Wilson, 1977; Scharkow, 2016.

A term from information retrieval, where feedback is conventionally divided into explicit and implicit. Explicit signals are the ones users generate intentionally — star ratings, surveys, thumbs up, or direct responses to prompts. Explicit signals are easy to collect but structurally unreliable: they are subject to social desirability bias, post-hoc rationalization, and the introspection limits documented by Nisbett and Wilson (1977). Users tend to report the identity they aspire to, not the behavior they perform.

Implicit Signals

Information retrieval (implicit feedback); Joachims et al., 2005, 2007

A term from information retrieval, where interaction data stands in for explicit relevance judgments; the explicit/implicit division predates the web-search work cited here. Implicit signals are generated as a byproduct of user behavior rather than as user testimony — a click, a pause, a return to search, a query reformulation. Because they require no self-report, implicit signals bypass the cognitive distortions that make explicit signals unreliable. They still require correction for known biases such as position bias before they carry reliable inferential weight.

Information Foraging

Pirolli & Card, 1999 (Psychological Review); adapted from optimal foraging theory in behavioral ecology

A theory from human-computer interaction, developed by Peter Pirolli and Stuart Card at Xerox PARC in the 1990s, adapting optimal foraging theory from behavioral ecology. It models human information-seeking as analogous to foraging for food. Users allocate attention toward sources where the expected rate of gain is high and abandon — leave the 'patch' — when yield drops below a threshold. In search, query reformulation is patch abandonment: the user evaluated what was served, found insufficient yield, and moved to a new search environment.

Dwell Time

Behavioral IR / web analytics

A measure from information retrieval and web analytics, with no single originating paper. The amount of time users spend on a page after arriving from search results. Longer dwell time can indicate engagement or utility, though interpretation depends heavily on context — a login page with high dwell time signals confusion, not satisfaction. Duration alone is not self-interpreting: it carries meaning only once normalized against the intent behind the query and the type of page serving it.

Goodhart's Law

Goodhart, 1975; the widely quoted phrasing is Strathern, 1997

A law from economics, stated by Charles Goodhart in a 1975 paper on UK monetary policy: any observed statistical regularity will tend to collapse once pressure is placed upon it for control purposes. The compressed version usually quoted — when a measure becomes a target, it ceases to be a good measure — is Marilyn Strathern's 1997 generalization of it, not Goodhart's own words; both are given here because the second is the one people recognize and the first is the one he wrote. The law is the reason a behavioral signal's usefulness to a retrieval system and its usefulness as an optimization target move in opposite directions: the more openly a signal is rewarded, the more effort is spent producing the signal rather than the thing it was standing in for. It is the general principle the Manipulation-Cost Framework applies to a specific case, and the reason this framework treats any single-metric target as structurally unsafe.

Implicit Behavioral Validation

Method: Joachims et al., 2005. Design evidence: US Patent 8,938,463 — which documents an intended design, not a confirmed deployment.

The framework's procedure for treating observed behavior as evidence about content usefulness without asking the user anything. The underlying method is implicit feedback — a long-established area of information retrieval in which interaction data stands in for explicit relevance judgments — and this entry does not claim it. What the framework adds is a reportability condition: a behavioral reading is not admissible unless its tier is named under the Behavioral Signal Hierarchy and it satisfies the Signal Interpretation Constraints, and a reading that fails either is discarded rather than discounted. Two things it deliberately does not assert. It does not establish that any particular engine uses any particular signal, or how heavily. And it does not treat implicit signals as more truthful than explicit ones — only as failing differently: explicit measures carry introspection and social-desirability limits, implicit measures carry presentation and context confounds, and neither is the criterion the other is scored against.

Contextual Normalization

Method: Chapelle & Zhang, 2009 — contextual click modeling. Procedure and reportability rule: Behavioral Responsivity Framework.

The process of adjusting behavioral signals against the context in which they were generated before drawing conclusions about content quality. The same signal carries different meaning in different environments: long dwell on a recipe page indicates engagement; long dwell on a checkout form indicates confusion. Reliable interpretation requires normalization for query intent, page type, device, and temporal patterns. This entry is the procedure; the conditions that make it necessary are stated under Signal Interpretation Constraints. Normalizing a measurement against the conditions that produced it is ordinary practice in behavioral measurement and in click modeling, and the framework claims none of it. What it specifies is narrower: which contextual factors have to be declared before a reading is reportable, and that an undeclared one invalidates the reading rather than merely weakening it.

