The Industry Consensus
There is a piece of advice so universal in Generative Engine Optimization that it has achieved the status of doctrine. It appears in venture capital research memos, in agency playbooks, on the blogs of every major marketing platform, and in the recommendations of solo consultants publishing to LinkedIn. The advice is this: structure your content with bullet points, headers, FAQ schemas, bold text, and scannable formatting to get cited by AI. It is the foundational recommendation of an entire industry. And it is architecturally wrong.
The uniformity is remarkable. Andreessen Horowitz, in a June 2025 article titled "How Generative Engine Optimization (GEO) Rewrites the Rules of Search" authored by partners Zach Cohen and Seema Amble, states that phrases like "in summary" or bullet-point formatting help LLMs extract and reproduce content effectively, and that generative engines prioritize content that is well-organized, easy to parse, and dense with meaning. Semrush, the dominant SEO platform, advises in its November 2025 GEO guide that structured content is more likely to be cited, telling marketers to use headings, bullet points, FAQs, and clear formatting to help Google and AI parse their content, and to include bullet points, lists, or tables to simplify how concepts are presented. HubSpot, in its October 2025 overview of generative engine optimization, cites Chris Long, VP of Marketing at Go Fish Digital, who says his team has noticed how AI-driven search tends to pull in content that is extremely structured via bulleted lists, structured headings, and general listicle-style articles. HubSpot's own analysis identifies content structure emphasis as a key factor, asserting that AI engines favor bullet points, numbered lists, and clear hierarchies that enable easy information extraction.
Neil Patel, arguably the most widely read individual marketer on the internet, argues in his August 2025 GEO guide that the discipline requires smart keyword usage, creating strong E-E-A-T signals, and producing content formats AI can process, with related articles emphasizing answer-ready formatting as a pillar of GEO success. Search Engine Land, in a February 2026 article, correctly identifies the underlying mechanism — that LLMs break content into chunks, convert those chunks into numerical representations, and retrieve the most relevant passages when assembling an answer — but then defaults to the same prescription, advising that GEO places more emphasis on content that is easy to extract and reassemble.
The pattern continues without variation across the rest of the industry. Incremys asserts in its 2026 GEO guide that pages with an H1-H2-H3 hierarchy are 2.8 times more likely to be cited, recommending that marketers highlight key information in bold, limit paragraphs to three or four sentences, and favor lists for groups of items, claiming that eighty percent of cited pages use lists. Reply.com advises that AIs, much like hurried humans, love lists, tables, and summaries, and urges marketers to use bullet points for features or steps and to include summary boxes or key takeaway sections. Directive Consulting instructs marketers to structure content for easy extraction by using clear header hierarchy focusing on H2 and H3, limiting paragraphs to under 120 words, incorporating numbered processes and bullet points for quick facts. MarketingProfs tells content marketers to add a structured TL;DR or FAQ section to each blog post and summarize the main insights in three to five bullet points. StoryChief states that marketers should prioritize concise paragraphs, bullet points, and numbered lists, arguing that generative engines are built on summarization models and they like content that can be summarized in two to three lines. Siege Media advises marketers to prioritize AI-friendly content structure by breaking content into short, scannable sections with clear headings, bullet points, and concise takeaways, and to make formatting machine-readable with clear headers, bullet points, tables, and definition lists. Levy Online prescribes that paragraphs should not exceed four lines and instructs marketers to incorporate numbered lists or bullet points wherever relevant.
This is not a fringe position held by a few contrarian agencies. It is the single most consistent recommendation across the entire GEO landscape. A venture capital firm, the dominant SEO platform, the leading marketing blog, the most famous individual marketer, and at least a dozen specialized agencies and consultancies have all converged on the same foundational claim: format your content with bullet points, headers, and structured markup to get cited by AI. The question is whether that claim survives contact with how AI citation systems actually work.
The Architectural Reality
The GEO industry's formatting consensus rests on an unstated assumption — that the way content enters an AI system is the same as the way content exits it. This assumption is false. There are two distinct architectural layers at play, and the industry has conflated them into one.
