What does the evidence establish about visibility in generative search?

Optimisation experiments, source attribution and the distance between citations and commercial outcomes

This retrospective review considers source versions available by 6 August 2026. The cutoff defines the evidence included; preparation and source checking took place on 21 October 2026.

Keywords: generative search; GEO; retrieval; citations; visibility measurement

1. Visibility is a family of outcomes

A webpage can be retrieved by a search system without appearing in its generated answer. It can appear as a citation without being clicked. A brand can be mentioned without receiving a recommendation, and a recommendation can occur without a link to the brand's own website. These are different forms of visibility. Treating them as one outcome makes an optimisation claim difficult to interpret and can disguise the distance between an observed citation and a business result.

This review asks what the available research establishes about improving visibility in generative search, and which conclusions remain outside that evidence. It combines a primary optimisation study with research on retrieval, attribution and answer quality, alongside official platform documentation. The aim is to distinguish a measured effect in a defined engine from general claims about all AI search systems. The latter differ in retrieval sources, query expansion, model behaviour and presentation.

2. Method and evidence selection

Targeted retrieval on 21 October 2026 covered generative engine optimisation, citation evaluation, retrieval-augmented generation and official search documentation. Thirteen references were retained. The accessible GEO manuscript was inspected for its experimental setup, visibility metrics, results and limitations. Official Google and Bing pages were consulted directly. Adjacent model studies were examined through author abstracts and publication records. No live search-visibility audit or intervention experiment was performed for this review.

The synthesis separates experimental findings from platform statements. Research describes a system and comparison under observed conditions. Documentation describes a provider's stated requirements or reporting behaviour at the time of access. Neither supplies a universal causal model of ranking. There was no comprehensive literature search, independent extraction or quantitative meta-analysis. The selected evidence is sufficient to clarify the measurement problem, but not to estimate an average commercial return from a named optimisation service.

3. From a page to an answer

3.1 Retrieval and generation

Retrieval-Augmented Generation combines learned parametric knowledge with access to external material for generation [9]. The architecture makes a useful distinction between obtaining evidence and producing an answer from it. A page's suitability for retrieval and its contribution to the final answer are separate events. The original research concerns knowledge-intensive NLP tasks, not a direct test of marketing websites or contemporary public search rankings.

WebGPT demonstrates a browser-assisted question-answering system trained with human feedback and required to gather references [7]. It provides evidence that searching and collecting support can be integrated into answer generation. It does not show that every generated citation is equally valuable to a publisher or that the system's preferred sources maximise consumer benefit. The user-facing answer and the publisher's visibility are related but different evaluation targets.

Google’s May 2025 launch account describes AI Mode dividing a question into subtopics and issuing multiple related searches [3]. Its site-owner advice connects AI experiences with established Search practices [2]. Neither account discloses every selection rule. The entered question and the information needs addressed within a response are therefore distinct analytical units. Relevance to one subtopic does not establish prominent placement throughout the answer.

3.2 Information available and information used

Lost in the Middle reports that model performance on selected long-context tasks can depend on where relevant information occurs [10]. That finding concerns the use of supplied context, not a direct ranking factor for public search. Its relevance is narrower: making information available to a model does not establish that it is used robustly. Retrieval coverage and answer support need separate measurement even when both occur within the same system.

This distinction also limits conclusions from inspecting a response alone. An uncited page may have been retrieved and ignored, never retrieved, or represented indirectly through another source. Without access to the engine's internal process, the response does not identify which event occurred. An external visibility study can measure what was displayed. It should avoid claiming knowledge of hidden retrieval steps unless the system exposes them.

4. The GEO experiment

4.1 Intervention and engine

Aggarwal and colleagues evaluate textual interventions in a generative-engine setup using a benchmark of 10,000 queries. Their main setup retrieves five sources and generates responses with GPT-3.5-turbo; they also examine a deployed engine. The interventions include adding source citations, quotations or statistics and changing presentation. The paper reports visibility improvements that vary by method and domain [1]. Those findings concern the tested engine and metrics rather than every later search product.

4.2 Visibility metrics

The study's position-adjusted word-count metric combines the amount of answer text associated with a source and the position of its citation. Its subjective-impression measures use model-based evaluation. These are measures of representation within generated responses. They are not observed click-through, purchase or brand-lift outcomes. The authors also note that engine evolution can require adaptation of methods and that the study did not evaluate effects on ordinary search rankings [1].

