1. Introduction
The working paper on the SARC Content Methodology sets out what the methodology claims and why those claims follow from linguistics and information retrieval. This page covers the practical side. It describes the checks I run on a draft for each of the four letters (Search, Authority, Relevance and Conversion), the changes those checks lead to, one published page checked against them, and the results recorded on client and own-site work. The paper is the theory. This page is the procedure.
The procedure needs its own page because a method is only testable if it can be applied the same way twice. If SARC existed only as a set of principles, two writers could both claim to follow it and produce very different pages, and no study could tell whether the method or the writer produced the result. Writing the checks down as a protocol is what turns professional judgement into something a PhD can examine, and something an industry team can adopt without me in the room.
The reason the method exists at all is that generative search changed what a page has to do. Retrieval-augmented systems fetch passages from an index and condition their answer on them (Lewis et al., 2020; Ram et al., 2023). Generative engines then present a single synthesised answer with a small number of citations instead of a ranked list of links, which is the shift Metzler et al. (2021) anticipated when they proposed replacing the index-and-rank model with models that answer directly. The effect on readers is measurable. In Pew Research Center’s (2025) analysis of real browsing data, Google users clicked a traditional result in 8% of visits when an AI summary appeared, against 15% when none did, and they clicked a link inside the summary itself in about 1% of visits.
A page now has two readers. The first is a retrieval system that splits the page into passages, scores them against a query, and decides whether any of them is worth quoting. The second is a person who may meet the page only as a cited fragment inside someone else’s answer, and who may never visit it. SARC is a way of writing for both readers at once. Search and Relevance mostly serve the first reader, and Authority and Conversion mostly serve the second, but every check below affects both.
A note on status. The checks are written as a working protocol drawn from my practice. Where a threshold or rule of thumb is my own convention rather than something the literature establishes, it is marked [CONFIRM] so that I can revise it before the page goes live. Client figures are taken only from results I have already reported, and any date or figure still to be checked is marked [DATE] or [VERIFY]. Every citation on this page is to work published between 2020 and 2026.
2. Where SARC sits in the writing workflow
SARC is not a drafting style. It is an editing pass applied to a complete draft, and it sits inside a six-stage workflow.
Stage 1: Brief. The brief names the page’s audience, the single action the page should lead to, and the question set: the five to ten natural-language questions the page must answer. The question set is the most important output of the brief, because every later check is scored against it.
Stage 2: Research. I collect the facts the page will carry: product specifications, prices, process steps, figures from the client’s own data, and outside sources for any claim the client cannot evidence. Each fact is logged with its source at this stage, because the Authority checks later depend on knowing where every claim came from.
Stage 3: Draft. The first draft is written for a human reader with no SARC rules applied. This is deliberate. Drafting to a checklist tends to produce stiff, keyword-shaped prose, and the checks work better as corrections to natural writing than as constraints on it.
Stage 4: SARC pass. The draft is edited letter by letter, always in the same order: Search, then Authority, then Relevance, then Conversion. The order matters. Search decides which questions each section answers. Authority then fills those sections with attributable evidence. Relevance cuts and reorders what the first two passes added, so that each section leads with its answer. Conversion comes last because the next step a reader is offered depends on what the page has by then become.
Every change made in this pass is recorded in a change log with five columns: the letter, the check number (for example S3 or A2), the text before, the text after, and a one-line reason. The log is the bridge between practice and research. It makes it possible to count which checks fire most often, to compare pages edited by different writers, and later to test whether particular kinds of change predict particular outcomes.
Stage 5: Publish. Before publishing I make sure the page is indexable, that its headings are real heading elements rather than styled paragraphs, and that structured data (for example FAQ or product markup) matches the visible text. These are standard technical checks, not part of SARC itself, but a page that cannot be crawled cannot be retrieved.
Stage 6: Measure. The question set from Stage 1 is rerun in the generative engines that matter for the client, at a fixed interval, and the results are recorded alongside search and conversion data. Section 9 describes this in more detail. The measurement stage closes the loop: a check that never seems to change an outcome is a candidate for removal from the protocol.
