sarcmethodology.com

Abstract

Generative search systems such as ChatGPT search, Perplexity and Google AI Overviews no longer return a ranked list of links. They retrieve passages, synthesise an answer and cite a small number of sources. This change moves the unit of competition from the web page to the passage, and it raises a question that information retrieval research has only begun to address: which properties of a text make it likely to be retrieved, trusted, extracted and cited, and do those same properties move the human reader who follows the citation to act? This article proposes the SARC Content Methodology, a four-component framework of Search, Authority, Relevance and Conversion, as a linguistically grounded answer. Each component is mapped to an established body of linguistic theory: lexical semantics and query–text alignment for Search; stance, evidentiality and attribution for Authority; relevance theory and information structure for Relevance; and persuasion and genre analysis for Conversion. Drawing on more than sixty sources across information retrieval, natural language processing, pragmatics, discourse analysis and persuasion research, the article derives thirteen testable propositions and sets out a mixed-methods research agenda combining corpus analysis, controlled rewriting experiments and cross-engine replication. The framework originated in professional content practice; its contribution here is to restate that practice as theory that can be falsified. Because most readers of a generative answer never visit its sources, the article also argues that being cited is not enough: the cited passage must give readers a reason to click, and the page must turn those visitors into buyers, because revenue comes from customers, not from appearances in AI answers. More broadly, it argues that generative search is, at bottom, a linguistic filter, and that linguistics is therefore well placed to explain it.

Keywords: generative engine optimization; retrieval-augmented generation; relevance theory; information structure; stance; evidentiality; genre; AI search; content methodology

1. Introduction

For roughly twenty-five years, being found online meant being ranked. A search engine matched a query against an index, scored documents and returned an ordered list of links; the user then chose which link to open (Broder, 2002; Manning et al., 2008). Visibility was a matter of position, and position was a matter of matching terms, earning links and satisfying technical criteria. The text itself mattered, but chiefly as a container of keywords.

Generative search changes this arrangement. Systems built on large language models retrieve candidate passages, place them in the model’s context window and generate a single synthesised answer, often with inline citations (Lewis et al., 2020; Nakano et al., 2021; Metzler et al., 2021). Aggarwal et al. (2024) describe these systems collectively as generative engines and observe that they answer queries by synthesising information from multiple sources. The user may never see a list of links at all. What they see is prose, assembled from fragments of other people’s prose, with a handful of small numbered markers pointing back to the sources that were chosen.

This shift has three consequences that motivate the present article.

First, the unit of competition has changed. A traditional search engine ranks documents; a generative engine selects passages. A page can contain the right answer and still be ignored if the answer is buried, ambiguous or dependent on context the model does not have. Research on long-context models shows that position within the input affects whether information is used at all: Liu et al. (2024) found that performance is often highest when relevant information appears at the beginning or end of the input and degrades when it sits in the middle. Shi et al. (2023) showed that irrelevant material in the context can substantially reduce accuracy. The internal organisation of a text therefore matters in a way it did not when the whole page was the unit of retrieval.

Second, the decision to cite is partly a judgement about language. Generative engines do not cite everything they retrieve. Studies of commercial systems found that only about half of generated sentences were fully supported by their citations, and roughly three quarters of citations supported the sentence they were attached to (Liu, Zhang, & Liang, 2023). Wan et al. (2024) found that models rely heavily on how relevant a document is to the query and largely ignore stylistic features that humans weigh, such as scientific references or a neutral tone. Xie et al. (2024) showed that models are highly receptive to external evidence that is coherent and convincing, even when it conflicts with what the model already holds. These findings suggest that the properties which earn a citation are linguistic properties: lexical overlap, coherence, directness, the marking of evidence. They also suggest that those properties are not the same ones that earn human trust.

Third, visibility and commercial value have separated. When a user reads a synthesised answer, the cited source may receive no visit. If a visit does come, it comes from a reader who has already been partly answered and who arrives with a narrower, more specific need. Content that is written only to be cited may therefore gain visibility and still fail to produce action. Any account of content for generative search has to address both audiences: the model that selects and the person who acts.

1.1 The research gap

A new literature on generative engine optimization (GEO) has begun to measure how edits to web content change its visibility in generative answers. Aggarwal et al. (2024) introduced the term, built a benchmark of queries and sources, and reported that some optimisation strategies could raise visibility by up to 40%, with effects varying by domain. Kumar and Lakkaraju (2024) showed that inserting an optimised text sequence into a product page could move that product to the top of an LLM’s recommendations. Nestaas et al. (2025) demonstrated that crafted content could manipulate production systems such as Bing and Perplexity to favour an attacker and disparage competitors.

This work is valuable, but it shares two limitations. It is largely computational, treating text as a string to be optimised rather than as language with structure and function. And it is largely adversarial or engineering-led, asking what works rather than why. Few studies ask which linguistic properties drive selection, how those properties relate to established theory, or whether the text that wins a citation also earns the click and serves the reader who arrives. This matters commercially as well as theoretically. When a generative answer satisfies the user, few go on to visit its sources (Pew Research Center, 2025), and a citation that produces no visit produces no customers and no revenue. Shah and Bender (2022) warned that generative systems risk flattening the varied ways people seek and evaluate information; a linguistic account of what these systems reward is one way to make their behaviour visible and open to critique.

1.2 The SARC framework

The SARC Content Methodology was developed in professional practice as a way of writing content that is found by search and AI systems and that converts the readers it brings. Its four components are:

1. Search: whether a text is retrieved at all for the queries people actually ask.
2. Authority: whether a system and a reader treat the text as a trustworthy source worth quoting and attributing.
3. Relevance: whether each passage answers the intent behind a query directly enough to be extracted.
4. Conversion: whether the citation earns a click, and whether the reader who arrives goes on to act by becoming a potential customer or a buyer.

In practice these components functioned as a checklist. This article reframes each one as a claim about measurable properties of text, grounds it in linguistic theory, and derives propositions that can be tested.

1.3 Contribution and structure

The article makes three contributions. It offers a theoretical synthesis that connects information retrieval and generative search research with pragmatics, systemic functional linguistics, discourse analysis and persuasion research. It provides an operational framework in which each SARC component is tied to features that can be annotated or measured in a corpus. And it sets out a research agenda with thirteen propositions and a three-phase design for testing them.

Section 2 reviews the technical and linguistic background. Sections 3 to 6 develop each SARC component in turn. Section 7 integrates them into a single model and states the propositions. Section 8 proposes methods. Section 9 discusses limitations and ethics and concludes.

2. Background

Generative search sits on four decades of information retrieval research and on a much shorter history of neural language modelling. This section sets out the parts of each that SARC depends on, then introduces the linguistic traditions the framework draws on.

2.1 From term matching to dense retrieval

Classical retrieval represents documents and queries as bags of terms and scores their overlap. Weighting schemes such as tf-idf reward terms that are frequent in a document but rare across the collection (Salton & Buckley, 1988). The probabilistic relevance framework, of which BM25 is the best-known instance, formalised this into a model of the probability that a document is relevant given its term statistics (Robertson & Zaragoza, 2009). In these systems, the words a writer chooses decide whether a document can be found. A page that never uses the user’s vocabulary is invisible, however good its content.

Early attempts to escape strict term matching used latent semantic analysis to map words and documents into a reduced space in which synonyms sit close together (Deerwester et al., 1990). Distributed word representations extended this idea, learning vectors from co-occurrence so that semantically related words have similar representations (Mikolov et al., 2013). The transformer architecture (Vaswani et al., 2017) and pre-trained encoders such as BERT (Devlin et al., 2019) made it possible to represent whole sentences and passages in context. Sentence-BERT adapted these encoders to produce embeddings that can be compared efficiently (Reimers & Gurevych, 2019), and dense passage retrieval showed that learned passage embeddings could outperform BM25 for open-domain question answering (Karpukhin et al., 2020). Neural re-rankers then score a shortlist of candidates more precisely, often by reading the query and passage together (Nogueira & Cho, 2019).

Two points from this history matter for SARC. First, modern pipelines usually split documents into passages before indexing them, so the passage, not the page, is what is retrieved. Second, most production systems combine lexical and dense signals. A text therefore benefits from matching both the literal words of a query and its meaning.

