Retrievability¶

A document-centric information-retrieval measure of how broadly and prominently a fixed system exposes an information object across a declared population of queries.

Version: v2 · 2026-08-30 · History
Domain-specific #: 2663
Origin domain: information retrieval
Subdomain: document centric evaluation
Aliases: Document retrievability, Document accessibility in information retrieval

Core Idea¶

Retrievability is a document-centric property induced by an information-retrieval system: it measures how many, or how much probability mass of, a declared population of queries expose a particular information object at ranks a user is willing to inspect. Azzopardi and Vinay introduced it to evaluate the access that a retrieval system affords to individual documents and to the collection as a whole, complementing the usual query-centric measures of relevance effectiveness.^[1]^[2]

For a document $d$, query set $Q$, query weight $o_q$, rank $k_{dq}$ of $d$ for query $q$, rank cutoff $c$, and a rank utility function $f$, the general score is

\[ r(d)=\sum_{q\in Q} o_q f(k_{dq},c). \]

In the simplest binary version,

\[ f(k,c)= \begin{cases} 1,& k\leq c,\\ 0,& k>c\text{ or the document is not returned.} \end{cases} \]

Then $r(d)$ is the weighted count of queries for which $d$ enters the top $c$ results. A graded, top-heavy $f$ can give more credit to rank 1 than rank $c$. The cutoff represents an examination horizon, not an intrinsic document property: a web-search user, a patent examiner, and a systematic searcher may tolerate very different depths.

Retrievability is relational. The same document can be highly retrievable under BM25 with one index and query distribution, poorly retrievable under another model or corpus representation, and absent under a third index. It is therefore incomplete to say “this document has retrievability 37” without naming $Q$, the weights, the retrieval system and index state, $f$, and $c$. The property belongs to the document-system-query protocol, even though it is reported per document.

Structural Signature¶

Recognition form: information object + fixed indexed collection + query population and weights + retrieval/ranking system + rank utility and examination cutoff -> document retrievability score; distribution of scores -> collection-level retrieval-bias profile.

The mandatory roles are:

Information object $d$. A document, document version, dataset, record, or other independently ranked unit whose exposure is being measured.
Indexed collection $D$. The candidate universe and its representation. A missing object necessarily has score zero, but inclusion alone does not make it retrievable.
Query population $Q$. The declared set, sample, or distribution of queries against which access is evaluated.
Query weights $o_q$. Equal weights model an artificial uniform query population; frequency or probability weights model a use distribution.
Fixed retrieval pipeline. Tokenization, indexing, matching function, field weights, length normalization, priors, and ranking determine $k_{dq}$.
Rank utility $f(k,c)$. It converts a ranked position into exposure utility and makes the assumed inspection behavior explicit.
Examination cutoff $c$. Results below the horizon count as unencountered in a binary protocol; it must fit the access task.
Per-object score $r(d)$. The sum or expectation over queries, not the rank under a single query.
Collection distribution. Ordering all $r(d)$ values exposes which objects or groups are systematically surfaced or suppressed.
Protocol audit. Query sampling, document granularity, index coverage, ties, zero-score objects, and uncertainty are reported before comparisons.

The defining invariant is conditional monotonicity: holding $Q$, weights, system, and $f$ fixed, moving $d$ upward for a query cannot lower its score when $f$ is nonincreasing in rank; adding an exposure-producing query of nonnegative weight cannot lower it. Changing the query frame or cutoff changes the measurement itself, not merely its precision.

What It Is Not¶

Retrievability is not relevance. A system can repeatedly surface an irrelevant or low-quality document, giving it high retrievability. Relevance asks whether an object satisfies an information need; retrievability asks whether the system exposes the object across the declared query population. Retrieval is a precursor to relevance judgment, not proof of it.

It is not precision, recall, mean average precision, normalized discounted cumulative gain, or reciprocal rank. Those are normally query-centric effectiveness measures that compare returned items with relevance judgments. Retrievability reverses the axis: it aggregates across queries for each document, often without relevance judgments, then examines the distribution over documents.^[1]

It is not binary indexability, crawlability, storage access, authorization, or file availability. A document can be indexed and permitted yet effectively invisible because almost no plausible query returns it within the examination horizon. Conversely, a high score does not guarantee the user can open, understand, trust, or use the item.

It is not broad findability or HCI discoverability. Those include navigation, signifiers, labels, exploratory interfaces, prior knowledge, and paths outside ranked search. Retrievability fixes a query-and-ranking protocol. Nor is it a rank: rank is query-specific, while retrievability aggregates ranks across a query population.

Finally, unequal retrievability is not automatically unfairness. A relevance-oriented system is designed to favor some documents for some queries. A Gini or Lorenz analysis describes exposure inequality induced by the protocol; a fairness verdict additionally requires protected groups, a justified normative baseline, and relevance or utility controls.