3. Search: can the page be found for the questions it answers?
The Search pass asks whether each question in the brief has a passage that a retrieval system can match to it. Two kinds of matching are at work. Lexical retrieval scores overlap between the words of a query and the words of a passage. Dense retrieval encodes both as vectors and scores closeness of meaning (Karpukhin et al., 2020). Production systems typically combine the two, so a page has to work for both, and the checks below are designed with that in mind.
The checks
S1. Map each question to a section. For every question in the brief, I identify the section of the draft that answers it. If no section does, I add one. If two sections half-answer the same question, I merge them. The result is a one-to-one map between questions and sections. Longer, conversational questions deserve special attention: Pew Research Center (2025) found that searches phrased as questions or full sentences were much more likely to produce an AI summary than one- or two-word searches, so the long forms are exactly the queries where a page competes to be cited.
S2. Put the query’s key terms in the answering section. The distinctive words of each question (the product category, the location, the regulation, the problem) must appear at least once in the heading or the first sentence of the section that answers it. This protects lexical matching. Dense retrievers are often assumed to have made exact wording irrelevant, but the BEIR benchmark showed that BM25, a purely lexical method, remained a robust baseline across eighteen datasets, and that dense models often underperformed it outside the data they were trained on (Thakur et al., 2021). Exact wording still earns its place. The check stops at one or two natural occurrences, however. Aggarwal et al. (2024) found that keyword stuffing was among the least effective of the visibility strategies they tested in generative engines, so repetition beyond what the sentence needs is removed.
S3. Make each sentence self-contained. I rewrite sentences that cannot be understood out of context. Pronouns and pointing words such as it, this, they and the above are replaced with the noun they refer to whenever the sentence opens a paragraph or carries a key fact. This check follows Chen et al. (2024), who found that indexing a corpus by propositions (short, self-contained statements of a single fact, with references resolved) outperformed passage-level and sentence-level indexing in dense retrieval and improved downstream question answering. A sentence that begins “It also handles this” may be retrieved as a chunk but carries no meaning once separated from its neighbours.
S4. Make the answering passage look like an answer. The first two or three sentences of each section are written in the form a good answer to the question would take: the direct answer first, then the condition or qualification. The reason comes from how dense retrieval works. Gao, Ma, et al. (2023) showed that retrieval improves when a system first generates a hypothetical answer document and then searches for real passages close to it in embedding space. Whether or not a given engine uses that technique, the finding shows that passages shaped like answers sit near answer-shaped representations of a query. A section that opens with background and reaches its answer in the fourth paragraph is, in vector terms, about the background.
S5. Name entities fully and consistently. Brand names, product names, places, standards and people are named in full at their first mention in each section, and always in the same form. Generative models know far less about rare entities than about common ones. Mallen et al. (2023) found that language models struggled with facts about less popular entities and that retrieval helped most in exactly those cases, and Kandpal et al. (2023) showed that a model’s ability to answer a factual question is strongly linked to how many relevant documents it saw in pre-training. Most of my clients are long-tail entities: a regional recycler, a niche software product, a specialist training provider. For them, the page has to carry the facts, because the model will not already have them.
S6. One topic per section, with a heading that names it. Each section covers one question, sits under a heading that names its topic in plain words, and runs to roughly 100 to 300 words [CONFIRM range]. Retrieval systems split pages into chunks, often at headings or at fixed lengths, and a section that covers three topics produces chunks that match none of them well.
Typical changes
In practice the Search pass most often adds a missing section for a question the client is regularly asked but the draft ignored. It also rewrites vague headings such as “Our approach” into headings that name the topic, such as “How secure data destruction works”. It replaces pronouns at the start of paragraphs with the noun they stand for, and it moves the answer sentence from the end of a section to its start. It rarely adds keywords for their own sake. When S2 finds a missing term, the fix is usually a more precise sentence rather than an extra phrase.
Why Search alone is not enough
The Search pass only gets a passage considered. Whether a generative engine quotes it, and whether a reader trusts it, depends on what the passage says and how it supports its claims. That is the job of the next pass.