2.2 Retrieval-augmented generation and generative engines

Large language models store a great deal of knowledge in their parameters, but that knowledge is uneven. Kandpal et al. (2023) found that a model’s accuracy on a factual question rises with the number of pre-training documents relevant to it, and that retrieval reduces this dependence. Models also generate fluent statements that are unsupported or false, a problem surveyed under the heading of hallucination (Ji et al., 2023).

Retrieval-augmented generation (RAG) addresses both problems by retrieving relevant passages at inference time and conditioning generation on them (Lewis et al., 2020). Retrieved text can simply be placed into the model’s context alongside the query, without changing the model itself (Ram et al., 2023). WebGPT showed that a model could be trained to browse the web, collect references and compose answers that cite them (Nakano et al., 2021). Metzler et al. (2021) argued that combining retrieval with language models could move search from pointing users towards answers to providing expert-quality answers directly, provided those answers can be justified by reference to sources.

Citation is therefore central to generative search, and it has become a research problem in its own right. Rashkin et al. (2023) proposed the Attributable to Identified Sources (AIS) framework for judging whether generated statements are supported by a source. Gao et al. (2023) built ALCE, a benchmark for evaluating citation quality, and found that even the best models lacked complete citation support about half of the time on one dataset. Liu, Zhang and Liang (2023) audited four commercial generative search engines and found that responses were fluent and appeared informative but frequently contained unsupported statements and inaccurate citations. They also observed that systems often copied or closely paraphrased the pages they cited.

This last finding matters for SARC. If systems prefer to lift sentences that can stand as answers, then the form of a sentence affects whether it is chosen.

2.3 Generative engine optimization

GEO research asks how a content creator can raise their visibility in generative answers. Aggarwal et al. (2024) formalised visibility metrics, released a benchmark of queries, and tested a set of editing strategies. Their results showed that strategies such as adding citations, quotations and statistics could improve visibility, and that effectiveness varied by domain.

Other studies have taken an adversarial view. Kumar and Lakkaraju (2024) used a catalogue of fictitious coffee machines to show that a strategically optimised text sequence on a product page could make an LLM recommend that product first. Nestaas et al. (2025) introduced preference manipulation attacks against production systems and plugin APIs. They also identified a prisoner’s dilemma: each attacker gains individually, but when several attack the same system, all lose visibility.

On the model side, Wan et al. (2024) built a dataset of contentious questions paired with conflicting real-world evidence. Perturbations that increased a text’s relevance to the query had a strong positive effect on whether models agreed with it, while many stylistic changes that influence humans had neutral or negative effects. Xie et al. (2024) found that models readily accept coherent external evidence but also show confirmation bias when some evidence agrees with their prior knowledge.

Taken together, this literature shows that generative engines can be influenced by the text they retrieve, that the influence is measurable, and that it can be abused. What it does not offer is a theory of why particular textual features matter. That is the gap SARC addresses.

2.4 Human information seeking and relevance

Generative engines serve people, and a long tradition in information science describes how people seek information. Taylor (1968) showed that an information need passes through stages, from a vague sense of a gap to a formal question, and that the question finally asked is a compromise with the system. Belkin (1980) described the starting point as an anomalous state of knowledge: users cannot specify exactly what they need because they do not yet know it. Broder (2002) classified web queries as navigational, informational or transactional, a taxonomy that still shapes how intent is modelled.

Relevance itself has been studied as a family of related notions rather than a single property. Saracevic (2007) reviewed thirty years of work and described a system of interdependent relevances, from the topical match between query and text to the situational usefulness of information to a particular person at a particular moment. Information foraging theory adds an economic view: people follow cues, or information scent, and abandon a source when the expected gain from staying falls below the gain from moving on (Pirolli & Card, 1999). Studies of click behaviour show that people trust search rankings heavily and tend to click what is shown first (Joachims et al., 2005).

These accounts describe the person on the other side of the generative answer. SARC’s Relevance and Conversion components depend on them.

2.5 Linguistic perspectives

Four linguistic traditions supply the theoretical core of SARC.

Pragmatics studies how meaning depends on context and inference. Grice (1975) proposed that conversation is governed by a cooperative principle and maxims of quantity, quality, relation and manner. Relevance theory reworked this into a cognitive account: an input is relevant to the extent that it yields positive cognitive effects for a small processing effort (Sperber & Wilson, 1995; Wilson & Sperber, 2004). Rashkin et al. (2023) draw directly on relevance theory’s notion of explicature to define what a generated sentence asserts.

Information structure concerns how sentences package given and new information, topic and focus (Chafe, 1976; Lambrecht, 1994; Krifka, 2008). Halliday (1967) introduced the distinction between theme and rheme. Daneš (1974) showed how themes chain across sentences to give a text its thematic progression. Clark and Haviland (1977) described a given–new contract in which writers signal what the reader already knows before adding something new.

Discourse and text linguistics describe how sentences hold together. Halliday and Hasan (1976) catalogued the cohesive devices of English. Rhetorical structure theory models texts as trees of relations such as elaboration, evidence and contrast (Mann & Thompson, 1988). Psycholinguistic models describe comprehension as the construction of a propositional representation from text (Kintsch & van Dijk, 1978; Kintsch, 1988).

Evaluation, stance and genre concern how writers position themselves and their claims. Work on stance, hedging, metadiscourse and appraisal describes how writers signal certainty, attribute knowledge and engage readers (Biber & Finegan, 1989; Hyland, 1998, 2005a, 2005b; Martin & White, 2005). Genre analysis describes how texts are organised into recognisable moves that serve a communicative purpose (Swales, 1990; Bhatia, 1993).

The next four sections apply these traditions to the four SARC components.

3. Search: lexical and semantic alignment

The Search component concerns the first gate a text must pass: retrieval. A passage that is never retrieved cannot be read, weighed or cited by a generative engine. In SARC, Search is defined as the degree to which a text’s wording aligns with the ways people phrase their needs, at both the lexical and the semantic level, and the degree to which each passage remains retrievable once it is separated from the page around it.

3.1 The vocabulary problem

People do not search in the vocabulary of experts. Taylor (1968) showed that the question a person finally puts to a system is a compromise, shaped by what they think the system can understand. Belkin (1980) argued that users often cannot name what they need because the need arises from a gap in their knowledge. The result is a persistent mismatch between the words searchers use and the words authors use. A cybersecurity vendor may write about “application security posture management” while a buyer asks how to “find security problems in my code before release”.

Lexical retrieval models make this mismatch costly. Under tf-idf and BM25, a document scores only for terms it contains (Salton & Buckley, 1988; Robertson & Zaragoza, 2009). Dense retrieval reduces the problem by mapping semantically similar texts to nearby points in a vector space (Karpukhin et al., 2020; Reimers & Gurevych, 2019), but it does not remove it. Dense models are trained on particular distributions of queries and passages, and they can miss rare terms, exact names and specialised vocabulary. This is why many production systems combine lexical and dense retrieval. For a writer, the practical implication is that a text should carry both the precise term and the everyday paraphrase.

Lexical semantics offers a way to describe this. Relations such as synonymy, hyponymy and meronymy organise vocabulary into networks. A query for “car insurance for new drivers” may be satisfied by a passage about “motor cover for learner and probationary drivers”, but only if the retrieval system recognises the relations between those terms. SARC’s Search component asks writers to map the lexical field of a topic deliberately: the expert term, its common paraphrases, its superordinate and subordinate terms, and the question forms in which it is typically asked.

3.2 Questions as the new query form

Generative engines encourage longer, conversational queries. Instead of typing two or three keywords, users ask full questions, often with situational detail. This brings query language closer to ordinary speech and therefore closer to the pragmatic structures linguists study. A question has a presupposition, a focus and an expected answer type (Krifka, 2008). “How long does a UK visa decision take?” presupposes that decisions take time, focuses on duration and expects an answer measured in days or weeks.

Texts that mirror question forms in their headings, and answer them in the first sentence beneath, align with this structure at two levels. The heading matches the query lexically. The answer sentence supplies the expected answer type. Broder’s (2002) taxonomy remains useful here: informational queries call for definitions, explanations and comparisons, transactional queries for steps and options, and navigational queries for a named destination. A text aligned for Search signals which type of need it serves.