Satisfaction Paradox

Structure adapted from the satisfaction paradox in quality-of-life research (Herschbach, 2004); proposed mechanisms informed by Guo et al., 2009 and Pan et al., 2007

A proposed search-specific subtype of Utility Divergence: a user reports satisfaction with a search episode while the observed behavior is inconsistent with a behavioral-success hypothesis stated in advance for resolving the underlying task. What separates it from the umbrella term is that it carries a proposed explanation — that position bias, cognitive ease, and post-hoc rationalization can hold a favorable judgment in place even where the need went unresolved. Those mechanisms are proposed, not demonstrated. Each is documented in its own literature; none of that establishes that any particular reported satisfaction was produced by any of them. The structure is adapted from the satisfaction paradox in quality-of-life research, where subjective assessment can diverge from independent indicators of a person's circumstances — the adaptation to search evaluation is ours, and that literature did not establish a search effect. Behavioral evidence does not directly measure utility, relevance, satisfaction, or task resolution, so what the term names is a disagreement between two fallible channels rather than between a report and the truth. It is unrelated to satisfaction of search, an established effect in radiology and visual search in which finding one target reduces detection of a second: that term concerns multi-target detection and premature search termination, this one concerns self-report read against behavioral evidence, and they share no mechanism.

Utility Divergence

Behavioral Responsivity Framework

The framework's umbrella category for a disagreement between what a person reports finding useful and what their behavior indicates when read against a behavioral-success hypothesis stated in advance. It is deliberately mechanism-neutral: it names the disagreement without asserting a cause, without quantifying its size, without establishing that the task failed, and without deciding which channel gives the truer account. Satisfaction Paradox is its search-specific subtype and does carry proposed mechanisms — every Satisfaction Paradox case is a Utility Divergence case, and not every Utility Divergence case is a Satisfaction Paradox. It is classified as a phenomenon rather than a metric, and it is one the framework cannot currently detect on its own: the reported side needs a survey, interview, or experiment this framework does not run, so what it can identify is the behavioral half alone, held for a later comparison. That the two channels come apart at all is established elsewhere — see Measurement Gap — not by this framework. Not to be confused with "utility divergence" as used in evidence retrieval (Zhang et al., 2023), where the term names a loss component measuring the gap between a verifier's judgments on gold versus retrieved evidence; the two senses share a name and nothing else.

Behavioral Success Hypothesis

Behavioral Responsivity Framework

An explicit, stated-in-advance rule identifying the observable action patterns expected to accompany a specified task outcome, for a defined class of sessions. It is a hypothesis, not a measurement: it says what behavior should look like if the information need were met, and it does not become an estimate of success until it is evaluated against an independent reference criterion. A hypothesis is reportable only when it states all six of: the outcome it concerns, the behavioral features used and their definitions, the weighting or convergence rule applied, the observation window, the contextual factors held constant or normalized for, and the alternative explanations that were not ruled out. Where no independent criterion is available, that absence is part of the report — the hypothesis stands untested and must not be described as a measure of satisfaction, relevance, or utility.

Intent-Response Alignment

The correspondence between the goal behind a query and what a page provides in response, as distinct from how well the page matches the query's keywords. It is treated as a qualitative property and deliberately not scored: a single alignment figure would require a pre-specified success outcome for every query class, and those outcomes are context-dependent enough that the number would carry more precision than the definition supports. Where the practical question is whether one page serves an intent better than another, the framework's answer is a controlled comparison at adequate sample size, not an index. Poor correspondence produces short clicks and query reformulations regardless of keyword match.

Behavioral Signal Hierarchy

Behavioral Responsivity Framework; informed by Joachims et al., 2005, 2007; Pirolli & Card, 1999

The structure the framework uses to classify behavioral signals and order them by how far each can be trusted. It separates explicit from implicit signals, then sorts the implicit ones into four tiers — Click, Engagement, Reformulation, and Longitudinal — by inferential reliability and manipulation resistance, not by confirmed ranking weight. The ordering reflects epistemology: higher tiers are harder to manufacture at scale, so they carry more inferential weight, while the ranking weight engines actually assign each tier remains publicly unconfirmed. Its working purpose is to block the common error of treating every behavioral signal as an equivalent input. The categories themselves are not a discovery — click, engagement, reformulation and longitudinal signals are familiar groupings in retrieval research; what the framework claims is the ordering, the reliability rule that produces it, and the refusal to collapse it into a score. This entry states the ordering; the argument for why the ordering should hold is the Manipulation-Cost Framework.