The first layer is the input side: parsing and retrieval. When an AI system crawls the web or performs Retrieval-Augmented Generation, it processes content into passages, converts those passages into vector representations, and retrieves the ones most relevant to a user's query. On this side, there is a grain of truth to the formatting consensus. Clear headers can delineate topic boundaries. Lists may be parseable into discrete items. Semantic HTML may assist chunking algorithms. The Search Engine Land article gets the mechanism right when it describes how LLMs break content into chunks and retrieve the most relevant passages when assembling an answer, noting that those retrieved chunks are then synthesized into a response, often without the surrounding context from the original page. To the extent that formatting helps a page get crawled and chunked effectively, it has value on the input side.
But even on the input side, the actual controlled research paints a different picture than the industry consensus suggests. The foundational academic study on GEO, which will be examined in detail in the next section, found that Keyword Stuffing — the closest analogue to structural formatting optimization — performed poorly as a GEO method. The interventions that actually improved visibility in generative engine responses were content-quality methods: adding statistics, citing credible sources, adding quotations from authorities, and improving prose fluency. Formatting changes were not among the top-performing GEO interventions in controlled experiments.
The second layer is the output side: generation and citation. This is where the industry consensus falls apart entirely, because the output side operates under completely different rules than the input side.
Consider the first architectural fact. Major AI assistants are explicitly designed to produce prose output, not formatted output. Their operational instructions direct them to avoid over-formatting responses with elements like bold emphasis, headers, lists, and bullet points, and to use the minimum formatting appropriate to make a response clear and readable. In typical conversations and when answering questions, they are instructed to keep their tone natural and respond in sentences and paragraphs rather than lists or bullet points. For reports, documents, and explanations, they are directed to write in prose and paragraphs without any lists. Web search responses specifically call for natural prose, minimal headers, and concise delivery. This means that even if your source content is a beautifully structured page of bullet points, the AI will dissolve that structure when generating its response. The formatting does not transfer. It cannot transfer. The system is designed to prevent it from transferring.
Consider the second architectural fact. The citation unit in AI systems is the individual sentence, not the section, not the list item, not the page. When an AI cites a source, it wraps individual claims in citation tags tied to specific sentence indices within source documents. The citation maps a document index to a sentence index. This is confirmed across multiple sources. Goodie's research on citation strategy for LLM brand mentions explicitly states that sentence-level citations are tied to a single sentence or claim, improving transparency and accountability. The academic paper "Training Language Models to Generate Text with Citations via Fine-grained Rewards" by Gao and colleagues states directly that they treat each sentence in the response as a unit to compute citation metrics, and their instruction to the model reads: always cite for any factual claim, cite at least one document and at most three documents in each sentence. Perplexity's architecture provides inline numbered citations mapped to specific claims within the response. The Geolyze analysis of how AI engines cite sources confirms that Perplexity uses in-text numbered references that maintain the claim-to-source bond at the sentence level. The architecture is consistent: one citation, one sentence, one claim.
Consider the third architectural fact. Claims in AI responses must be paraphrased, not reproduced. AI citation systems require that claims be stated in the system's own words, never as exact quoted text. Even short phrases from sources must be reworded. The AI is not reproducing your bullet point. It is extracting the factual claim from your bullet point, rewording that claim, and embedding it in a prose sentence within its own response. The formatting is gone. The structure is gone. What survives is the underlying factual assertion, stripped of its original presentation.
Consider the fourth architectural fact. No element of your page's formatting appears in the AI's output. A bullet point from your source does not appear as a bullet point in the response. A header does not appear as a header. Bold text does not appear as bold text. Your FAQ structure does not appear as an FAQ. All of this is dissolved. The AI response is a new document composed in the AI's own voice, with its own structure, citing specific claims from your content at the sentence level.
The synthesis of these four facts is damning for the industry consensus. The GEO industry is optimizing for the input side and calling it optimization for the output side. The question "will the AI read my page more easily?" is being treated as equivalent to "will the AI cite my content in its response?" These are different questions with entirely different answers. A page of dense, well-written prose containing one citable factual claim per sentence may generate more citations than a beautifully formatted page of bullet points in which each bullet is a vague, non-specific generality — because the citation mechanism does not care about the bullet point. It cares about whether the sentence contains a discrete, specific, attributable factual claim that can survive paraphrasing.
Empirical Evidence
The foundational academic research on Generative Engine Optimization comes from a paper titled "GEO: Generative Engine Optimization" by Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan, and Ameet Deshpande, representing researchers from Princeton University, Georgia Tech, The Allen Institute for AI, and IIT Delhi. The paper was published at ACM SIGKDD 2024, the premier international conference on knowledge discovery and data mining. It remains the most rigorous controlled study of what content-level interventions actually improve visibility in generative engine responses.