3.3 Passage independence and the cohesion problem

Retrieval systems usually split documents into passages before indexing them. This creates a specific linguistic risk. Cohesion, in Halliday and Hasan’s (1976) sense, depends on ties that link sentences together: reference (pronouns, demonstratives), substitution, ellipsis, conjunction and lexical cohesion. These ties make a text read smoothly as a whole. When a passage is cut out of its document, however, the ties that point outside the passage break.

Consider a passage that begins: “It also reduces this risk by about a third.” A human reader of the full page knows what “it” and “this risk” refer to. A retrieval system that sees only this passage does not, and neither does a generative model that receives it as one of several fragments. The passage contains the answer but not the words that would let it be matched to the question or attributed to a topic.

SARC therefore treats passage independence as part of Search. Each passage that is meant to be found should name its subject explicitly, restate the key entity rather than relying on a pronoun across a section break, and avoid conjunctions that depend on a preceding paragraph. This is not an argument against cohesion. It is an argument for local cohesion, where ties resolve inside the unit that will be retrieved. In Kintsch and van Dijk’s (1978) terms, each passage should allow a reader to build a coherent propositional representation without access to the rest of the text.

3.4 Definitional and entity-centred sentences

Some sentence types are especially easy to retrieve and reuse. Definitional sentences of the form “X is a Y that Z” state an entity, place it in a category and give a distinguishing feature. They match “what is X” queries lexically and supply exactly the answer type such queries expect. Entity-centred sentences, which name a product, organisation, person or concept as grammatical subject, give retrieval systems a clear anchor.

The long-tail findings of Kandpal et al. (2023) add a further reason to attend to this. Language models answer questions about frequently mentioned entities from their parameters, but they rely more on retrieval for rare ones. For specialised topics, niche products and new concepts, the retrieved passage may be the model’s only source of information. A clearly stated definition on a small company’s page may therefore have a larger effect on the generated answer than a similar definition about a well-known subject.

3.5 Limits of the Search component

Search alignment is necessary but not sufficient. Wan et al. (2024) found that perturbations which raised a document’s relevance to the query had a strong positive effect on model agreement, but their notion of relevance went beyond keyword overlap. Excessive repetition of query terms, the old practice of keyword stuffing, may help lexical retrieval while harming readability and trust. Search therefore sets the conditions for selection without determining it. Whether a retrieved passage is used depends on the next two components.

3.6 Operationalising Search

For research purposes, Search can be measured through several features of a passage:

• lexical overlap with a set of real user queries, measured with BM25 or a similar score;
• semantic similarity to the same queries, measured with sentence embeddings;
• coverage of the topic’s lexical field, measured against a list of expert terms and paraphrases;
• the proportion of sentences whose referring expressions resolve within the passage;
• the presence of explicit definitional or entity-centred sentences.

These features lead to three propositions:

P1. Passages that contain both the expert term and the common paraphrase used in real queries are retrieved and cited more often than passages that contain only one.

P2. Passages whose referring expressions resolve within the passage are cited more often than otherwise similar passages that depend on context outside the passage.

P3. For definitional queries, passages containing an explicit definitional sentence are cited more often than passages that convey the same information indirectly, and the effect is larger for long-tail entities.

4. Authority: stance, evidentiality and attribution

The Authority component concerns whether a retrieved text is treated as a source worth relying on. In SARC, Authority has two layers. Source authority is the standing of the author or organisation. Textual authority is the set of linguistic choices through which a text shows where its claims come from and how certain they are. SARC focuses on textual authority, because it is the layer a writer controls directly and the layer a model can read.

4.1 Credibility in persuasion and web research

The study of credibility begins with source effects. Hovland and Weiss (1951) found that the same message was judged more favourably when attributed to a trustworthy source, although the difference faded over time. Pornpitakpan’s (2004) review of five decades of research confirmed that highly credible sources are generally more persuasive, while noting many moderating conditions.

Online, credibility judgements are fast and heuristic. Fogg’s (2003) prominence–interpretation theory holds that people first notice certain elements of a website and then interpret them; an element that is not noticed cannot affect the judgement. Metzger (2007) reviewed models of online credibility and showed that users rarely carry out the systematic checks recommended to them. Metzger and Flanagin (2013) described the cognitive heuristics people use instead, such as reputation, endorsement, consistency with other sources and expectancy violation. Sundar’s (2008) MAIN model attributes credibility judgements to cues arising from modality, agency, interactivity and navigability.

Expert evaluators behave differently. Wineburg and McGrew (2019) compared historians, students and professional fact checkers. Historians and students often stayed on a website and judged it by its own features, while fact checkers left the site quickly and read laterally, opening other tabs to see what independent sources said about it. Lateral reading is a useful analogy for generative engines. A system that retrieves several passages and compares them is, in effect, reading laterally. A claim that agrees with other retrieved sources, or that can be checked against them, is easier for the system to use.

4.2 Evidentiality and the marking of sources

Linguistics has a precise vocabulary for how speakers mark the source of their knowledge. Evidentiality refers to the grammatical or lexical marking of information source: whether something was seen, inferred, reported or assumed (Chafe & Nichols, 1986). Some languages encode evidentiality in obligatory grammatical morphemes; English does so mainly through lexical and phrasal means such as “according to”, “research shows”, “we found” and “reportedly” (Aikhenvald, 2004).

For a generative engine that must attribute statements, explicit evidential marking has a clear function. Rashkin et al. (2023) define attribution in terms of whether a statement can be verified against an identified source. A sentence such as “In a 2023 audit of four generative search engines, only about half of generated sentences were fully supported by their citations” carries its evidence with it. A sentence such as “Most AI answers are unreliable” does not. The first can be lifted into an answer and cited cleanly; the second can only be paraphrased or ignored.

The GEO literature points in the same direction. Aggarwal et al. (2024) reported that strategies such as citing sources, adding quotations and adding statistics were among those that improved visibility in generative responses. Read through the lens of evidentiality, each of these strategies adds explicit evidential marking to a passage.

4.3 Stance, hedging and calibrated certainty

Writers also signal how committed they are to a claim. Biber and Finegan (1989) identified styles of stance in English based on markers of certainty, doubt and affect. Hyland (1998) showed that hedges such as “may”, “suggests” and “appears to” are central to scientific writing, allowing authors to present claims with appropriate caution. Hyland (2005b) later proposed a model of stance and engagement in which writers use hedges, boosters, attitude markers and self-mention to position themselves, and use reader pronouns, directives and questions to engage the audience. Du Bois (2007) described stance as a single act with three parts: evaluating an object, positioning the self and aligning with others.

Appraisal theory adds the dimension of engagement, which concerns how a text acknowledges other voices (Martin & White, 2005). A monoglossic statement presents a claim as undisputed fact. A heteroglossic statement opens space for other views, either by attributing the claim to a source or by acknowledging alternatives. Hunston and Thompson (2000) showed that evaluation runs through texts of every kind, not only openly persuasive ones.

These frameworks suggest a hypothesis about calibration. A text that boosts unsupported claims (“this is the best tool on the market”) offers little that a careful system can attribute and may look promotional. A text that hedges every statement offers nothing definite to extract. Between these lies calibrated certainty: claims stated firmly where evidence is attached, and hedged where it is not. SARC predicts that calibrated texts are more useful to generative engines, because they supply extractable assertions together with the evidence that licenses them.

4.4 Argument structure

Toulmin’s (1958) model of argument divides a reasoned case into a claim, the data that support it, the warrant connecting data to claim, and qualifiers and rebuttals that limit it. Rhetorical structure theory similarly treats evidence and justification as relations between parts of a text (Mann & Thompson, 1988). Both frameworks give a way to annotate whether a passage contains a claim alone, or a claim with its support. SARC’s Authority component asks that important claims appear in full argumentative form within a single passage, so that the support is retrieved with the claim.

4.5 Where human and machine authority diverge

There is good reason to think that models and people do not weigh authority in the same way. Wan et al. (2024) found that models relied heavily on relevance and largely ignored features humans find important, including scientific references and a neutral tone. Xie et al. (2024) showed that models accept coherent, convincing external evidence even when it contradicts their prior knowledge, which means that a well-formed but false passage may be persuasive to a model. Nestaas et al. (2025) found that content which disparaged competitors could cause a system to ignore them. Liu, Zhang and Liang (2023) found that systems’ citations often failed to support the sentences they were attached to.