4. Authority: can every claim be traced to a source?
The Authority pass asks whether each claim on the page can be checked. Generative engines are poor at this on their own account. When Liu et al. (2023) audited four commercial generative search engines, only about half of the generated sentences were fully supported by their citations, and about a quarter of the citations did not support the sentence they were attached to. Gao, Yen, et al. (2023) found the same gap in a controlled benchmark: even strong models often failed to give complete citation support for what they wrote. Pages that make their own evidence explicit make that gap smaller, and they make it easier for an engine to attribute an answer to them rather than to a competitor.
The linguistic basis for this pass is evidentiality and stance: how a text signals where its knowledge comes from and how strongly it commits to a claim. The measurement basis is the attribution standard proposed by Rashkin et al. (2023), under which a statement counts as attributable only if a reader would agree that an identified source supports it.
The checks
A1. Make a claim inventory. I list every factual claim in the draft (figures, comparisons, outcomes, definitions and statements about regulation) and mark each one as sourced, client-evidenced or unsupported. Unsupported claims are either given a source, rewritten as clearly marked opinion, or cut.
A2. Attach the source in the same sentence. Where a claim has a source, the source is named in the sentence that makes the claim, not in a footnote or a links list at the bottom of the page. A retrieval chunk may contain only that sentence. If the attribution sits elsewhere, the chunk carries an unattributed claim.
A3. Prefer specific figures, named sources and direct statements. Where a precise figure or a named source exists, it replaces vague phrasing. “Many businesses” becomes the figure from the survey that measured it, and “experts recommend” becomes the named body that issued the guidance. This check draws on Aggarwal et al. (2024), whose experiments found that adding citations, quotations from relevant sources and statistics were among the most effective ways to increase a source’s visibility in generative engine responses, with gains of up to about 40% on their benchmark. They also found that effects varied by domain, which is why A3 is a preference and not a quota.
A4. Mark the basis of first-party claims. Claims drawn from the client’s own experience are marked as such, with their scope stated: the period, the number of cases and where the figure comes from, rather than “typically” or “usually”. Calibrated hedging is kept. Overstated certainty is reduced. A first-party claim with a stated scope is attributable, while the same claim without one is just an assertion.
A5. Keep facts consistent across the site. The same fact (a price, a turnaround time, a certification, a founding date) must be stated in the same way on every page that mentions it. Xie et al. (2024) found that language models are highly receptive to coherent external evidence, but also that when the evidence they are given is mixed, they lean towards whichever part agrees with what they already hold. Contradictions across a site give an engine grounds to discount the site or pick the wrong version.
A6. No manufactured authority. SARC excludes hidden text, invented reviews, fabricated statistics and text written to instruct a model rather than inform a reader. This is an ethical rule, but it is also a practical one. Kumar and Lakkaraju (2024) and Nestaas et al. (2025) both showed that text strings added to a page can manipulate which product or source an LLM recommends, and the second study showed that this kind of manipulation leaves every publisher worse off once everyone does it. Engines can be expected to defend against these techniques, while honest evidence does not need defending. There is also a cost to readers. Si et al. (2024) found that people over-rely on AI explanations even when the explanations are wrong, so a page that feeds an engine false confidence harms the people the page is meant to serve.
A caution the protocol takes seriously
The evidence on authority is not one-sided. Wan et al. (2024) found that when language models judged conflicting web evidence, they relied heavily on how relevant a page was to the question and largely ignored features humans treat as signs of credibility, such as references and a neutral tone. I read this as a reason to keep Authority and Relevance as separate checks. Authority markers may do more for verification and for the human reader than for the model’s choice of source. Relevance, covered in the next section, may do more for selection. Whether that division holds is one of the questions the PhD is designed to test.
Typical changes
The Authority pass most often converts general claims into sourced ones and moves sources from the foot of the page into the sentence. It adds the scope to first-party figures and removes superlatives that cannot be evidenced, such as “the leading” or “the best”. It also brings figures into line where two pages on the same site disagree.