The findings of this paper align precisely with the architectural argument presented above and diverge sharply from the industry consensus.
Keyword Stuffing, the traditional SEO structural approach and the optimization technique closest in spirit to the formatting advice that dominates GEO guides, performed poorly. Despite being one of the most common SEO optimization techniques in practice, it did not meaningfully improve visibility in generative engine responses. This finding alone should give pause to anyone recommending structural and formatting changes as the primary lever for GEO.
The top-performing methods were universally content-quality methods, not formatting methods. Statistics Addition — the practice of embedding specific quantitative statistics into content — produced up to a forty percent improvement in visibility, making it the single most effective GEO intervention tested. Cite Sources, the practice of adding citations to credible external sources, produced significant improvement. Quotation Addition, embedding credible quotes from authorities, also produced significant improvement. Fluency Optimization, the practice of making prose clearer and better-written, produced a fifteen to thirty percent improvement. The combination of Fluency Optimization and Statistics Addition was the single best-performing pair of any two methods tested. Easy-to-Understand, the practice of simplifying language for clarity, also performed well, producing a fifteen to thirty percent improvement.
Every one of these winning methods operates at the sentence level. Statistics Addition changes what individual sentences say by embedding specific numbers. Cite Sources changes what individual sentences say by adding attributions to authorities. Fluency Optimization changes how individual sentences are written by improving their clarity and readability. None of these methods involve restructuring a page with headers, bullet points, or FAQ schemas.
Perhaps most revealing is what the paper did not test. The nine methods evaluated were Authoritative, Statistics Addition, Keyword Stuffing, Cite Sources, Quotation Addition, Easy-to-Understand, Fluency Optimization, Unique Words, and Technical Terms. None of them are "restructure your content with bullet points and FAQ schemas." The entire category of formatting-based optimization that dominates GEO industry advice was absent from the controlled research — not because the researchers overlooked it, but because their framework focused on content-level interventions that operate on what sentences say and how they say it. The methods that worked in controlled experiments were about the substance and quality of individual claims, not about the visual layout of the page.
The paper's own framing confirms this thesis. The authors write that generative engines value not only content but also information presentation. But critically, "information presentation" in their findings means fluency, clarity, and statistical specificity. It does not mean visual formatting like headers and bullet points. The industry has taken a general principle — that AI values well-presented information — and extrapolated it to a specific prescription about formatting without testing whether that specific prescription actually works. The Princeton and Georgia Tech team did test content-level interventions. The winners were all about sentence-level quality.
What is perhaps more revealing than the academic research is what happens when you interrogate the AI systems themselves. In February 2026, a researcher posed a direct question to ChatGPT: what are your rules for Generative Engine Optimization? The initial response was a textbook reproduction of the industry consensus. ChatGPT listed structured, scannable formatting as a core rule, advising the use of headings, bullet points, tables, step-by-step sections, and FAQs, and asserting that well-structured content increases extraction accuracy and citation likelihood. It recommended answer-first design, question-based headers, and content that works in snippets and converts into bullets. In other words, the AI itself, when asked casually, repeated the same formatting-first advice that dominates the GEO industry.
But the researcher pressed further. The next question was pointed: do bullet points help because the AI model itself prefers them, or because users more clearly define their intent when they format that way? And do you break down both sentences and bullet lists into similar internal representations anyway? ChatGPT's answer shifted significantly. It acknowledged that bullet points help primarily because they reduce ambiguity and isolate intent at the input level, stating explicitly that it is less about the model liking bullets and more about clearer semantic segmentation, lower parsing complexity, and better intent separation. It confirmed that internally, both long sentences and bullet lists are transformed into semantic representations, but that the quality of that transformation depends on how clean the input structure is. This is the input-side benefit the formatting consensus correctly identifies, and it is the only benefit ChatGPT could articulate.
The researcher then pushed to the deepest level, asking ChatGPT to run a controlled internal experiment: given a user with very clear intent expression, does formatting still produce better results? ChatGPT designed a test comparing a cleanly written paragraph against a bullet-pointed version of identical content. Its conclusion was unambiguous. When intent clarity is already high, formatting does not dramatically improve semantic understanding. Both formats decompose into similar semantic cores internally. The difference, ChatGPT said, lies only in structural weighting and output organization. It characterized formatting as a variance reducer, not a comprehension enhancer. It stated that the model's understanding layer is nearly identical whether the input arrives as prose or as bullets.