Scope of Application¶

The home scope is evaluation and auditing of ranked information-retrieval systems. It applies to web and enterprise search, digital libraries, web archives, legal and patent retrieval, news archives, scholarly search, and integrated search over publications, datasets, or other record types. The unit need not be a conventional text page, but it must be an object that a stable retrieval pipeline ranks for queries.

The construct is especially useful when relevance judgments are scarce or the question is collection exposure rather than average topical effectiveness. Web archives illustrate this: multiple versions of a page, crawl time, and retrieval-model choices can make some parts of the archive easier to reach than others. Samar and colleagues measured per-version retrievability at several cutoffs and used score distributions, Lorenz curves, and Gini coefficients to compare induced bias.^[3]

Patent prior-art search supplies a high-recall setting. Query-term mismatch can leave many patents at very low or zero retrievability; query expansion and corpus organization can be evaluated partly by whether they expand the query catchments of otherwise hidden patents.^[4]

Modern vector or retrieval-augmented systems can instantiate the abstraction, but only when the evaluation defines the object unit, query population, rank utility, and cutoff. “Used in RAG” by itself establishes no retrievability protocol. Generation quality, grounding, answer faithfulness, and end-to-end task success are downstream questions.

Clarity¶

A complete retrievability result answers: retrievability of which object version, in which collection and index, under which query distribution, retrieval model, weights, rank utility, and cutoff? These are not optional reproducibility details. They define the property.

The cleanest diagnostic contrasts two axes:

Fix a query $q$ and compare documents: this is a ranked result and may support query-centric effectiveness evaluation.
Fix a document $d$ and aggregate over queries: this is document retrievability.

The second axis reveals a system behavior hidden by mean query performance. Two retrieval models can have similar effectiveness on judged topics while exposing the rest of the collection very differently. Conversely, a model with a more equal retrievability distribution need not have the best relevance effectiveness; Wilkie and Azzopardi found a practical relationship between bias, length normalization, and performance, but did not make equality synonymous with optimal ranking.^[5]

Zero must also be interpreted carefully. It can mean that the document was not indexed, no sampled query matched it, matching queries ranked it below $c$, or the query generator failed to express its concepts. Those diagnoses have different remedies.

Manages Complexity¶

A collection with $N$ documents and $M$ queries produces up to $N\times M$ ranks. Retrievability compresses each document's column of ranks into one score, and then compresses the $N$ scores into a distribution that can be compared across systems. This makes otherwise invisible exposure structure tractable: inaccessible tails, overexposed domains, document-length effects, source bias, temporal-version bias, and parameter sensitivity.

The compression supports three levels of analysis:

Object level: identify documents with zero or unusually low $r(d)$.
Group level: compare mean or distributional retrievability by domain, source, date, type, language, or another independently justified grouping.
Collection level: use a Lorenz curve or an inequality statistic such as the Gini coefficient to summarize score concentration.

That hierarchy is useful but lossy. A single Gini value cannot say which documents are hidden, whether exposed documents are relevant, or whether inequality is justified. A single $r(d)$ cannot say which query intents expose the object. The score manages complexity by enabling drill-down, not by replacing it.

Abstract Reasoning¶

The formula licenses precise counterfactuals. Holding the evaluation frame fixed, lowering an object's ranks increases its retrievability under a top-heavy utility. Increasing $c$ weakly increases every binary score, but may change relative order and the apparent inequality of the collection. Upweighting frequent queries makes the score closer to expected exposure for observed users; equal weighting instead asks about breadth across the chosen query vocabulary.

It also localizes interventions. A zero caused by absent indexing calls for crawl or ingestion repair. A zero caused by term mismatch calls for representation, expansion, or semantic matching. Low scores caused by length normalization or field weighting call for ranking calibration. Low scores concentrated in a document group call for a feature and representation audit, but correlation between a feature and $r(d)$ is not proof that the feature caused the bias.

The main inferential hazard is query-frame dependence. Enumerating all possible queries is infeasible, so studies simulate queries from document language, sample query logs, or build topic-centered sets. Each route observes a different access population. Query simulation may privilege verbose documents and the collection's own vocabulary. Logs represent actual use but omit unexpressed needs and inherit the deployed system's feedback. Jahani, Azzopardi, and Sanderson found that simulation-based estimates from a digital-library index were not strong predictors of practical retrievability through an external web search engine, with submitted queries explaining substantial variation.^[6] A retrievability claim must therefore travel with its query frame.

Knowledge Transfer¶

The structure transfers within information access without metaphor. A web page, patent, archived page version, dataset record, and scholarly article each becomes the node; candidate queries become sources; the retrieval model supplies the friction field; rank supplies access cost; $c$ supplies the tolerance horizon; and $r(d)$ measures the weighted query catchment. This mapping is why the specific live parent Access Catchment is more accurate than a generic “measurement” parent.