The outcome definition changes the meaning of an improvement. A source may receive a larger share of words in an answer because it supplies a detailed explanation, while attracting no additional visit. Another source may be represented briefly but answer a commercially important question. A common numerical visibility score cannot establish which has greater business value without a separate model or observation of that value. The primary experiment should therefore be read as an answer-representation study.

Likewise, a relative change needs its baseline and denominator. Improving a source's share within a generated answer differs from increasing the probability that any answer cites it at all. An experiment can hold the retrieved source set fixed while studying how the generator uses the text. A live deployment may change retrieval at the same time. These interventions address different stages and cannot be assigned the same interpretation merely because both are described as optimisation.

5. What platform documentation establishes

5.1 Eligibility and ranking

The May 2025 site-owner advice identifies crawl access, functioning pages and indexable content as technical prerequisites [2]. These are eligibility conditions, not a controlled estimate of an optimisation effect. Technical access can be assessed separately from selection for a particular response. The platform account supplies context for visibility experiments while leaving their external validity unresolved.

The dated Search Essentials announcement provides the wider technical and policy context for inclusion in Google Search [5]. Its role in the evidence base is to identify the framework within which the AI-feature accounts sit. A research paper demonstrating improved visibility in a separate engine cannot override those platform requirements. Equally, a platform's broad best-practice statement does not negate a measured intervention effect in another specified experimental system. The sources answer different questions.

5.2 Preview controls and structured data

The May 2025 advice also addresses preview controls and alignment between structured data and visible content [2]. Later claims about a specific AI optimisation file, schema or reporting interface are excluded from this evidence snapshot. A technical convention and a measured visibility effect are different propositions. Implementing machine-readable information does not alone demonstrate an increase in citations, visits or commercial outcomes; those endpoints require observations under the actual search and query conditions.

Structured-data documentation explains how machine-readable markup describes page information and can support eligible search features [6]. Its presence should remain consistent with visible content. A structured-data implementation is a technical property, while increased generative citation is an empirical outcome. The cited documentation does not supply a universal percentage gain from adding markup, nor does the absence of such a gain imply that clear data representation has no other use.

6. Attribution quality is distinct from citation quantity

ALCE evaluates generated answers along fluency, correctness and citation-quality dimensions. Gao and colleagues show that providing citations does not eliminate incomplete support in the systems they evaluated [8]. The research is directly relevant to interpreting visibility: a source can be cited in a response that makes unsupported or misleading claims. Counting the citation alone cannot determine whether the brand or information was represented accurately.

The Attributable to Identified Sources framework specifies a method for evaluating whether generated statements about the external world are supported by identified material [12]. It distinguishes the existence of a reference from the relationship between that reference and the claim. For a marketing visibility assessment, that distinction permits an audit of the actual assertion associated with a brand. A mention accompanied by a citation can still be inaccurate, incomplete or attached to the wrong source.

FActScore breaks long-form output into atomic claims and evaluates the portion supported by a reliable source [11]. Its empirical setting includes generated biographies, so its reported scores should not be transferred to brand recommendations. The methodological contribution is useful: a paragraph can mix supported and unsupported statements, making a single all-correct judgement too coarse. A brand audit can therefore distinguish correct product facts, unsupported comparisons and mistaken details within the same answer.

KILT evaluates knowledge-intensive tasks together with provenance against a shared knowledge source [13]. That design illustrates the value of a defined reference corpus when evaluating evidence use. Public search is less controlled: sources change, pages disappear and the engine's corpus is not fully visible. A comparison of citations over time must therefore record enough context to distinguish a change in the answer system from a change in the available evidence.

7. Measurement in a changing engine

7.1 Query sampling

A visibility study needs a query population. Branded navigational questions, generic category questions and detailed comparison questions place different demands on an answer system. A collection chosen only because it already produces favourable mentions cannot estimate visibility across the broader population of interest. The sampling rule and its relationship to real user demand should be stated before the resulting share of mentions is interpreted.

Repeated prompts can reveal response variation under a fixed testing protocol. They cannot, by themselves, establish how frequently real people ask those questions. A study with a thousand repetitions of a narrow query remains narrow in substantive scope. Query volume, linguistic variation and purchase relevance require separate evidence. A useful report therefore distinguishes the number of generated responses from the number and representativeness of the underlying information needs.