These findings complicate any simple claim that “authoritative” text is cited more. They also define an important research question. If the features that earn a model’s citation differ from those that earn a reader’s trust, then a text optimised for one may fail the other. SARC’s position is that Authority must serve both audiences. Evidential marking and calibrated stance should help a system attribute claims, and they should also help the human reader who arrives through the citation to judge them. Whether this dual effect holds is an empirical matter.

4.6 Operationalising Authority

Textual authority can be annotated and measured through:

• the density of evidential markers per claim, distinguishing attributed, data-backed and unmarked assertions;
• the ratio of hedges to boosters, and whether boosters co-occur with evidence;
• the proportion of claims presented with supporting data in the same passage, using Toulmin or RST categories;
• the presence of specific, checkable details such as dates, figures, named studies and named organisations.

Three propositions follow:

P4. Passages that mark the source of their claims explicitly are cited more often than passages that make the same claims without evidential marking.

P5. The relationship between stance and citation is non-linear: passages with calibrated certainty are cited more often than passages that are either heavily hedged or that boost claims without evidence.

P6. Authority cues that raise human credibility judgements, such as credentials and scholarly references, have a weaker effect on model citation than on human trust, so that model and human rankings of the same passages diverge.

5. Relevance: relevance theory and information structure

The Relevance component concerns what happens after retrieval: whether a passage answers the question directly enough to be extracted and used. Search decides whether a passage enters the model’s context. Relevance decides whether the model draws on it once it is there. In SARC, Relevance is defined as the ratio of useful answer to processing effort within a passage, together with the placement of that answer where a reader or model will find it first.

5.1 Two traditions of relevance

The word “relevance” carries different meanings in information science and in linguistics, and SARC draws on both.

In information science, relevance is the basic criterion by which retrieval is judged. Saracevic (2007) showed that it is not a single property. Topical relevance concerns whether a document is about the subject of the query. Situational relevance concerns whether it helps a particular person with a particular task at a particular moment. A passage can be topically relevant and situationally useless, for example by explaining a concept at length when the user wanted a number.

In linguistics, relevance theory offers a cognitive definition. Sperber and Wilson (1995) propose that an input is relevant to an individual when processing it yields a positive cognitive effect, such as answering a question, strengthening a belief or correcting an error. Relevance increases with the size of the effect and decreases with the effort required to obtain it. Communication works because hearers expect an utterance to be relevant enough to be worth processing (Wilson & Sperber, 2004). Grice (1975) captured a related intuition in the maxims of quantity (say as much as needed and no more), relation (be relevant) and manner (be clear and orderly).

The effect–effort trade-off maps closely onto how generative engines handle retrieved text. A retrieved passage that answers the query in its first sentence offers a large effect for little processing. A passage that reaches the answer only after several sentences of background offers the same effect for greater effort, and it carries more material that could mislead or distract.

5.2 Evidence from model behaviour

Three lines of evidence support this mapping. Liu et al. (2024) showed that language models use information best when it sits at the beginning or end of their input and worst when it sits in the middle. Although their study concerns the position of documents within a long context, it establishes that position affects use. Shi et al. (2023) found that adding irrelevant information to a problem substantially reduced model accuracy, which suggests that text surrounding an answer is not neutral. Wan et al. (2024) found that relevance to the query was the dominant factor in whether a model agreed with a piece of evidence.

Taken together, these findings imply that a passage should put its answer early and keep unrelated material to a minimum. This is relevance theory’s effect–effort principle restated for machine readers.

5.3 Information structure: answer first

Information structure gives a finer account of what “answer first” means. Halliday (1967) distinguished the theme of a clause, its point of departure, from the rheme, what is said about it. Halliday and Matthiessen (2014) describe how the choice of theme orients the reader to what the clause is about. The given–new contract holds that writers should present known information before new information, so that readers can attach the new to something they already have (Clark & Haviland, 1977). Chafe (1976) and Lambrecht (1994) described how speakers mark information as given or new, active or inactive in the hearer’s mind.

Krifka (2008) linked information structure to questions. A question sets up a set of alternative answers, and the focus of an answer is the part that picks one of them. In “The decision usually takes three weeks”, the focus is “three weeks”, because it answers the implied question “How long does the decision take?” When a heading poses a question and the following sentence takes the question’s topic as its theme and supplies the focus immediately, the passage’s information structure matches the query’s.

SARC’s Relevance component therefore recommends a specific pattern. A section begins with a question or a topic heading. The first sentence takes the topic as theme and gives the answer as focus. Supporting detail, qualifications and examples follow. In rhetorical structure theory, the nucleus comes before its satellites (Mann & Thompson, 1988). This pattern is familiar from journalism’s inverted pyramid. Its new justification is that it suits the way both readers and models extract information.

5.4 Thematic progression and topical focus

Daneš (1974) described patterns of thematic progression. In constant progression, successive sentences share the same theme. In linear progression, the rheme of one sentence becomes the theme of the next. In derived progression, themes are drawn from a superordinate topic. Texts that switch themes frequently, with no clear progression, are harder to follow.

For passage-level retrieval this has a practical consequence. A passage with a stable theme is about one thing, and its embedding will sit close to queries about that thing. A passage that drifts across several themes will produce an averaged representation that matches none of them well. Kintsch and van Dijk (1978) proposed that readers build a macrostructure, a summary of what a text is about. A passage with a clear macrostructure gives both human readers and retrieval systems a single, stable topic.

5.5 Information scent and the human reader

Relevance also governs whether a human reader stays. Information foraging theory holds that people follow cues that predict where useful information lies and leave when those cues weaken (Pirolli & Card, 1999). A reader who arrives from a generative answer has a narrower question than a reader who arrived from a list of links, because the answer has already covered the basics. Headings that echo the reader’s question and first sentences that confirm they have found the right place preserve information scent. Text that repeats what the generative answer already said, or that buries the next useful detail, weakens it.

This is where Relevance begins to connect with Conversion. The same structure that lets a model extract an answer lets a human reader confirm quickly that the page is worth their time.

5.6 Operationalising Relevance

Relevance can be measured through:

• the position of the first answer-bearing sentence within a passage, relative to a question or topic heading;
• the proportion of sentences in a passage that bear on the target query, as judged by annotators or by a similarity threshold;
• the type and consistency of thematic progression, annotated following Daneš (1974);
• whether the first sentence’s theme matches the topic of the heading and whether its focus supplies the expected answer type.

Three propositions follow:

P7. Passages that give the answer in their first sentence are cited more often than passages that convey the same answer later, holding content constant.

P8. Holding length constant, passages with a higher proportion of query-relevant sentences are cited more often than passages diluted by background or tangential material.

P9. Passages with constant or linear thematic progression are cited more often than passages whose themes shift without a clear pattern.

6. Conversion: persuasion, genre and reader action

The Conversion component concerns the step from being cited to earning revenue: whether a cited passage earns the click, and whether the human reader who follows it acts. In SARC, Conversion is defined as the likelihood that this reader takes a meaningful next step, such as signing up, making an enquiry, downloading a resource or buying. Search, Authority and Relevance can be studied by observing what generative engines do. Conversion can only be studied by observing what people do. It is included in the framework because visibility without clicks and buyers produces no revenue for the organisations that produce content, and because the readers generative search sends are different from those traditional search sent.

6.1 A different kind of visitor

In traditional search, a user scanned a list of links and clicked one, often near the top (Joachims et al., 2005). Many arrived with a broad need and little prior knowledge of the topic. In generative search, the user has usually read a synthesised answer before clicking. They know the basics. They click because they want something the answer did not give: more detail, a specific option, evidence, a price or a way to act.

The elaboration likelihood model offers a way to think about this reader. Petty and Cacioppo (1986) proposed that persuasion follows two routes. When people are motivated and able to think about a message, they take the central route and weigh the quality of its arguments. When they are not, they take the peripheral route and rely on cues such as source attractiveness or the number of arguments. A reader who has already read an AI answer and chosen to click through is likely to be more involved than an average searcher. SARC therefore predicts that, for these readers, argument quality and specificity matter more than peripheral cues.

This also ties Conversion to situational relevance (Saracevic, 2007). A page that repeats the generative answer gives this reader nothing new. A page that picks up where the answer stopped, with the detail, comparison or next step they came for, serves their situation.