The method also transfers from object audit to system comparison. Freeze the corpus and query protocol, compute scores under multiple ranking systems, compare score distributions, and then inspect which features or groups move. Only the substrate-specific representation and access task change.

Transfer stops at unranked or non-query discovery unless an equivalent exposure protocol is explicitly constructed. Browsing menus, signage, social recommendation, and serendipitous encounter may produce findability, but calling them retrievability without a query population and rank utility discards the construct's identity. The generic lesson in those settings belongs to access catchment, discoverability, or coverage/reachability.

Examples¶

Canonical¶

Consider four documents $D=\{d_1,d_2,d_3,d_4\}$ and three equally weighted queries $Q=\{q_1,q_2,q_3\}$ with cutoff $c=2$. Suppose $d_1$ has ranks $(1,4,2)$, $d_2$ has $(2,1,\infty)$, $d_3$ has $(3,2,1)$, and $d_4$ is never returned. With binary utility, the scores are

\[ r(d_1)=2,\quad r(d_2)=2,\quad r(d_3)=2,\quad r(d_4)=0. \]

Map the roles: the documents are nodes; queries are candidate sources; the fixed engine produces ranks; rank is access cost; $c=2$ is the inspection horizon; and each score is the size of the document's query catchment. The example shows why a single rank is not retrievability and why index inclusion is insufficient. It also shows that equal nonzero scores can conceal different exposing queries. If $q_1$ carries weight ten and the others weight one, the documents no longer have equal expected exposure even though their unweighted breadth is equal.

Applied / In Practice¶

Web archive search. Samar and colleagues evaluated a Dutch web archive whose pages occurred in multiple temporal versions.^[3] The object was a page version; the collection was the archive index; simulated query sets were the access population; TF-IDF, BM25, and language-model variants supplied alternative ranking fields; and cutoffs from shallow web-like inspection to deeper search supplied tolerance horizons. Per-version $r(d)$ values were aggregated into Lorenz and Gini views. The study showed that retrieval models and version handling induced different exposure distributions and that higher-retrievability documents were easier to find in known-item search.

The example also marks the boundary. A page version that receives high $r(d)$ has broad ranked exposure under that protocol; it is not thereby the most relevant version for a historian, the most authoritative version, or the fairest result. Those judgments require additional evidence.

Structural Tensions¶

Breadth vs. realism of the query frame. A huge synthetic query set covers vocabulary broadly but may poorly represent actual demand; a real log reflects demand but inherits past system exposure and excludes unissued needs. Diagnostic: report both generation mechanism and sensitivity to alternative query frames.
Exposure equality vs. retrieval effectiveness. Equal scores make the collection uniformly reachable, while relevance ranking deliberately concentrates exposure on likely answers. Diagnostic: treat Gini as descriptive and compare it with relevance effectiveness rather than optimizing equality alone.
Shallow encounter vs. deep professional search. A small $c$ approximates ordinary web attention; legal or patent work may inspect far deeper. Diagnostic: set $c$ from the task and show cutoff sensitivity.
Object granularity vs. duplicate amplification. Treating each archived version as a separate document can make version-rich resources dominate; collapsing by URL can erase meaningful temporal change. Diagnostic: declare the identity unit and compare version-aware alternatives.
Relevance-free audit vs. semantic blindness. Retrievability can be computed without judgments, enabling large audits, but it cannot distinguish useful exposure from repeated surfacing of poor material. Diagnostic: pair exposure analysis with judged relevance or downstream utility where conclusions require quality.
Diagnostic correlation vs. causal attribution. Document length, inlinks, or source may correlate with low scores, yet another indexed feature may drive both. Diagnostic: intervene on the ranking feature or run controlled comparisons before claiming cause.

Structural–Framed Character¶

Retrievability is structural with aggregate framed score $0.14$. Once the corpus, query population, weights, retrieval pipeline, rank utility, and cutoff are declared, the score is a formal result. Its recognition does not depend on institutional authority or moral evaluation.

The nonzero framed residue lies in evaluation design. Choosing $Q$ decides whose possible queries count; choosing $c$ encodes an inspection horizon; choosing equal or frequency weights distinguishes lexical breadth from expected exposure. Those choices make every reported value frame-relative. They do not turn the construct into a normative verdict: “unequal” becomes “unfair” only after a separate fairness standard is supplied.

Structural Core vs. Domain Accent¶

The structural core is an access catchment: a node, a population of possible sources, a friction-bearing medium, a tolerance horizon, and the set or weighted mass of sources that reach the node. That skeleton transfers to clinics, transit stops, servers, and other resources.

The domain accent is exact and substantial. Documents and versions are the nodes; text or vector queries are sources; indexing and ranking define the medium; result rank is friction; the examination cutoff is the horizon; query frequency supplies weights; and relevance, precision/recall, query simulation, Lorenz curves, and search-model parameters supply the information-retrieval apparatus. Removing those roles leaves generic access catchment rather than retrievability.

