We analyzed 170+ AI-cited sources across 5 content types. Each one follows a different formula.

Methodology: how we ran this research

We want to be transparent about what we did and how, because the findings only matter if the methodology holds up.

Query selection. We selected 40 queries across four categories: how-to questions (e.g., “how to ferment vegetables at home”), what-is questions (e.g., “what is retrieval augmented generation”), comparison queries (e.g., “Shopify vs WooCommerce”), and product/service queries (e.g., “best accounting software for freelancers”). We also analyzed 7 long-form guide articles separately to validate cross-type patterns.

Multi-platform testing. Each query was run through four AI platforms: Perplexity, ChatGPT with Browse, Google AI Overviews, and Google AI Mode. We collected every cited source URL from every response. Most published AEO studies test one or two platforms. We tested four because our earlier client work showed that what gets cited on Perplexity often gets ignored by ChatGPT, and Google AI Mode and AI Overviews behave differently from each other despite both being Google products.

Source analysis. Across all queries, we collected 170+ unique URLs that were cited by at least one AI platform. From these, we identified 14 “champion” pages: sources cited by 3 or 4 AI platforms simultaneously for the same query. These cross-platform champions are the most interesting data points. A page that gets cited by ChatGPT alone might be lucky. A page cited by Perplexity, ChatGPT, AI Mode, and AI Overviews is doing something structurally right. We analyzed each champion in depth: HTML structure, first paragraph construction, section format, bold text patterns, table presence, and entity density.

Response mapping. We captured full responses from all four platforms with complete inline citation mapping, tracking which source was cited at which point in the answer, and what information from that source was used. Perplexity was especially valuable here because it shows numbered source references inline, making it possible to trace exactly which passage came from which URL.

Cross-analysis. We compared citation patterns across content types to identify which structural elements predicted citation for each type. All research was conducted in March 2026.

One note: this research captures a snapshot. AI platforms update their retrieval and citation behavior. The structural patterns we describe have been consistent across three months of testing, but they are not permanent laws. They are the current rules of the game.

Finding 1: How-to articles and the answer capsule

When we ran how-to queries through Perplexity, a formula emerged immediately. 10 out of 10 Perplexity answers to how-to questions started with the same structure: “You [verb] by [method 1], [method 2], [method 3].” Every time. The sources that matched this “answer capsule” format in their own content were the ones getting cited. Sources that opened with a personal story before reaching the actionable content were skipped.

The most interesting example was fermentedfoodlab.com.

This site was cited by all four platforms we tested: Perplexity (twice), Google AI Mode (twice), Google AI Overviews (twice), and ChatGPT. On paper, it should not have been. The HTML hierarchy is objectively terrible: the site uses all H1 tags. No H2s, no H3s. That is the worst possible heading structure by any SEO standard. The content is 8 years old. And the opening of most articles is a personal story about the author's experience.

Yet it was a 4-platform champion. Why?

When we looked at the section structure, every individual section in FermentedFoodLab's articles is self-contained. Each one starts with a bold keyword label followed by a specific instruction: “Use fresh ingredients. Don't use old, soggy cabbage...” and “Keep everything submerged. If vegetables float above the brine...” Each section makes sense completely on its own, without needing the sections above or below it for context.

Compare this to a source we found that opened with “In my experience as a chef for 15 years, the art of fermentation is deeply personal...” This article contained recipes that were arguably better and more detailed. It was not cited by any of the four AI platforms. Not once.

The difference is not content quality. The difference is extractability. AI platforms do not read an article top to bottom like a human does. They scan for sections that can answer a specific sub-question independently. FermentedFoodLab's bold-label-plus-instruction format creates what we started calling “information islands”: pieces of content that stand alone.

We call this the FermentedFoodLab Paradox. It proves that content structure correctness (proper H2/H3 hierarchy, clean HTML) does not predict AI citation. Self-contained, specific, declarative sections predict AI citation. The two are not the same thing, and confusing them sends you optimizing the wrong elements.