The researcher then pushed further still, into territory that inverts the GEO consensus entirely. If a user expresses intent clearly and includes explicit weighting of subtopics, would a paragraph actually produce more accurate results than bullet points? ChatGPT's response was remarkable. It acknowledged that bullet-point formatting can introduce what it called artificial salience equalization. Because bullet lists create structural symmetry — each item occupying equivalent visual and token-level space — they generate a pattern-level signal that the listed items are equally important conceptual units. Even when the user specifies explicit weights, ChatGPT explained, the repeated structural framing of a list creates parallel syntax that biases toward parallel conceptual salience. A paragraph, by contrast, can encode hierarchy through language itself: primary concern, secondary role, comparatively minor factor. ChatGPT stated that this linguistic gradient creates stronger salience signals than formatting symmetry, and that in hierarchical reasoning tasks, paragraphs may actually outperform bullet points because they preserve relational fidelity between concepts rather than implying separability. It summarized the core tradeoff precisely: bullet points optimize for coverage completeness and modular clarity, while paragraphs optimize for hierarchical reasoning and nuanced weighting. Formatting, it concluded, is not just cosmetic — it modifies the statistical prior for how concepts relate. Bullets create an independence prior. Paragraphs create a relational prior.
This admission goes beyond the claim that formatting is merely neutral on the output side. It suggests that in certain cases, the GEO industry's foundational recommendation may be actively counterproductive even on the input side. If bullet-point formatting can flatten the salience hierarchy that a well-written paragraph preserves, then advising content creators to convert their prose into lists is not just irrelevant to citation — it may be degrading the very quality of the content that citation systems evaluate.
The researcher then directed the interrogation to its logical endpoint: how does all of this apply when the AI is deciding what to cite? Can formatting influence salience gradients in a way that makes a hierarchical paragraph more likely to be cited than a bullet list containing the same data? ChatGPT's response addressed citation mechanics directly for the first time. It explained that when selecting sources to cite, the model retrieves candidate sources, encodes their content into internal representations, compares those representations against the user's intent and the reasoning trajectory being formed, and selects the source that best supports the claim being generated. Citation, it stated, is not chosen purely by keyword match — it is chosen based on semantic alignment, claim support strength, confidence calibration, and internal reasoning coherence. Formatting enters this process because it affects how claims are segmented, how relationships between concepts are encoded, and how easily a statement can be extracted as a supportable proposition.
ChatGPT then ran the comparison that the GEO industry has never run. Given two sources with identical factual content — one presented as a bullet list of parallel independent claims, the other as a hierarchical paragraph encoding causal relationships and relative importance — which is more likely to be cited? The answer depends on what kind of response the AI is generating. If the AI is constructing a hierarchical explanation involving causal reasoning, ordering, or dependency between concepts, the hierarchical paragraph integrates more smoothly into the reasoning trajectory, which increases what ChatGPT called citation coherence. If the AI is generating an enumerated list of independent facts, the bullet list provides cleaner atomic propositions that are easier to extract. The critical insight is this: citation probability increases when the structure of the source matches the structure of the answer being generated. Since AI systems overwhelmingly generate prose responses — not bullet lists — this structural alignment principle implies that prose sources have a systematic citation advantage that the GEO industry has entirely overlooked.
ChatGPT was careful to qualify this as a second-order effect, not a primary ranking signal. Authority, relevance, claim support strength, and confidence alignment remain the dominant factors in citation selection. Formatting, it said, affects integration, not truth assessment. But in edge cases where two sources are otherwise equal in authority and relevance, structural alignment between source and answer acts as a tiebreaker. The implication is clear: if you are choosing between presenting your content as bullet points or as well-structured prose, and the AI that will cite you generates prose responses, the prose version has a marginal structural advantage — not because the AI aesthetically prefers it, but because it integrates more smoothly into the reasoning trajectory the AI is already building.