The portable skeleton is already a prime. The residual is therefore correctly domain-specific, not a duplicate prime and not a mere alias for findability.

Retrievability strictly instantiates prime:access_catchment. For a fixed document, queries are the candidate sources that may reach it; the retrieval system is the friction-weighted medium; rank is access cost; the cutoff is the tolerance horizon; and $r(d)$ is the size or weight of the reachable source set.

It is related to prime:search_and_retrieval, which supplies query matching and ranking but does not aggregate exposure per document. prime:sampling_representativeness governs how a sampled $Q$ supports inference to a target query population. prime:coverage_reachability becomes relevant when the design obligation is that every required document have at least one exposing query, but ordinary retrievability does not assert complete coverage. prime:measurement explains why the reported value requires an operational protocol.

domain_specific:discoverability is a close sibling. Discoverability covers content, actions, and features located through cues and bounded exploration; retrievability is narrower, formalizing ranked-query exposure for information objects. Neither term should be used as an unrestricted alias for the other.

Relationships to Other Abstractions¶

Current abstraction Retrievability Domain-specific

Parents (1) — more general patterns this builds on

Retrievability is a kind of Access Catchment Prime

Retrievability strictly instantiates prime:access_catchment.

Hierarchy path (1) — routes to 1 parentless root

Retrievability → Access Catchment → Threshold

Neighborhood in Abstraction Space¶

Retrievability sits in a sparse region of the domain-specific corpus (84^th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Unclustered & Miscellaneous (1565 abstractions)

Nearest neighbors

Extended Boolean model — 0.82
Postings List — 0.81
Retrieval-augmented generation — 0.80
Inverted index — 0.80
Relevance (information retrieval) — 0.80

Computed from structural-signature embeddings · 2026-09-08

Not to Be Confused With¶

Findability: the broad ability to locate an object across search, navigation, metadata, links, and other pathways. Retrievability is one ranked-IR component.
Discoverability: an HCI property of whether actions, features, or content can be located through system cues and exploration. It need not use a query population or rank aggregation.
Searchability: whether content can be indexed and queried. A searchable document can still have near-zero retrievability.
Index coverage or crawlability: whether the object enters the index. Inclusion is a prerequisite, not sufficient exposure.
Relevance: whether an object answers an information need. A highly retrievable object can be irrelevant.
Recall: the fraction of relevant objects retrieved for one query or topic. Retrievability aggregates exposure of one object over queries.
Popularity or click-through rate: observed attention includes presentation, attractiveness, trust, and user behavior after ranking; retrievability may be estimated without clicks.
Retrieval probability: a normalized retrievability score can support a probability interpretation under a query distribution, but the raw cumulative score is not automatically a calibrated probability.
Algorithmic fairness: exposure inequality can inform fairness analysis but does not identify protected groups, deservingness, or a fair baseline by itself.
Accessibility in disability and inclusive design: Azzopardi and Vinay borrowed accessibility language for information access; this is not WCAG conformance or assistive-technology usability.

References¶

[1] Leif Azzopardi and Vishwa Vinay, “Retrievability: An Evaluation Measure for Higher Order Information Access Tasks,” in Proceedings of the 17^th ACM Conference on Information and Knowledge Management (2008), 561–570. https://doi.org/10.1145/1458082.1458157 registry ↩a ↩b

[2] Leif Azzopardi and Vishwa Vinay, “Accessibility in Information Retrieval,” in Advances in Information Retrieval, LNCS 4956 (2008), 482–489. https://doi.org/10.1007/978-3-540-78646-7_46 registry ↩

[3] Thaer Samar, Myriam C. Traub, Jacco van Ossenbruggen, Lynda Hardman, and Arjen P. de Vries, “Quantifying Retrieval Bias in Web Archive Search,” International Journal on Digital Libraries 19, no. 1 (2018): 57–75. https://doi.org/10.1007/s00799-017-0215-9 registry ↩a ↩b

[4] Shariq Bashir and Andreas Rauber, “Improving Retrievability of Patents in Prior-Art Search,” in Advances in Information Retrieval, LNCS 5993 (2010), 457–470. registry ↩

[5] Colin Wilkie and Leif Azzopardi, “Relating Retrievability, Performance and Length,” in Proceedings of SIGIR 2013 (2013), 937–940. https://doi.org/10.1145/2484028.2484145 registry ↩

[6] Hamed Jahani, Leif Azzopardi, and Mark Sanderson, “Measuring the Retrievability of Digital Library Content Using Analytics Data,” Journal of the Association for Information Science and Technology 75, no. 11 (2024): 1233–1248. https://doi.org/10.1002/asi.24886 registry ↩