5. Relevance: does each passage answer its question, and only its question?
The Relevance pass asks whether each section answers its question early and without distraction. Its linguistic basis is relevance theory, the idea that readers expect a text to repay their attention, and information structure, the ordering of known information before new information within sentences and paragraphs. The retrieval evidence for this pass is unusually direct.
The checks
R1. Answer in the first sentence. The first sentence under each heading answers the question the heading names. Background, history and caveats follow it rather than precede it. Position matters for models as well as people. Liu et al. (2024) found that language models used information best when it appeared at the beginning or end of their input, and performed markedly worse when the relevant passage sat in the middle. A page cannot control where its passage lands in an engine’s context window, but it can control where the answer lands in the passage.
R2. Remove distractors. Sentences that are on the topic but do not help answer the section’s question are cut or moved. These include company history in a how-to section, generic statements about the industry, and repeated calls to action in the middle of explanations. Shi et al. (2023) showed that adding irrelevant context to a problem substantially reduced the accuracy of large language models. Cuconasu et al. (2024) found something more pointed for retrieval-augmented systems: passages that a retriever ranked highly and that were related to the query, but did not contain the answer, harmed the model’s answers. Near-miss text is the hardest kind of distractor because it looks relevant. R2 is aimed at exactly that.
R3. One intent per page, one question per section. If a page tries to serve both someone comparing providers and someone looking for a definition, it is split, or one intent is moved to a linked page. Relevance is judged against a question, and a page with two audiences answers each of them half as well.
R4. Order given before new. Within sentences, I start with the term the reader brought (usually a word from the query) and end with the new information. “Secure data destruction means…” rather than “Under our certified process, which we have refined over many years, the data on your devices is destroyed…”. This keeps the query term in the prominent opening position and puts the payload in the stressed final position, which is also where readers expect new information.
R5. Prefer concrete over abstract. Abstract wording (“solutions”, “leverage”, “seamless”) is replaced with what actually happens: what is done, to what, how long it takes and what it costs. Concrete language does more than read well. Packard and Berger (2021) found that when customer service agents used more concrete language, customers were more satisfied and spent more. Concrete wording also carries more matchable content for retrieval, because it names the things people search for.
R6. Match the reader’s stage. An informational query gets an explanation, a comparison query gets criteria and differences, and a commercial query gets price, process and next steps. When the draft gives a sales pitch in answer to a “what is” question, the pitch is moved to the end of the section or to a different page.
Why Relevance carries so much weight
If Wan et al. (2024) are right that models weigh relevance to the question above stylistic signs of credibility, then Relevance may be the letter with the strongest single effect on whether a passage is selected. It is also the pass that most often makes a page shorter. In my experience a first draft carries a good deal of material written for the client’s own satisfaction rather than for the reader’s question, and that is the material R2 removes.
Typical changes
The Relevance pass most often moves the answer sentence to the top of a section and cuts introductory paragraphs that restate the heading. It splits sections that answer two questions, replaces abstract service language with concrete process descriptions, and moves promotional sentences out of explanatory sections and into the conversion area of the page.
6. Conversion: does the page lead somewhere, even when it is only quoted?
The Conversion pass asks what a reader can do next, and whether the page still works for the business when only a fragment of it reaches the reader. The second half of that question is new. When a page is cited inside a generative answer, most readers never visit it: in Pew Research Center’s (2025) data, links inside AI summaries were clicked in about 1% of visits. Shah and Bender (2022) argued that answer-style search interfaces cut the reader off from the source and its context. For a business, that means the cited fragment may be the only part of the page a prospective customer ever reads.
The linguistic basis for this pass is persuasion and genre: the conventions readers expect from a product page, a guide or a comparison, and the moves within those genres that lead a reader from understanding to action. Berger et al. (2020) set out the two directions in which text matters in marketing: it reflects the person or firm that produced it, and it affects the person who reads it. Conversion is the letter concerned with the second.