The researcher then tested the strongest version of this claim — asking whether, since most users write in prose, AI models would systematically prefer to cite prose sources over bullet-point sources. ChatGPT pushed back. It stated plainly that models do not inherently prefer prose findings over bullet findings simply because users write in prose. Bullet lists are extremely common in training data — documentation, blogs, summaries, product pages, encyclopedic entries — and structured formats are statistically familiar patterns, not anomalies. There is no built-in prose preference bias in a simple sense. The model does not reason that the user wrote prose, so it should cite prose. That kind of surface symmetry bias, ChatGPT explained, is weak compared to semantic alignment. The structural alignment effect is conditional: hierarchical reasoning may slightly favor hierarchical prose sources, while enumerated outputs may slightly favor bullet-form sources. But the dominant variable is the match between reasoning structure and source structure, not between user formatting and source formatting. And overriding all of this, ChatGPT reaffirmed, is the quality of the claims themselves. A bullet list that clearly states a specific causal relationship is stronger than a prose paragraph that vaguely implies one. Clarity dominates format.
This qualification is important because it prevents the argument from overreaching — and in doing so, it sharpens the thesis rather than weakening it. The claim is not that prose is universally superior to bullet points for citation. The claim is that the GEO industry has elevated formatting to the primary recommendation for AI citation, and that emphasis is unfounded. What actually determines citation is the specificity, clarity, and attributability of individual claims. Format is, at best, a marginal tiebreaker in edge cases. The industry has built its entire optimization framework around a variable that even the AI systems themselves, when interrogated, rank below relevance, authority, claim support, and confidence alignment.
The arc of this entire exchange, spanning five escalating rounds of interrogation, is itself evidence. When asked casually about GEO, even an AI system reproduces the industry consensus — structured formatting, bullet points, scannable content. When forced to examine its own processing, it concedes that formatting is an input-side convenience that both formats decompose past. When pressed on whether formatting can distort processing, it acknowledges that bullet points can flatten salience hierarchies that prose preserves. When directed to examine citation mechanics, it reveals that structural alignment between source and answer influences citation probability. And when the strongest version of the resulting claim is tested, the AI qualifies it precisely: format is a conditional, marginal effect subordinate to clarity and relevance. At every level of depth, the interrogation moved further from the industry consensus. The AI never endorsed formatting as a primary driver of citation. Under sustained questioning, it consistently ranked it last.
Anyone can verify the output-side claim through direct observation. Take any bullet-pointed page that gets cited by ChatGPT, Perplexity, or Claude. Look at the actual response the AI produces. The bullet points are gone. The headers are gone. The bold text is gone. What appears is a prose sentence paraphrasing a factual claim from the source, with a citation link or footnote attached. This is universally observable across every major AI platform. Perplexity produces flowing prose paragraphs with inline numbered citation references. ChatGPT with web browsing produces prose paragraphs with sentence-end link citations. Claude with web search produces prose with citation attribution tags. Google AI Overviews produce prose summaries with source links. None of them reproduce the source's bullet points, headers, or formatting in their output. The formatting is dissolved. The factual claims survive.
What Actually Gets Cited
If formatting is not the determining factor in AI citation, then what is? The answer lies in four properties of content that align with how AI citation systems actually operate at the architectural level. These are not formatting recommendations. They are principles of sentence-level factual engineering.
The first property is fact singularity. Each sentence should contain exactly one discrete, citable factual claim. A sentence that bundles three claims together is difficult to cite because the citation system maps one citation to one sentence. A sentence that contains zero factual claims — filler, transitional language, opinion without supporting evidence — will never be cited because there is nothing specific to attribute. Compare a sentence like "Our product offers many benefits including ease of use, great support, and competitive pricing" with "WidgetCo's deployment time averaged 4.2 days across 340 enterprise installations in 2025, according to their published implementation data." The first sentence contains three vague claims bundled together, none of them specific enough to cite individually. The second sentence contains exactly one precise, verifiable, attributable claim. The second sentence is citation-ready. The first is not. This has nothing to do with whether either sentence appears as a bullet point or as prose.
The second property is contextual self-containment. A sentence must make complete sense without the sentences around it. AI systems extract individual passages and reassemble them in a new context — the context of the AI's own response, surrounded by claims drawn from other sources entirely. A sentence that begins with "This means that..." or "As mentioned above..." or "Building on the previous point..." requires context that will not be present when the AI lifts it from your page and places it into its response. Self-contained sentences can be extracted and paraphrased without losing meaning. Context-dependent sentences cannot. This is why dense, self-contained prose often outperforms loosely connected list items that rely on a header for their meaning.