6.2 Genre and the moves of persuasive text

Genre analysis studies how texts that serve a shared purpose follow recognisable structures. Swales (1990) described research article introductions as sequences of moves, such as establishing a territory, identifying a niche and occupying it. Bhatia (1993) extended the approach to professional and promotional genres, describing sales letters and similar texts as sequences of moves such as establishing credentials, introducing the offer, offering incentives and soliciting a response.

Content that aims to convert can be analysed the same way. A typical SARC-aligned page answers the question (a Relevance move), shows evidence (an Authority move), introduces an option or solution, addresses likely objections and invites a specific next step. Genre analysis gives a way to annotate these moves and to test whether their presence and order relate to conversion.

Register is also relevant. Biber (1988) showed that texts vary along several dimensions, including involved versus informational production and explicit versus situation-dependent reference. Content written for search has often drifted towards a generic informational register. Content written for conversion may need to recover some involvement, by addressing the reader directly and responding to their likely situation.

6.3 Engagement and directive speech acts

Hyland’s (2005a, 2005b) work on metadiscourse and engagement describes how writers bring readers into a text through reader pronouns, questions, directives and appeals to shared knowledge. Speech act theory classifies utterances by what they do: assert, question, promise or direct (Searle, 1969). A call to action is a directive. Its success depends on felicity conditions, including that the reader is able to perform the act and has a reason to.

SARC treats the call to action as a linguistic object that can be studied. A directive that names a concrete action, states what the reader gets and arrives after the reader’s question has been answered is more likely to meet its felicity conditions than one that appears before the reader has a reason to act.

6.4 Concreteness and specificity

Recent consumer research shows that small linguistic choices can affect behaviour. Packard and Berger (2021) analysed customer service interactions and found that when employees used more concrete language, customers were more satisfied and more likely to buy, because concrete language signalled that the employee was listening. Berger and Milkman (2012) found that online content which evoked high-arousal emotions was more likely to be shared, which shows that the emotional texture of a text affects what readers do with it.

Concreteness also links Conversion back to Authority. Specific figures, named examples and precise descriptions of what will happen next are both evidence for a claim and signals of competence. SARC predicts that concrete language serves both the model, by supplying attributable detail, and the reader, by signalling attention to their needs.

Readability is a further consideration. Flesch (1948) proposed one of the first formulas linking sentence length and word length to ease of reading. Such formulas are crude, but they capture an effort cost that relevance theory also predicts. A reader who has to work hard to find the next step is less likely to take it.

6.5 The tension with citation

Conversion introduces a tension into the framework. Generative engines may cite a passage and satisfy the user without sending a visit. The more completely a passage answers the question, the more extractable it is and the less reason the reader may have to click. SARC addresses this by separating the passage that answers from the page that converts. The answer passage is written to be complete and citable. The surrounding page is written to extend that answer for the reader who wants more. Whether this separation works in practice is an open question and one of the main reasons Conversion is part of the research agenda.

6.6 Visibility is not revenue: earning the click and the buyer

The tension described in Section 6.5 is sharper than it first appears. In Pew Research Center’s (2025) analysis of browsing data from 900 U.S. adults, Google users clicked a traditional search result in 8% of visits to pages that showed an AI summary, compared with 15% of visits to pages without one, and clicked a link inside the summary itself in about 1% of visits. An organisation can therefore appear in AI Overviews thousands of times and earn nothing from it. Revenue does not come from appearances in an answer. It comes from people who visit, become potential customers and then buy.

SARC therefore gives the Conversion component two jobs. The first is to earn the click. A cited passage has to be complete enough to be quoted, but it should also signal that more value lies on the page behind it: a specific option, a price, a comparison, a tool or a next step that the generated answer cannot provide. This is information scent in Pirolli and Card’s (1999) sense, a cue that tells the reader that following the link will repay the effort. The organisation and its offer should also be named inside the extractable passage, not in a neighbouring sentence the engine may drop (see Section 3.3), because answer-style interfaces tend to separate readers from the sources behind an answer (Shah & Bender, 2022). A reader who does not know whose page the answer came from has no reason to visit it.

The second job is to turn the visitor into a buyer. The reader who clicks through from a generative answer has already read a summary, so the page must extend that answer rather than repeat it (Section 6.1). It must state the offer in concrete terms (Packard & Berger, 2021), place evidence next to the claims that support the decision (Section 4), answer the questions a buyer asks before committing, and end in one clear next step. A visit that does not lead to an enquiry, a sign-up or a purchase is still only visibility.

Being named in an answer may also lead some readers to search for the organisation later, and this can be tracked. SARC does not treat it as a substitute for the click, however. The measure of the framework is not how often a page is cited, but how many cited readers visit and how many of those visitors buy. SARC predicts that passages which signal further value behind the citation earn more clicks without being cited less, and that pages which extend the answer convert those clicks at a higher rate. The measures for testing this are set out in Section 6.7.

6.7 Operationalising Conversion

Conversion can be measured through behavioural data and annotated text features:

• conversion rates for sessions referred from generative engines, compared with other referral sources, where partner organisations can share data;
• the presence and order of genre moves on the landing page;
• the concreteness of the language, measured with established concreteness norms;
• the form, position and specificity of directives;
• the proportion of the page that extends rather than repeats the likely generative answer.

Because being cited does not by itself produce revenue (Section 6.6), these measures are organised around the path from citation to purchase: the share of generative answers to a fixed set of commercial questions that cite or name the organisation, recorded by putting the same questions to each engine at fixed intervals, as in Liu et al.’s (2023) audit of generative search engines; the click-through rate from those answers to the page, where analytics can identify visits referred by generative engines; the share of those visitors who become potential customers through an enquiry, sign-up or trial; and the share who buy, together with the revenue they generate. Branded search volume can be tracked as a secondary indicator, but the framework is judged on clicks, customers and revenue, not on appearances alone.

Two propositions follow:

P10. Among readers referred by generative engines, pages that extend the cited answer with new, situation-specific detail convert at higher rates than pages that repeat it.

P11. Concrete, specific language on landing pages is associated with higher conversion among generative-search referrals, consistent with the concreteness effects reported by Packard and Berger (2021).

7. An integrated model and testable propositions

The four SARC components are not independent checklists. They describe successive filters that a text passes through on the way from publication to reader action, and each filter is governed by a different combination of linguistic properties.

7.1 Four filters

The model can be stated as a sequence of conditional probabilities:

1. Retrieval (Search). Given a query, is the passage retrieved into the candidate set? This depends mainly on lexical and semantic alignment and on passage independence.
2. Selection and citation (Authority and Relevance). Given that it is retrieved, is the passage used in the generated answer and cited? This depends on how directly it answers the query, how little irrelevant material it carries and how clearly its claims are attributed.
3. Click-through (Relevance and Authority). Given that it is cited, does the user follow the citation? This depends on whether the cited passage signals that more useful information lies behind it, and on whether the source appears credible.
4. Action (Conversion). Given that the user arrives, do they take a meaningful next step? This depends on whether the page extends the answer, how concretely it speaks to the reader’s situation and how well its directives are placed.

The probability that a piece of content produces an action for a given query is the product of these four conditional probabilities. A weakness at any stage lowers the whole product. This multiplicative structure explains a pattern familiar from practice: content that ranks well but converts poorly, or content that is well written for humans but never retrieved, has failed at a single stage while succeeding at others.

7.2 Interactions between components

The components overlap in the linguistic features they rely on. Some of these overlaps are complementary, and some create tension.

Complementary features. Explicit, self-contained passages serve Search, because they resolve their own references, and Relevance, because they reduce the effort of extraction. Concrete, specific language serves Authority, because it supplies checkable detail, and Conversion, because it signals attention to the reader (Packard & Berger, 2021). Answer-first structure serves Relevance for models and preserves information scent for human readers (Pirolli & Card, 1999).

Tensions. Two tensions stand out. First, complete answers favour citation but may reduce click-through, because the reader is satisfied without visiting. Second, the features that models weigh may differ from those that readers weigh, as the evidence reviewed in Section 4.5 suggests (Wan et al., 2024). A text tuned only for machines may be cited and then fail to persuade the reader who arrives. A text tuned only for readers may persuade well and never be cited.