Manipulation-Cost Framework

Behavioral Responsivity Framework

An ordinal ordering of behavioral signals by how costly they are to manufacture at scale under a stated threat model, on the reasoning that a signal which is harder to fake carries more inferential weight. This is an ordering, not a function: the framework claims the rank order, not a measured relationship between cost and reliability, and no units are attached to either. Click farms can generate Tier 1 signals; manufacturing coherent longitudinal patterns requires also generating the underlying utility that motivates them. A scored instrument — rating each signal on production cost, scalability, detection risk, and dependence on genuine task completion — is a stated design goal rather than a current capability; specifying it requires signal-level research this framework has not yet carried out. It stands in a fixed relation to the Behavioral Signal Hierarchy: the Hierarchy is the ordering, this entry is the criterion that generates it.

Signal Interpretation Constraints

Behavioral Responsivity Framework

The conditions that have to hold before a behavioral signal can be read as evidence of anything. Three are load-bearing. Meaning is not a property of the signal: the same measurement carries a different reading depending on query type, page type, device, and the intent behind the visit, so a value cannot be interpreted without the context that produced it. Signals are therefore not additive — summing or averaging a measure across contexts discards the very information that made each instance interpretable, and produces a number that describes no situation. And interpretation is comparative rather than absolute: a dwell time is long or short only against comparable pages under comparable conditions, so a reading with no stated comparison class is not a reading. These constraints state what makes an interpretation invalid; Contextual Normalization is the procedure applied in response to them.

Behavioral Information Organization

Behavioral Responsivity Framework

The framework layer that translates behavioral requirements into decisions about how information is structured, sequenced, and connected. Where the Behavioral Foundations layer explains why people seek information, and Signal Architecture describes what retrieval systems can observe of that seeking, this layer asks a different question: given a documented behavioral requirement, what information must exist, how should it relate to other information, and in what order should a person meet it. It governs the Representation stage of the Retrieval Loop — the stage at which a behavioral model becomes an actual information system.

Decision Knowledge Flow

Behavioral Responsivity Framework; Field Report 001

The sequence in which a page removes a reader's uncertainty — the concepts and evidence it presents, ordered by what the next step of the decision requires. A page is modeled as a decision journey rather than as a document, on the premise that the same content in a different sequence resolves the decision differently. It pairs with the page's knowledge graph and answers a different question: the graph defines what is connected; the Decision Knowledge Flow defines when each connection becomes relevant.

Evidence-First

Behavioral Responsivity Framework; Field Report 001

A page satisfies Evidence-First when its densest verifiable technical content — the specification values, rated capacities, compatibility constraints, and named standards a qualified reader would use to assess or disqualify the offering — is reachable in the initial viewport under at least one declared desktop or mobile viewport configuration, without interaction. Stated this way, it is a binary structural property of the page, checkable by inspection. It is deliberately not a claim about outcomes: the framework does not assert that this placement improves ranking, conversion, satisfaction, comprehension, or decision quality. Any such effect requires a controlled comparison. What counts as the densest verifiable content must be declared before inspection and evaluated in relation to the page's task and audience. Where a page contains no such content, the property does not apply. Related concepts: Anchoring Effect; Cognitive Miser. These may motivate hypotheses about evidence placement but are not requirements of Evidence-First.

Intra-Page Ontology

Behavioral Responsivity Framework; Field Report 001

The model of the concepts, entities, claims, and evidence relationships expressed within a single page: what the page covers, how its concepts connect, what each claim asserts, and which verifiable evidence supports, qualifies, or limits that claim. It is distinguished from a site-level ontology, which models relationships among pages, topics, and site entities. The framework models both scales deliberately. Conventional site architecture often prioritizes the relationships between pages while leaving the semantic and evidential relationships within a page implicit, editorial, or unformalized. Schema markup, internal links, headings, page layout, tables, citations, and component structure are representations of the intra-page ontology; they are not the ontology itself.

Retrieval Coverage

Behavioral Responsivity Framework; Field Report 001

The proportion of a defined query set for which a domain's own pages are cited as evidence in a generated answer. Rankings measure where a page appears in a list of links; Retrieval Coverage measures whether the page is built into the answer itself. It is measured under a fixed protocol — the same queries, repeated runs, every miss recorded — so results stay comparable across time and across clients. It is a retrieval outcome, not a traffic outcome, and it does not replace rankings: it measures an earlier stage of visibility that rankings do not describe.