Finding 2: What-is articles and the first-sentence definition

For what-is queries, the pattern was even sharper. Every what-is champion in our dataset puts the definition in the first sentence of the article. Not the second paragraph. Not after an introduction. Sentence one.

The formula is consistent: “[Entity] is [category] that [mechanism].” That is it.

AWS documentation, Conductor's marketing glossary, Weaviate's technical docs. All champions. All place the definition inline in sentence one. ChatGPT and Perplexity both extracted these definitions and cited them with inline references.

NVIDIA published a detailed, well-written explainer on the same topic. It opened with company context and background before reaching the definition. The result: scattered citations for specific technical details, but never cited as the primary definition source. The definition itself was attributed to sources that led with it.

Headings told a similar story. H2 headings that contain the entity name (“How does retrieval augmented generation work?”) were cited. H2 headings with generic labels (“Process overview” or “Technical details”) were not. Across our entire what-is dataset, the correlation was 100%. Every cited H2 section in what-is articles contained the entity name in the heading.

What-is articles also need what we started calling the “mechanism section”: a walkthrough of how the thing works, broken into 3–5 named steps. Benefits need to be standalone subsections, not buried in paragraphs. And one finding that kept repeating: limitations and balanced views increase citation likelihood. Sources that acknowledged drawbacks or limitations alongside benefits were cited more consistently than sources that presented only the positive case.

One detail that stood out: analogies get extracted. NVIDIA's analogy comparing the technology to a specific real-world process was pulled and used by ChatGPT. GeeksForGeeks' restaurant analogy for a database concept was extracted the same way. Across our dataset, when a source contained a clear analogy, that analogy appeared in the AI's response 7 out of 9 times. Analogies are citation magnets for what-is content.

Finding 3: Comparison articles and the verdict

This was the finding that surprised us most in its consistency.

ChatGPT always starts its comparison answers with a verdict. Always. The format is: “[X] is best for [use case]. [Y] is best for [use case].” Before any analysis, before any feature breakdown, before any context. Verdict first.

We tested 10 comparison queries. All 10 ChatGPT responses followed this pattern. All 10.

The sources that get cited first in comparison answers are the ones that have a TL;DR verdict at the top of their page. In one comparison query, aiproductivity.com was cited as inline source [3] and fortune.com as inline source [1]. Both had clear verdict statements in their opening paragraphs. Sources with thorough, balanced analysis but no verdict statement appeared later in the citations or not at all.

Tables versus prose was the most binary finding in our entire research. Comparison tables are extracted. Prose comparisons are never cited. Across all 10 comparison queries, every time ChatGPT presented a feature comparison, it came from a table in the source material. We found zero instances of ChatGPT constructing a comparison from prose paragraphs, even when those paragraphs contained the same information as a competing source's table.

“Choose X if... / Choose Y if...” decision frameworks are quoted nearly verbatim when present. This is different from the verdict. The verdict says who wins. The decision framework says when each option wins. Both get cited, but through different mechanisms. The verdict gets cited at the top of the response. The decision framework gets cited in the recommendation section.

One pattern we did not expect: ChatGPT identifies what we call the “philosophical difference” in every comparison. For Shopify vs WooCommerce, the philosophical difference is convenience versus ownership. For HubSpot vs Salesforce, it is simplicity versus configurability. Sources that name this core trade-off explicitly (“The real choice here is between convenience and ownership”) get that exact framing cited. Sources that describe the trade-off without naming it lose the citation to a source that does.

Balanced honesty helps. Avidlyagency published a comparison piece about CRM platforms that included a section titled “When should you pay for Salesforce?” In that section, they honestly recommended the competing platform for specific scenarios. That section was cited by ChatGPT. Recommending your competitor's product, in the right context, is a citation signal. It reads as trustworthy to the model.

Finding 4: Case studies and the headline metric

ChatGPT treats case studies as data sources, not as stories. It extracts four elements: the headline metric, the company name, the mechanism, and the timeframe. Everything else is ignored.

We tracked what ChatGPT pulled from 12 case studies across our dataset. In every instance, the extraction followed the same pattern. Headline metric plus company name plus mechanism plus timeframe. That is the complete citation.