The checks
C1. One primary next step. Each page has one primary action (book a collection, start a trial, request a quote) that matches the intent identified in the brief. Secondary links are allowed, but the primary step is stated in the same words wherever it appears. A page that offers five equal options asks the reader to do the business’s sorting for it.
C2. State the offer concretely. The offer says exactly what the reader gets, what it costs or whether it is free, how long it takes, and what happens after they act. A closing line such as “Get in touch to learn more” is replaced with a sentence that names the step, what the reader sends or receives, and when. This applies the same evidence as R5: Packard and Berger (2021) found that concrete language increased both customer satisfaction and purchase behaviour.
C3. Put proof next to the promise. Each claim that supports the decision (a result, a certification, a turnaround time) is placed next to the action it supports, and carries the attribution required by A2. Proof that sits on a separate testimonials page does not help the reader at the moment of deciding.
C4. Make the citable fragment carry the brand. In the sentences most likely to be quoted (usually the first sentence or two of each section, after the Relevance pass), the business is named, together with what it does and where. A generic definition can be quoted without the reader ever learning who wrote it. A sentence that names the business, what it offers and who it serves carries the brand inside the fragment. This check is the main way SARC links visibility to conversion. If a page is quoted but not visited, the brand still reaches the reader, and that is often the step that leads to a later branded search.
C5. Answer the questions people ask before buying. Common objections and pre-purchase questions (price, contract length, data handling, what happens if something goes wrong) are answered on the page, in their own short sections, each written to the Search and Relevance checks. These sections serve conversion and retrieval at once, because pre-purchase questions are exactly the long, specific queries that generative engines answer.
C6. Keep generated claims about the brand in check. Generative engines sometimes state things about a business that are wrong: an outdated price, a service it does not offer, a location it has left. Ji et al. (2023) survey how widespread this kind of fabrication is in language generation. The Conversion pass therefore includes a check of how engines currently describe the client on the brief’s questions, and any recurring error is answered directly on the page in a plain, attributable sentence. This is the least developed check in the protocol, and an obvious target for the research.
Typical changes
The Conversion pass most often reduces several competing calls to action to one, and rewrites generic buttons and closing lines as concrete offers. It adds a short section answering the most common objection, moves a key proof point from a separate page to sit beside the action, and puts the brand name into the opening sentences of sections that previously spoke only in general terms.
7. Worked example: a published page checked against SARC
This example uses a real, published page: It is the page behind the result reported in Section 8: it reached the first Google result within three days of publishing.
The analysis below covers only features visible on the published page. It maps each feature to the check it satisfies. It does not show the draft before editing, so it is a check of the outcome rather than a record of the changes. The before-and-after record belongs in the box at the end of this section.
| Feature of the published page | Check | Why it matters |
|---|---|---|
| The title is the comparison question itself: | S2, S4 | The query’s key terms sit in the title, and the page is framed as the answer to that query |
| A summary sits at the top of the article | R1 | The answer comes before the detail, in the position models use best |
| The opening states that the two terms are used as synonyms and then says plainly that they are not the same job | R1, R4 | It starts from the reader’s assumption (given) and corrects it (new) |
| Each term is defined in its own section, starting with “What Is (THE MAIN KEYWORD)?” | S1, S6, R3 | One question per section, under a heading that names it |
| The First section lists who does the work | R5, S5 | Concrete roles instead of abstract description, and named entities a searcher may use |
| A “Key Differences at a Glance” comparison table | S4, R6 | A comparison query gets criteria and differences, in a form that can be lifted directly |
| The credential is named in full, and tied to the topic | S5, C1, C4 | The provider’s own offer is named inside the content, so it travels with quoted passages |
| A named author and a visible update date | A4 | The basis and recency of the page are stated |
Which of these features were added in a SARC pass, and which were in the first draft, is the part only the change log can show.