SARC’s claim is that these tensions can be managed through structure. Passages that answer are written for extraction. The page around them is written for the reader who wants more. The final proposition states this claim in testable form:

P12. Content that scores highly on all four SARC components produces more reader actions per query than content that scores highly on any one or two components alone, and no single component predicts action as well as the four combined.

A further proposition follows from the argument in Section 6.6 that being cited is not enough to produce revenue.

P13. Cited passages that signal further value on the page, such as a price, a comparison, a tool or a next step that the generated answer does not provide, earn a higher click-through rate from generative answers than passages that do not, without being cited less often (Pirolli & Card, 1999; Pew Research Center, 2025).

7.3 Summary of propositions

The table below gathers the thirteen propositions with the theory each draws on and the main measure it would require.

No.	Component	Proposition (short form)	Theoretical basis	Main measure
P1	Search	Expert term plus common paraphrase raises citation	Lexical semantics; IR term weighting	Query overlap; lexical-field coverage
P2	Search	Passage-internal reference resolution raises citation	Cohesion; text comprehension	Proportion of resolved referring expressions
P3	Search	Explicit definitions raise citation, more for long-tail entities	Information structure; long-tail knowledge	Presence of definitional sentence; entity frequency
P4	Authority	Explicit evidential marking raises citation	Evidentiality; attribution	Evidential markers per claim
P5	Authority	Calibrated certainty beats heavy hedging and unsupported boosting	Stance; hedging; appraisal	Hedge–booster ratio; booster–evidence co-occurrence
P6	Authority	Human credibility cues affect models less than people	Source credibility; web credibility heuristics	Model citation vs human trust ratings
P7	Relevance	Answer in first sentence raises citation	Relevance theory; theme and focus	Position of first answer-bearing sentence
P8	Relevance	Less irrelevant material raises citation	Relevance theory; Gricean quantity	Share of query-relevant sentences
P9	Relevance	Stable thematic progression raises citation	Thematic progression; macrostructure	Annotated progression type
P10	Conversion	Extending the answer beats repeating it	Situational relevance; elaboration likelihood	Novelty of page content relative to answer; conversion rate
P11	Conversion	Concrete language raises conversion	Concreteness effects; engagement	Concreteness score; conversion rate
P12	Integrated	All four components together predict action best	Multiplicative filter model	Composite SARC score; actions per query
P13	Conversion	Passages that signal further value earn the click	Information scent; situational relevance	Click-through rate from generative answers; citation rate

The propositions are deliberately narrow. Each names a linguistic feature that can be annotated, a direction of effect and an outcome that can be observed. Several of them, notably P5 and P6, are framed so that existing evidence could just as easily support the opposite result. This is intentional. A framework that cannot fail a test is not yet a theory.

8. Research agenda and methods

The propositions call for a mixed-methods design that moves from observation to controlled experiment to replication. This section outlines a three-phase programme suitable for doctoral research, together with the annotation scheme and analytical approach it requires.

8.1 Phase 1: Corpus analysis of cited and uncited passages

The first phase asks whether the linguistic features named in P1–P9 differ between passages that generative engines cite and comparable passages that they retrieve or could retrieve but do not cite.

Query set. Queries would be drawn from realistic sources rather than invented. Possible sources include anonymised query reports from partner websites, publicly visible question suggestions in search interfaces, and existing benchmarks such as the query set released by Aggarwal et al. (2024). Queries would be stratified by Broder’s (2002) intent types and by domain, since GEO effects vary by domain.

Data collection. Each query would be submitted to several generative engines. Because outputs are not deterministic, each query would be repeated several times and the frequency of each citation recorded. For each cited URL, the passage most likely to have been used would be identified by comparing the generated sentence to the source page, following the close-paraphrase observations of Liu, Zhang and Liang (2023). Comparison passages would be drawn from pages that rank for the same query in traditional search but are not cited.

Annotation. Passages would be annotated for the features listed in Sections 3.6, 4.6 and 5.6. Some features can be computed automatically: lexical overlap, embedding similarity, passage length, readability and concreteness. Others require trained human annotation: evidential marking, hedges and boosters, Toulmin structure, thematic progression and answer position. A coding manual would be developed from the frameworks of Hyland (2005b), Martin and White (2005), Daneš (1974) and Mann and Thompson (1988). Inter-annotator agreement would be reported, and the manual revised until agreement is acceptable.

Analysis. Citation (yes or no, or citation frequency across repeats) would be modelled with mixed-effects regression, with linguistic features as predictors and random effects for query, domain and engine. Controls would include signals of source authority that the text does not determine, such as domain age or link counts where available. This phase establishes associations but not causes.

8.2 Phase 2: Controlled rewriting experiments

The second phase tests causality by changing one SARC feature at a time while holding content constant.

Materials. Passages would be rewritten in matched pairs or sets. For P7, the same answer would be placed first in one version and third in another. For P4, the same claim would appear with and without evidential marking. For P5, versions would vary the balance of hedges and boosters. Each rewrite would be checked by annotators to confirm that only the target feature changed.

Environment. Experiments would run first in a controlled retrieval-augmented pipeline using open models, a fixed retriever and a fixed corpus into which the rewritten passages are inserted. This allows precise measurement of retrieval rank, inclusion in the answer and citation, as in the controlled designs of Aggarwal et al. (2024), Wan et al. (2024) and Xie et al. (2024). Attribution quality can be evaluated with frameworks such as AIS (Rashkin et al., 2023) and the citation metrics of Gao et al. (2023).

Human component. To test P6, the same passages would be rated by human participants for credibility and usefulness. Model preferences and human ratings could then be compared directly to find where they diverge.

8.3 Phase 3: Live replication and conversion

The third phase tests whether controlled effects hold in live generative engines and whether they reach the human reader.

Cross-engine replication. Selected rewrites would be published on partner or purpose-built websites and monitored across ChatGPT search, Perplexity and Google AI Overviews over several months. Live testing is noisy, because engines change and indexing is slow, so this phase would look for consistency of direction rather than precise effect sizes.

Clicks, customers and revenue. P10, P11 and P13 require behavioural data. With partner organisations, sessions referred by generative engines would be identified, and each page would be followed along the path from citation to purchase: how often it is cited, its click-through rate from generative answers, the share of its visitors who become potential customers through an enquiry, sign-up or trial, and the share who buy, together with the revenue they generate. Pages would be compared on whether their cited passages signal further value (P13), how much they extend the cited answer (P10) and how concrete their language is (P11). Where possible, pages would be varied through A/B tests. Where that is not possible, observational comparisons with appropriate controls would be used, and their limits stated.

Integration. P12 would be tested by computing a composite SARC score for each page and comparing its predictive power for reader action with the power of each component alone.

8.4 Expected contributions of the programme

If the propositions hold, the programme would provide a linguistic explanation for GEO findings that are currently described only in engineering terms. If they fail, the failures would be informative in their own right. For example, if P6 is confirmed and model and human judgements of authority diverge sharply, that finding would bear on the reliability of generative search as a source of information, a concern raised by Shah and Bender (2022) and by Liu, Zhang and Liang (2023). Either outcome would give linguistics a clearer role in a field that has so far been studied mainly as an engineering problem.

9. Limitations, ethics and conclusion

9.1 Limitations

The framework has five main limitations.

First, generative engines are moving targets. Retrieval pipelines, ranking signals and generation models change frequently and without notice. Effects observed in one period may not hold in the next. The research design addresses this by grounding propositions in properties of language that are likely to persist, such as the effect–effort trade-off, rather than in the behaviour of a particular system, and by testing in open, controlled pipelines as well as live engines.

Second, commercial systems are opaque. It is rarely possible to know which passage of a cited page was used or why. Phase 1 must therefore infer the source passage from the generated text. This introduces measurement error that should be reported.

Third, the framework has so far been developed for English. Evidentiality, information structure and stance are expressed very differently across languages (Aikhenvald, 2004). A Search or Authority feature that matters in English may have a different form, or no direct equivalent, elsewhere. Extending SARC to other languages is a natural direction for later work.

Fourth, SARC originated in professional practice, and its author has a practitioner’s interest in its success. The research design guards against confirmation bias by stating propositions that could fail, by testing several of them against existing evidence that points the other way (notably P6), and by separating annotation from hypothesis testing where possible.