Here is what ChatGPT ignores in case studies: founder stories, detailed methodology breakdowns, customer testimonials, product feature descriptions, and company background. We are not saying these elements are useless for human readers. We are saying AI does not cite them. The gap between what makes a case study compelling to a human and what makes it citable by AI is enormous.

One example illustrates this well. A Calendly case study detailed the company's growth story across multiple phases. ChatGPT extracted three things: “grew to 10 million users,” “product-led growth,” and the three growth phases (individual, team, enterprise adoption). The case study was approximately 3,000 words. ChatGPT used roughly 15 words from it. About 80% of the content was invisible to the citation process.

A finding with practical implications: collection pages beat single case studies for citation volume. A page titled “10 HubSpot case studies with numbers” had each individual case extracted as a separate citation. One page generated multiple citations in a single response, each pointing to a different case on the same URL. A standalone case study generates, at most, one citation.

The most reliable citation trigger in case studies is the one-line takeaway. When the source includes an explicit summary statement (“Key result: 340% increase in conversion rate within 90 days”), that line gets cited verbatim. When the source does not include one, ChatGPT writes its own summary. And when ChatGPT writes its own summary of your case study, it sometimes gets the numbers or the mechanism wrong. The source that controls its own one-line takeaway controls how AI represents its results.

Finding 5: Product roundups and named downsides

For “best X” and product roundup queries, ChatGPT consistently produces ranked lists with “Best for: [audience]” labels. That format is non-negotiable from the AI's perspective. It will reorganize any source's content into that structure.

Price is always included when the source provides it. 8 out of 9 sources with pricing data had their prices extracted and included in ChatGPT's response. The one exception was a source where pricing was on a separate page linked from the roundup rather than on the roundup page itself.

The finding that defined this category: named downsides per product increase citation likelihood. Netcodesign published a web design agency roundup where each entry included specific downsides (“limited e-commerce functionality,” “higher starting price than competitors”). Those downsides were cited. Pure-positive reviews with no downsides mentioned were skipped in favor of sources that included them.

This inverts normal marketing instincts. For a product roundup to be cited by AI, it needs to say negative things about the products it covers. Not harsh or dismissive, but specific and honest. “The platform is only ~20% of success” from rosieparsons got cited as what we call a “trust citation.” ChatGPT used it to calibrate expectations in its response.

“What actually matters” evaluation framework sections get cited consistently. TGG-accounting published a roundup of accounting software with an upfront section explaining the evaluation criteria they used. ChatGPT cited that framework separately from the individual product evaluations. The criteria section and the product section generated independent citations.

Physical product roundups and SaaS roundups have different extraction patterns. For physical products, specifications tables are what get extracted. For SaaS products, named integrations (which tools it connects with) are the main extracted element. For service-based businesses, lifestyle fit statements (“best for freelancers who handle fewer than 50 clients”) are what AI pulls out. The same content type, but the extracted element shifts depending on what is being compared.

The universal pattern across all 5 types

Across all five content types, several patterns appeared in every single champion page we analyzed.

Self-contained sections. Every champion page (14 out of 14) has sections that make sense without context. You can read any section in isolation and get a complete thought. We started calling these “information islands.” This is the single most consistent predictor of AI citation in our dataset. It does not matter whether the content is a how-to, a what-is, a comparison, a case study, or a product roundup. If the sections are self-contained, citation probability goes up.

Named entities over generic categories. Every champion (14 out of 14) uses specific names instead of generic references. “Salesforce” instead of “enterprise CRM.” “React” instead of “frontend framework.” “Calendly” instead of “scheduling tool.” When a source uses a generic category where a specific name could go, the AI cites a different source that uses the name.

Actionable beats analytical. We tracked 25+ analytical sections across our dataset. Sections about theory, trends, market psychology, or strategic thinking. Zero were cited. Every single cited section across all five content types tells the reader what to do or what is true in concrete terms. Not one cited section told the reader what to think about.