8. Client and own-site results
The results below come from work I have already reported publicly or to clients. They are listed so that a reader can see the range of settings in which the practices behind SARC have been used. They are not evidence that SARC caused these outcomes. None of these projects had a control group. Several came before the method had its current name and its numbered checks, and other factors (domain age, competition, budget, product quality) were not held constant. Establishing causation is the purpose of the PhD, not of this page.
| Client or site | Date | Work | Reported result | Note |
|---|---|---|---|---|
| Wavel AI (AI voice and video generation platform) | Full-time Nov 2022 – Aug 2023; part-time Sep 2023 – Jun 2024 | 100+ blog posts and landing pages | About 3,000 sign-ups every two months; content contributed to $15K in revenue | Dates from my employment record |
| Cybersecurity training provider | Article dated | Blog on the Main Keyword | First Google result within three days of publishing | Date as shown on the article page; see Section 7 |
| European SaaS client (name withheld) | Reported October 2026 | 140+ articles | 35,000+ AI impressions | Cumulative figure as of the reporting date |
| snehamukherjee.info | Four months, reported October 2026 | 15 articles | About 3,000 AI impressions | Own site |
| snehamukherjee.info and written-intelligence.com combined | As of October 2026 | All published articles | About 10,087 AI impressions | Own sites |
| IT Recycling Scotland (MGH Scotland, IT asset disposal and electronics recycling) | December 2024- March 2026 | Topic clusters and conversion-led pages | Organic visitors grew from 0 to 100K | Client Work |
| (public sector) | December 2018 – June 2019 | Public-sector communication system | System still in use | Durability outcome, not a traffic metric |
How to read these figures
Firstly, “AI impressions” needs a precise definition before it can be compared across sites. It should state which platform or tool reported the figure, what counted as an impression (an appearance as a cited source, or a mention of any kind), and over what period. Without that definition the figure is a useful signal in practice but not yet a research measure.
Which of these projects used the full SARC pass, and which used only parts of it, should be recorded against each row. The distinction matters, because a project that used only Search and Relevance cannot show anything about Authority or Conversion.
9. Measuring SARC
Each letter makes a different prediction, so each needs its own measure. The scheme below is how I measure SARC in practice. It is also the starting design for the doctoral study.
| Letter | What it predicts | Practical measure |
|---|---|---|
| Search | The page is retrieved for the questions in its brief | Share of brief questions for which the page appears as a source in each engine; search impressions for those queries |
| Authority | Claims on the page can be verified, and engines attribute answers to it | Share of claims with an in-sentence source (from the A1 inventory); share of engine answers that cite the page when they use its facts |
| Relevance | The cited passage is the one that answers the question | Whether the passage an engine quotes is the section mapped to that question in S1; position of the answer sentence within the section |
| Conversion | Visibility turns into action, even without a click | Primary-action completions; branded search volume; share of engine answers that name the brand |
The engine-side measures follow the approach Liu et al. (2023) used to audit generative search engines: a fixed set of questions is put to each engine, and each answer is recorded together with the sources it cites and whether those sources support it. In practice I run the brief’s question set at a fixed interval, in the same engines, from a clean session, and log each answer, each citation and whether the brand is named. Because engine output varies from run to run, each question is run more than once per interval, and the measure is the share of runs in which the page is cited, not a single yes or no.
Attribution is judged by the standard of Rashkin et al. (2023): a citation counts as supporting only if a reader would agree that the cited page backs the sentence it is attached to. This matters because, as Liu et al. (2023) found, a citation and support are often not the same thing.
For research rather than client reporting, the design adds what client work cannot provide: matched pages. Pages on the same site, of similar age and topic, are split into a group that receives the SARC pass and a group that does not, and the change log from Section 2 records exactly what was done to each treated page. Comparing the groups over the same period separates the effect of the method from site-wide changes. Because the log records each change by check number, the same design can also ask which checks do most of the work, and whether the separation between Authority and Relevance suggested by Wan et al. (2024) holds in live engines.
Limits of the protocol
The protocol has known limits, and stating them is part of making it testable.
First, it is written for informational and commercial pages in English. The Search and Relevance checks rely on how English marks given and new information and on how English queries are phrased, and they may need different rules for other languages.