7.2 Time and configuration

The date, interface, location and any known model version are part of the observation. Search systems change retrieval sources, presentation and generation methods. A before-and-after difference can therefore occur without a publisher changing its page. Conversely, a page intervention can coincide with an engine update. The observation is a valid description of what appeared, but identifying the effect of the intervention requires a comparison that addresses those other changes.

Matched queries and repeated runs improve interpretability by reducing some sources of variation. A contemporaneous comparison group can help distinguish a broad engine change from a page-specific change. These are design principles for a causal comparison, not a claim that the current review conducted one. Without such a design, an increase after editing remains a temporal association rather than a demonstrated effect of the edit.

7.3 What official reporting can show

Bing's February 2026 announcement describes citation counts, cited pages and visibility trends across supported AI experiences. It explicitly distinguishes citation frequency from importance, ranking and placement [4]. This is useful platform evidence because it defines what the reported quantity means. It also sets a boundary: a citation count does not directly measure the number of people influenced, the prominence of a brand or the value of resulting visits.

Reporting interfaces evolve. The dated accounts do not establish the contents of every later reporting interface. Report definitions must be matched to the account and observation period. A measurement process should retain the report name and field definitions used. Combining a platform citation count, an external prompt-panel score and website referral traffic under one label would lose the different denominators each quantity uses.

8. From visibility to a marketing outcome

A cited page can receive traffic, but the citation and the visit are different events. A visitor may arrive already informed by the answer or may be checking a claim. Neither behaviour determines the visitor's subsequent purchase. An observed referral can be useful operational evidence while leaving incrementality unresolved: the person might otherwise have found the same brand through another route. The reviewed visibility literature does not estimate that counterfactual for a particular business.

Brand representation also has a qualitative dimension. A model can name a brand in response to a relevant question but describe a discontinued product, infer an unsupported benefit or place the brand in an inappropriate category. More mentions can therefore coexist with poorer information quality. An assessment that records the associated claim and its source support can distinguish expanded exposure from improved representation [8,11,12].

The value of a source mention depends on the decision context. A brief reference in an information-seeking answer and a detailed comparison during product selection may have different practical significance. Assigning weights to those contexts is a marketing measurement decision requiring evidence about the audience and task. The reviewed papers do not supply universal weights that turn every citation into an equivalent unit of demand.

9. Content interventions and the limits of inference

A publisher can improve factual clarity, source support and technical access for readers regardless of whether a generative engine rewards the change. The causal claim that a particular edit increases visibility still requires observation. Adding a statistic is especially dependent on validity: the number must be relevant, correctly sourced and represented with its denominator. A visibility objective does not make an invented or context-free statistic useful evidence.

Evidence-rich content also differs from citation decoration. A reference should support the adjacent claim, while the claim should answer the reader's actual question. A long list of sources can increase the appearance of authority without improving that relationship. Attribution research supplies a way to inspect support, but does not show that adding references mechanically produces ranking gains. Quality of evidence and visibility remain related questions that require separate evaluation.

The engine's objective and the publisher's objective may diverge. A system can produce a satisfactory answer while sending fewer users to a source, or cite a source to qualify a claim rather than endorse it. An optimisation study that measures answer representation provides useful information about the engine's use of content. It should not be recast as a guarantee that the publisher benefits economically from every additional unit of representation.

10. Limitations and conclusion

The central optimisation evidence concerns specified historical engines and a defined benchmark. Adjacent retrieval and factuality research clarifies mechanisms and measurement but does not directly estimate public-search visibility gains. Dated platform accounts were checked on 21 October 2026 and selected within the historical evidence boundary. Later revisions of rolling documentation are not incorporated. This review conducted no site experiment, measured no live brand share of answers and calculated no commercial return. Several empirical papers were accessed only through abstracts, limiting detailed methodological appraisal.

The available evidence supports careful measurement of source representation within generated answers and shows that textual interventions can affect it in tested systems. It does not establish a universal optimisation formula or a direct equivalence between citations and business growth. A defensible marketing assessment identifies the engine, query population, observed representation, factual support and downstream outcome separately. Improvements can then be attributed to the level actually measured, with their scope preserved.

Source access and appraisal register

The GEO manuscript [1] was inspected for setup, metrics, results and limitations. Dated platform accounts [2], [3], [4] and [6] support the attributed statements; reference [5] was checked through its official publication record. Adjacent research references 7–13 were consulted through author abstracts and publication records. No live engine experiment or independent replication was conducted. Platform accounts describe their stated operation, rather than establish a causal estimate for every website.

References