Fifth, conversion data are difficult to obtain and rarely public. Tests of P10, P11 and P13, and any estimate of revenue, will depend on partner organisations willing to share click, conversion and sales data, and may be limited in scale. Their results should be treated as indicative until replicated.

9.2 Ethics: legibility, not manipulation

Any framework for influencing what AI systems cite raises ethical questions. Kumar and Lakkaraju (2024) and Nestaas et al. (2025) showed that generative engines can be manipulated with text designed to bias their outputs, and Nestaas et al. warned that widespread manipulation could degrade results for everyone. Xie et al. (2024) showed that models can be persuaded by coherent evidence even when it is false. These findings mean that techniques which make text more citable could also make misleading text more citable.

SARC takes a clear position on this. Its aim is to make accurate, useful content more legible to both machines and people, not to deceive either. The framework’s features are, for the most part, features of good writing in the ordinary sense: saying clearly what something is, showing where claims come from, answering the question asked and telling readers what they can do next. Features that work only by exploiting a system’s weaknesses, such as hidden instructions, invisible text or disparagement of competitors, fall outside SARC and would be treated as a separate object of study.

The research programme itself must also be conducted responsibly. Live experiments should not publish false information. Pages created for testing should be accurate and clearly owned. Partner data should be anonymised and used with consent. Findings that reveal ways to manipulate systems should be disclosed responsibly, following the practice of Nestaas et al. (2025), who notified affected parties before publishing.

9.3 Conclusion

Generative search has changed what it means for a text to be found. The unit of competition is now the passage, the act of selection is a judgement made by a language model, and the reader who arrives has already been partly answered. These changes make the properties of language central to visibility in a way they were not when search meant a ranked list of links.

This article has proposed the SARC Content Methodology as a linguistic account of those properties. Search draws on lexical semantics and cohesion to explain retrieval. Authority draws on evidentiality, stance and argument structure to explain attribution. Relevance draws on relevance theory and information structure to explain extraction. Conversion draws on persuasion research and genre analysis to explain what readers do next. Together, the four components form a sequence of filters whose combined effect determines whether a text produces action. Appearing in an AI answer is only the first of these filters. Revenue depends on the later ones: a cited passage that earns the click, and a page that turns the visitor into a buyer.

The framework began as a practitioner’s method. Its value as research depends on whether its propositions survive testing. The thirteen propositions and the three-phase design set out here are intended to make that testing possible. If they hold, they will give a principled explanation for what generative engines reward. If they fail, they will show where the intuitions of practice and the behaviour of these systems part ways. In either case, generative search is a filter that operates on language, and linguistics has much to say about how it works.

References

Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K., & Deshpande, A. (2024). GEO: Generative engine optimization. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (pp. 5–16). ACM. https://doi.org/10.1145/3637528.3671900

Aikhenvald, A. Y. (2004). Evidentiality. Oxford University Press.

Belkin, N. J. (1980). Anomalous states of knowledge as a basis for information retrieval. Canadian Journal of Information Science, 5, 133–143.

Berger, J., & Milkman, K. L. (2012). What makes online content viral? Journal of Marketing Research, 49(2), 192–205. https://doi.org/10.1509/jmr.10.0353

Bhatia, V. K. (1993). Analysing genre: Language use in professional settings. Longman.

Biber, D. (1988). Variation across speech and writing. Cambridge University Press.

Biber, D., & Finegan, E. (1989). Styles of stance in English: Lexical and grammatical marking of evidentiality and affect. Text, 9(1), 93–124. https://doi.org/10.1515/text.1.1989.9.1.93

Broder, A. (2002). A taxonomy of web search. ACM SIGIR Forum, 36(2), 3–10. https://doi.org/10.1145/792550.792552

Chafe, W. L. (1976). Givenness, contrastiveness, definiteness, subjects, topics, and point of view. In C. N. Li (Ed.), Subject and topic (pp. 25–55). Academic Press.

Chafe, W., & Nichols, J. (Eds.). (1986). Evidentiality: The linguistic coding of epistemology. Ablex.

Clark, H. H., & Haviland, S. E. (1977). Comprehension and the given–new contract. In R. O. Freedle (Ed.), Discourse production and comprehension (pp. 1–40). Ablex.

Daneš, F. (1974). Functional sentence perspective and the organization of the text. In F. Daneš (Ed.), Papers on functional sentence perspective (pp. 106–128). Academia.

Deerwester, S., Dumais, S. T., Furnas, G. W., Landauer, T. K., & Harshman, R. (1990). Indexing by latent semantic analysis. Journal of the American Society for Information Science, 41(6), 391–407. https://doi.org/10.1002/(SICI)1097-4571(199009)41:6<391::AID-ASI1>3.0.CO;2-9

Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2019). BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (pp. 4171–4186). Association for Computational Linguistics. https://doi.org/10.18653/v1/N19-1423

Du Bois, J. W. (2007). The stance triangle. In R. Englebretson (Ed.), Stancetaking in discourse: Subjectivity, evaluation, interaction (pp. 139–182). John Benjamins.

Flesch, R. (1948). A new readability yardstick. Journal of Applied Psychology, 32(3), 221–233. https://doi.org/10.1037/h0057532

Fogg, B. J. (2003). Prominence-interpretation theory: Explaining how people assess credibility online. In CHI ’03 Extended Abstracts on Human Factors in Computing Systems (pp. 722–723). ACM. https://doi.org/10.1145/765891.765951

Gao, T., Yen, H., Yu, J., & Chen, D. (2023). Enabling large language models to generate text with citations. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (pp. 6465–6488). Association for Computational Linguistics. https://doi.org/10.18653/v1/2023.emnlp-main.398

Grice, H. P. (1975). Logic and conversation. In P. Cole & J. L. Morgan (Eds.), Syntax and semantics: Vol. 3. Speech acts (pp. 41–58). Academic Press.

Halliday, M. A. K. (1967). Notes on transitivity and theme in English: Part 2. Journal of Linguistics, 3(2), 199–244. https://doi.org/10.1017/S0022226700016613

Halliday, M. A. K., & Hasan, R. (1976). Cohesion in English. Longman.

Halliday, M. A. K., & Matthiessen, C. M. I. M. (2014). Halliday’s introduction to functional grammar (4th ed.). Routledge.

Hovland, C. I., & Weiss, W. (1951). The influence of source credibility on communication effectiveness. Public Opinion Quarterly, 15(4), 635–650. https://doi.org/10.1086/266350

Hunston, S., & Thompson, G. (Eds.). (2000). Evaluation in text: Authorial stance and the construction of discourse. Oxford University Press.

Hyland, K. (1998). Hedging in scientific research articles. John Benjamins.

Hyland, K. (2005a). Metadiscourse: Exploring interaction in writing. Continuum.

Hyland, K. (2005b). Stance and engagement: A model of interaction in academic discourse. Discourse Studies, 7(2), 173–192. https://doi.org/10.1177/1461445605050365

Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., Ishii, E., Bang, Y. J., Madotto, A., & Fung, P. (2023). Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12), Article 248. https://doi.org/10.1145/3571730

Joachims, T., Granka, L., Pan, B., Hembrooke, H., & Gay, G. (2005). Accurately interpreting clickthrough data as implicit feedback. In Proceedings of the 28th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (pp. 154–161). ACM. https://doi.org/10.1145/1076034.1076063

Kandpal, N., Deng, H., Roberts, A., Wallace, E., & Raffel, C. (2023). Large language models struggle to learn long-tail knowledge. In Proceedings of the 40th International Conference on Machine Learning (Vol. 202, pp. 15696–15707). PMLR.

Karpukhin, V., Oğuz, B., Min, S., Lewis, P., Wu, L., Edunov, S., Chen, D., & Yih, W. (2020). Dense passage retrieval for open-domain question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (pp. 6769–6781). Association for Computational Linguistics. https://doi.org/10.18653/v1/2020.emnlp-main.550

Kintsch, W. (1988). The role of knowledge in discourse comprehension: A construction-integration model. Psychological Review, 95(2), 163–182. https://doi.org/10.1037/0033-295X.95.2.163

Kintsch, W., & van Dijk, T. A. (1978). Toward a model of text comprehension and production. Psychological Review, 85(5), 363–394. https://doi.org/10.1037/0033-295X.85.5.363

Krifka, M. (2008). Basic notions of information structure. Acta Linguistica Hungarica, 55(3–4), 243–276. https://doi.org/10.1556/ALing.55.2008.3-4.2

Kumar, A., & Lakkaraju, H. (2024). Manipulating large language models to increase product visibility (arXiv:2404.07981). arXiv. https://doi.org/10.48550/arXiv.2404.07981

Lambrecht, K. (1994). Information structure and sentence form. Cambridge University Press.

Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Küttler, H., Lewis, M., Yih, W., Rocktäschel, T., Riedel, S., & Kiela, D. (2020). Retrieval-augmented generation for knowledge-intensive NLP tasks. In Advances in Neural Information Processing Systems 33 (pp. 9459–9474).

Liu, N. F., Lin, K., Hewitt, J., Paranjape, A., Bevilacqua, M., Petroni, F., & Liang, P. (2024). Lost in the middle: How language models use long contexts. Transactions of the Association for Computational Linguistics, 12, 157–173. https://doi.org/10.1162/tacl_a_00638

Liu, N. F., Zhang, T., & Liang, P. (2023). Evaluating verifiability in generative search engines. In Findings of the Association for Computational Linguistics: EMNLP 2023 (pp. 7001–7025). Association for Computational Linguistics. https://doi.org/10.18653/v1/2023.findings-emnlp.467

Mann, W. C., & Thompson, S. A. (1988). Rhetorical structure theory: Toward a functional theory of text organization. Text, 8(3), 243–281. https://doi.org/10.1515/text.1.1988.8.3.243

Manning, C. D., Raghavan, P., & Schütze, H. (2008). Introduction to information retrieval. Cambridge University Press.

Martin, J. R., & White, P. R. R. (2005). The language of evaluation: Appraisal in English. Palgrave Macmillan.

Metzger, M. J. (2007). Making sense of credibility on the Web: Models for evaluating online information and recommendations for future research. Journal of the American Society for Information Science and Technology, 58(13), 2078–2091. https://doi.org/10.1002/asi.20672

Metzger, M. J., & Flanagin, A. J. (2013). Credibility and trust of information in online environments: The use of cognitive heuristics. Journal of Pragmatics, 59, 210–220. https://doi.org/10.1016/j.pragma.2013.07.012

Metzler, D., Tay, Y., Bahri, D., & Najork, M. (2021). Rethinking search: Making domain experts out of dilettantes. ACM SIGIR Forum, 55(1), Article 13. https://doi.org/10.1145/3476415.3476428

Mikolov, T., Sutskever, I., Chen, K., Corrado, G. S., & Dean, J. (2013). Distributed representations of words and phrases and their compositionality. In Advances in Neural Information Processing Systems 26 (pp. 3111–3119).

Nakano, R., Hilton, J., Balaji, S., Wu, J., Ouyang, L., Kim, C., Hesse, C., Jain, S., Kosaraju, V., Saunders, W., Jiang, X., Cobbe, K., Eloundou, T., Krueger, G., Button, K., Knight, M., Chess, B., & Schulman, J. (2021). WebGPT: Browser-assisted question-answering with human feedback (arXiv:2112.09332). arXiv. https://doi.org/10.48550/arXiv.2112.09332

Nestaas, F., Debenedetti, E., & Tramèr, F. (2025). Adversarial search engine optimization for large language models. In International Conference on Learning Representations (ICLR 2025). https://doi.org/10.48550/arXiv.2406.18382

Nogueira, R., & Cho, K. (2019). Passage re-ranking with BERT (arXiv:1901.04085). arXiv. https://doi.org/10.48550/arXiv.1901.04085

Packard, G., & Berger, J. (2021). How concrete language shapes customer satisfaction. Journal of Consumer Research, 47(5), 787–806. https://doi.org/10.1093/jcr/ucaa038

Petty, R. E., & Cacioppo, J. T. (1986). The elaboration likelihood model of persuasion. Advances in Experimental Social Psychology, 19, 123–205.

Pew Research Center. (2025, July 22). Google users are less likely to click on links when an AI summary appears in the results. https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/

Pirolli, P., & Card, S. (1999). Information foraging. Psychological Review, 106(4), 643–675. https://doi.org/10.1037/0033-295X.106.4.643

Pornpitakpan, C. (2004). The persuasiveness of source credibility: A critical review of five decades’ evidence. Journal of Applied Social Psychology, 34(2), 243–281. https://doi.org/10.1111/j.1559-1816.2004.tb02547.x

Ram, O., Levine, Y., Dalmedigos, I., Muhlgay, D., Shashua, A., Leyton-Brown, K., & Shoham, Y. (2023). In-context retrieval-augmented language models. Transactions of the Association for Computational Linguistics, 11, 1316–1331. https://doi.org/10.1162/tacl_a_00605

Rashkin, H., Nikolaev, V., Lamm, M., Aroyo, L., Collins, M., Das, D., Petrov, S., Tomar, G. S., Turc, I., & Reitter, D. (2023). Measuring attribution in natural language generation models. Computational Linguistics, 49(4), 777–840. https://doi.org/10.1162/coli_a_00486

Reimers, N., & Gurevych, I. (2019). Sentence-BERT: Sentence embeddings using Siamese BERT-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (pp. 3982–3992). Association for Computational Linguistics. https://doi.org/10.18653/v1/D19-1410

Robertson, S., & Zaragoza, H. (2009). The probabilistic relevance framework: BM25 and beyond. Foundations and Trends in Information Retrieval, 3(4), 333–389. https://doi.org/10.1561/1500000019

Salton, G., & Buckley, C. (1988). Term-weighting approaches in automatic text retrieval. Information Processing & Management, 24(5), 513–523. https://doi.org/10.1016/0306-4573(88)90021-0

Saracevic, T. (2007). Relevance: A review of the literature and a framework for thinking on the notion in information science. Part II: Nature and manifestations of relevance. Journal of the American Society for Information Science and Technology, 58(13), 1915–1933. https://doi.org/10.1002/asi.20682

Searle, J. R. (1969). Speech acts: An essay in the philosophy of language. Cambridge University Press.

Shah, C., & Bender, E. M. (2022). Situating search. In Proceedings of the 2022 ACM SIGIR Conference on Human Information Interaction and Retrieval (pp. 221–232). ACM. https://doi.org/10.1145/3498366.3505816

Shi, F., Chen, X., Misra, K., Scales, N., Dohan, D., Chi, E. H., Schärli, N., & Zhou, D. (2023). Large language models can be easily distracted by irrelevant context. In Proceedings of the 40th International Conference on Machine Learning (Vol. 202, pp. 31210–31227). PMLR.

Sperber, D., & Wilson, D. (1995). Relevance: Communication and cognition (2nd ed.). Blackwell.

Sundar, S. S. (2008). The MAIN model: A heuristic approach to understanding technology effects on credibility. In M. J. Metzger & A. J. Flanagin (Eds.), Digital media, youth, and credibility (pp. 73–100). MIT Press.

Swales, J. M. (1990). Genre analysis: English in academic and research settings. Cambridge University Press.

Taylor, R. S. (1968). Question-negotiation and information seeking in libraries. College & Research Libraries, 29(3), 178–194.

Toulmin, S. E. (1958). The uses of argument. Cambridge University Press.

Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., & Polosukhin, I. (2017). Attention is all you need. In Advances in Neural Information Processing Systems 30 (pp. 5998–6008).

Wan, A., Wallace, E., & Klein, D. (2024). What evidence do language models find convincing? In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) (pp. 7468–7484). Association for Computational Linguistics. https://doi.org/10.18653/v1/2024.acl-long.403

Wilson, D., & Sperber, D. (2004). Relevance theory. In L. R. Horn & G. Ward (Eds.), The handbook of pragmatics (pp. 607–632). Blackwell.

Wineburg, S., & McGrew, S. (2019). Lateral reading and the nature of expertise: Reading less and learning more when evaluating digital information. Teachers College Record, 121(11), 1–40. https://doi.org/10.1177/016146811912101102

Xie, J., Zhang, K., Chen, J., Lou, R., & Su, Y. (2024). Adaptive chameleon or stubborn sloth: Revealing the behavior of large language models in knowledge conflicts. In International Conference on Learning Representations (ICLR 2024).