Second, the engines change faster than any study can run. A check that helps today may stop helping when a retrieval pipeline changes, as the swing from ranked links to generated answers already shows (Metzler et al., 2021; Pew Research Center, 2025). For that reason the protocol ties each check to a linguistic or retrieval mechanism rather than to the observed behaviour of one product. The mechanisms (lexical overlap, self-contained propositions, the position of the answer, attribution) are more stable than any single engine.
Third, several checks rest on my own conventions rather than on published findings: the size of the question set, the section length in S6, and the interval and number of runs in the measurement scheme. They are reasonable starting values, not established ones, and the research should test them rather than assume them.
Fourth, the protocol is applied by a person, and different editors may apply the same check differently. The change log makes this visible but does not remove it. Measuring agreement between two editors who apply SARC to the same draft is one of the first things the PhD needs to do, because a method that two trained people apply differently cannot be the cause of a consistent result.
Finally, SARC deliberately excludes techniques that work by manipulating a model rather than informing a reader (Kumar & Lakkaraju, 2024; Nestaas et al., 2025). That choice may cost visibility in the short term. It is made because the method is meant to be used by businesses that need to be trusted over years, and because a method that relies on exploits stops working as soon as the exploit is closed.
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). Association for Computing Machinery. https://doi.org/10.1145/3637528.3671900
Berger, J., Humphreys, A., Ludwig, S., Moe, W. W., Netzer, O., & Schweidel, D. A. (2020). Uniting the tribes: Using text for marketing insight. Journal of Marketing, 84(1), 1–25. https://doi.org/10.1177/0022242919873106
Chen, T., Wang, H., Chen, S., Yu, W., Ma, K., Zhao, X., Zhang, H., & Yu, D. (2024). Dense X retrieval: What retrieval granularity should we use? In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (pp. 15159–15177). Association for Computational Linguistics. https://doi.org/10.18653/v1/2024.emnlp-main.845
Cuconasu, F., Trappolini, G., Siciliano, F., Filice, S., Campagnano, C., Maarek, Y., Tonellotto, N., & Silvestri, F. (2024). The power of noise: Redefining retrieval for RAG systems. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval (pp. 719–729). Association for Computing Machinery. https://doi.org/10.1145/3626772.3657834
Gao, L., Ma, X., Lin, J., & Callan, J. (2023). Precise zero-shot dense retrieval without relevance labels. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) (pp. 1762–1777). Association for Computational Linguistics. https://doi.org/10.18653/v1/2023.acl-long.99
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
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
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 (Proceedings of Machine Learning Research, 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
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
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). Curran Associates.
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
Mallen, A., Asai, A., Zhong, V., Das, R., Khashabi, D., & Hajishirzi, H. (2023). When not to trust language models: Investigating effectiveness of parametric and non-parametric memories. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) (pp. 9802–9822). Association for Computational Linguistics. https://doi.org/10.18653/v1/2023.acl-long.546
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
Nestaas, F., Debenedetti, E., & Tramèr, F. (2025). Adversarial search engine optimization for large language models. In Proceedings of the Thirteenth International Conference on Learning Representations (ICLR 2025).
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
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/
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
Shah, C., & Bender, E. M. (2022). Situating search. In Proceedings of the 2022 Conference on Human Information Interaction and Retrieval (pp. 221–232). Association for Computing Machinery. 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 (Proceedings of Machine Learning Research, Vol. 202, pp. 31210–31227). PMLR.
Si, C., Goyal, N., Wu, T., Zhao, C., Feng, S., Daumé III, H., & Boyd-Graber, J. (2024). Large language models help humans verify truthfulness – except when they are convincingly wrong. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) (pp. 1459–1474). Association for Computational Linguistics. https://doi.org/10.18653/v1/2024.naacl-long.81
Thakur, N., Reimers, N., Rücklé, A., Srivastava, A., & Gurevych, I. (2021). BEIR: A heterogeneous benchmark for zero-shot evaluation of information retrieval models. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks 1.
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
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 Proceedings of the Twelfth International Conference on Learning Representations (ICLR 2024).