The third property is attribution precision. Sentences that include specific sources, named entities, concrete numbers, or verifiable claims are more likely to be cited because the AI can verify the claim against the source material and attribute it with confidence. Vague assertions without specifics — "many experts agree" or "studies show" — are less useful to citation systems than precise attributions with names, dates, and figures. The Princeton and Georgia Tech research confirmed this directly. Statistics Addition was the single top-performing GEO method. Adding specific, attributable numbers to sentences made those sentences dramatically more likely to appear in generative engine responses. The mechanism is clear: a sentence with a specific statistic from a named source gives the AI something concrete to cite. A sentence without specifics gives it nothing to anchor a citation to.
The fourth property is paraphrase stability. A factual claim must survive being restated in different words. The AI will not reproduce your sentence verbatim. It will paraphrase it. If the meaning of a claim depends on the exact wording — if it relies on a pun, an ambiguity, a nuance that disappears when reworded — it is less likely to survive the citation process intact. Claims that can be stated in multiple equivalent ways without losing their core meaning are more citation-stable. The claim "WidgetCo reduced deployment time by thirty-seven percent in 2025" can be paraphrased as "According to WidgetCo, deployment times fell by over a third last year" without losing its factual content. The meaning is paraphrase-stable. A clever turn of phrase or a carefully constructed rhetorical flourish, however impressive to a human reader, may not survive the paraphrasing step that the AI citation system requires.
These four properties — fact singularity, contextual self-containment, attribution precision, and paraphrase stability — constitute the actual optimization target for content that wants to be cited by AI. They are all sentence-level properties. None of them have anything to do with whether a page uses bullet points or prose, whether paragraphs are three sentences or six, or whether the page includes an FAQ schema. A page of well-crafted prose in which every sentence embodies these four properties will be more citation-ready than a page of bullet points that are vague, context-dependent, unattributed, and paraphrase-fragile. The real work of GEO is not formatting. It is sentence-level factual engineering.
Implications
The practical consequences of the formatting paradox are straightforward, and they require most brands to redirect their GEO efforts.
The first and most immediate action is to audit content for factual density per sentence rather than formatting compliance. Count the number of specific, self-contained, attributable factual claims per hundred words. A page that scores highly on this metric — regardless of whether those claims are presented as bullet points, as prose paragraphs, or as anything else — is more likely to generate AI citations than a page that scores poorly but is impeccably formatted with H2 tags and FAQ schema. Factual density per sentence is a better predictor of citation likelihood than any formatting checklist currently circulating in the GEO industry.
The second action is to reconsider the reflexive instinct to break prose into lists. Dense, well-written prose with one fact per sentence is at least as citation-ready as a bullet-pointed page, and unlike bullet points, prose matches the output format that AI systems actually produce. When an AI cites your content, it will produce a prose sentence. You are writing to be paraphrased, not to be visually scanned by a human reader skimming your page. Optimizing for scannability optimizes for the human reading experience, which may have value for its own reasons, but it does not optimize for AI citation.
The third action is to add statistics and citations to your content aggressively. The Princeton and Georgia Tech research confirms that this is the highest-leverage GEO intervention available. Specific numbers, named sources, and credible citations transform vague assertions into citation-ready claims. Every sentence that contains a specific, attributable statistic is a sentence that an AI system can cite with confidence.
The fourth action is to invest in fluency. The second-highest performing GEO method in the controlled research was Fluency Optimization — making prose clearer and better-written. This is the opposite of the prevailing industry recommendation to break prose into scannable chunks. Better writing, not more formatting, is what the evidence supports.
The fifth and most important action is to stop optimizing for the wrong side of the architecture. Formatting may help your page get parsed on the input side, but it does not make your content more likely to be cited on the output side. The GEO industry has built its foundational recommendation on a confusion between these two layers, and brands that follow that recommendation without question are investing their optimization effort in the wrong place.
A note on nuance is warranted. This paper does not argue that formatting is useless. There may be marginal benefits to structured formatting for crawlability and parsing. Headers that clearly delineate topics may help chunking algorithms identify relevant passages. A well-organized page is easier for both humans and machines to navigate. But the industry has elevated formatting to the primary recommendation for AI citation optimization, and that emphasis is architecturally unfounded. The actual citation mechanism operates at the sentence level, in paraphrased prose, and what determines whether your content gets cited is the specificity, credibility, and self-containment of individual factual claims — not whether those claims were presented as bullet points or as flowing paragraphs. The GEO industry has been solving the wrong problem. The real optimization target has been the sentence all along.