Bold keyword + explanation. 12 out of 14 champion pages use a pattern where a bolded keyword or phrase is followed by an explanation. “Product-led growth. A go-to-market strategy where the product itself drives acquisition...” This pattern appears in how-to articles (FermentedFoodLab), what-is articles (AWS documentation), case studies, and roundups. It seems to serve as a parsing anchor for AI models, marking where a new discrete piece of information begins.

Tables beat prose for any comparison. Whenever information is presented in both table and prose format across competing sources, the table version gets cited. 100% of the time in our dataset. This held for feature comparisons, pricing comparisons, and specification comparisons. Not once did a prose comparison get cited when a table-format version existed on a competing page.

What surprised us: things that do not matter

Several things we expected to matter turned out to have zero correlation with AI citation in our dataset.

HTML hierarchy does not predict citation. FermentedFoodLab uses all H1 tags. It is a 4-platform champion. We found no correlation between correct H2/H3 nesting and citation rate. Pages with perfect semantic HTML were cited at the same rate as pages with flat or broken heading structures, as long as the content within sections was self-contained.

Content age does not matter for evergreen topics. Content from 2015 was cited alongside content from 2025 for topics that do not have a strong time component. An 8-year-old fermentation guide was cited more than a 6-month-old one. For time-sensitive queries (market data, pricing, software versions), recency mattered. For everything else, it did not.

Content length has no correlation. We saw citation champions at 800 words and at 4,000+ words. Short content was not penalized. Long content was not rewarded. The cited content was the content that had extractable, self-contained sections, regardless of total word count.

Answer-first placement is not required. The common AEO advice to “put the answer in the first paragraph” is an oversimplification. FermentedFoodLab's actual answer content starts on the third screen. The definition in some what-is champions is in the first sentence, yes, but in how-to champions, the actionable content can be several scrolls deep. What matters is not where the content is on the page, but whether it is structured as a self-contained unit when the AI reaches it.

Schema markup type has no effect. All 14 champion pages had only basic CMS-generated schema. No FAQPage markup. No HowTo structured data. No custom schema of any kind. We are not saying schema markup is useless for other purposes. We are saying that in our dataset of 170+ sources, advanced schema markup types did not correlate with citation rate.

FAQ sections are invisible to ChatGPT. This one deserves emphasis. Our dataset contained 39 FAQ sections across analyzed pages. ChatGPT cited zero of them. Not one FAQ section, out of 39, was cited. FAQ sections appear to be treated as supplementary rather than primary content. The questions-and-answers format that was built specifically for Google's featured snippets does not translate to AI citation.

What this means for brands publishing content

The mistake we see most often is treating “content optimization for AI” as a single activity. It is not. Each content type has its own citation formula, and applying the wrong one means you are optimizing for nothing.

If you are publishing a comparison page and it does not have a verdict in the first paragraph plus a comparison table, AI will skip it. It does not matter how good your analysis is. It does not matter how balanced your perspective is. Without a verdict and a table, that comparison page is structurally invisible to AI citation.

If you are publishing a case study, the headline metric matters more than the story. You can write a beautiful narrative about your client's transformation. ChatGPT will extract the number, the company name, the mechanism, and the timeframe. If those four elements are not clearly present and easy to extract, the case study generates zero citations.

If you are publishing a product roundup and every entry is positive, you are signaling to AI models that your content is promotional rather than evaluative. Named downsides are not a weakness in your content. They are a citation trigger.

The content types your brand publishes determine which structural patterns you need. A B2B SaaS company publishing comparison pages needs verdict-first formatting and feature tables. An agency publishing case studies needs headline metrics and one-line takeaways. An e-commerce brand publishing product roundups needs honest downsides and specifications tables.

Understanding which type of content you are publishing, and what structural pattern that type requires, is the difference between being cited and being invisible. It is not one formula. It is five.

Get your AI visibility snapshot

Not sure how AI sees your brand right now? Our AI Visibility Report tests 10 real customer queries in ChatGPT, audits your homepage structure, and gives you 10 prioritized recommendations ranked by impact and difficulty. One-time. No subscription. €80.

Get a Far & Wide AI Visibility Report →