ChatGPT search, Perplexity, Gemini and Google AI Overviews all read structured pages differently from how human readers do. They scan, extract, weigh and cite. Most crypto content is built for human scanning — and gets bypassed by AI extraction layers entirely.
We tracked which content patterns earned citations across 14 client engagements through 2024–2025, on a 22-prompt monitor that ran weekly across the four major AI platforms. The patterns repeat. Six of them, ranked by impact.
Why does paragraph 1 get cited 60% more than paragraph 4?
AI systems extract direct answers from sections by pattern-matching the start of the section against the question. If the first sentence of an H2 block answers the question in 20–30 words, the extractor pulls that sentence with high confidence. If the first sentence introduces context and the answer arrives in paragraph 3, the extractor either pulls the wrong content or skips the section entirely.
The structural rule we run on every long-form piece: every H2 starts with a direct-answer sentence, ≤30 words, that resolves the question on its own. The expansion follows in paragraphs 2–4. Expert nuance, edge cases, exceptions to the rule come last, not first.
This rule is harder to follow than it sounds, because the editorial impulse for human readers is to set the scene before answering. Trained writers will resist it. AI graders reward you for breaking the convention.
What does a named-expert byline actually move?
We ran a 90-day controlled experiment in early 2025 on a single crypto-licensing client’s blog: 12 pieces under “Editorial Team” byline, 12 equivalent pieces under a named expert with full schema.org Person markup and verifiable LinkedIn sameAs.
The named-byline pieces averaged 1.6× the AI-citation rate over a 60-day post-publication tracking window. They also retained ~80% of their citation rate through one Google algorithmic event during that period; the anonymous pieces dropped to ~50% citation share.
The signal compounds in YMYL niches because AI search systems explicitly weight expertise verifiability when picking sources for finance, legal, medical and crypto-adjacent content. The schema.org Person properties that move the needle: jobTitle, worksFor, knowsAbout, alumniOf, and sameAs linking to LinkedIn plus at least one independent third-party profile (a conference talk, a regulator submission, a publication interview).
How specific does a number need to be?
AI tools systematically prefer specific numbers and named sources over generalisations. They cite content that gives them a number they can quote; they skip content that says “most exchanges” or “the industry has seen significant growth”.
Bad: Most crypto exchanges struggle with KYC compliance.
Good: Four of the seven MiCA-licensed CASPs we audited in March 2026 had KYC processes that took longer than 72 hours from sign-up to first deposit, against a 24-hour benchmark in their own marketing copy.
The second version makes a citable claim. Even if the number is wrong by a small margin, the structural move (count out of total, time period, comparison to a stated benchmark) gives the AI tool something to anchor to. Vague claims get skipped because the extractor has nothing to attribute.
This is also where the regulatory-review pass matters. Every specific number you cite needs to be defensible in case a regulator quotes it back. We log sources for every number that appears in client content, kept in a per-client research log, with a provenance link for each.
What is the de-AI editorial pass actually catching?
GPT-3.5 and GPT-4-default writing has predictable lexical signatures that AI graders detect and downweight. The vocabulary stack is consistent enough that it can be caught with a regex: “navigate”, “leverage” (as a verb), “harness”, “delve into”, “in today’s fast-paced”, “in the realm of”, “it’s worth noting”, “comprehensive”, “robust”, “unlock the potential of”, “game-changer”, “revolutionise”.
Our de-AI editorial pass (we call it the Klimakov pass internally) strips these systematically. It also adds back the human signals that AI grader systems use as authenticity markers: em-dashes, contractions, “But/And/So” sentence starters, mixed paragraph lengths (some 1-sentence paragraphs, some 5-sentence paragraphs in the same block), specific dates instead of “recently”, and at least one mild opinion per major section.
The pass takes 30–60 minutes per long-form piece for a trained editor. It is not optional. We have run it on every deliverable since mid-2024 because the alternative — AI-default content — gets demoted in AI graders’ rankings inside three months of publication.
What does FAQPage schema actually do for AI citations?
Most crypto content has FAQ blocks. Most of them are not marked up with FAQPage schema. The schema markup roughly doubles the rate at which AI tools extract individual FAQ items as citation-ready snippets, because the markup tells the extractor where the question ends and the answer begins without it having to guess from heuristics.
The FAQ direct-answer rule applies harder than to body H2s: the answer has to fit in ≤30 words, with the option of a 2–3 sentence depth-expansion paragraph after. The schema only marks up the first 30-word answer; the expansion is for human readers.
Topics that earn the most citations from FAQPage schema in crypto are predictable: cost questions (how much does X cost), timeline questions (how long does X take), qualification questions (do I need Y to do X), comparison questions (which is better, X or Y). Theoretical or definitional FAQs (“what is a blockchain”) earn fewer citations because AI tools have higher-authority sources for those answers already.
How does source citation discipline change AI extraction?
Inline links to primary sources — regulator publications, named studies, original whitepapers, dated press releases — earn substantially more AI citations than text-only “industry reports show” framing. The citation extractor uses inline links as a confidence signal that the content is grounded in something verifiable.
Crypto content has the advantage that primary sources are usually available: ESMA technical advice on MiCA, FCA financial-promotion regime guidance, SEC interpretive releases, EBA ITS on prudential treatment. We link directly to the regulator’s official URL, not to a third-party explainer of the same document. Direct links to authoritative sources get treated differently by extraction layers than two-step links via aggregators.
The discipline that goes with this: every regulatory or numerical claim in client content has to have a primary-source link or an inline data attribution. If we cannot find a primary source for a claim, the claim does not ship.
What does the editorial workflow actually look like?
Every long-form piece goes through five stages.
Brief. Lead editor (Anastasiia for crypto, sometimes a senior editor she trains) writes a brief that includes target queries, source corpus, named expert who will be the byline, regulatory review flags. The brief is reviewed before drafting.
Outline. Outline includes proposed H2 questions in their final form, each with the proposed direct-answer first sentence draft and 1–2 supporting points. This is the structural commitment.
Draft. Senior writer produces the draft from outline. AI tools (Claude, GPT) are used for research and outline support; the draft itself is written by the named editor.
Editorial pass. Second editor runs the de-AI Klimakov pass on the draft, plus a structural verification (every H2 has a 30-word direct answer first; Quick Facts table is intact; FAQ block has 5+ items with schema-ready phrasing).
Compliance review. For YMYL content (most crypto pieces), the regulatory review happens after editorial. Compliance lead checks for banned phrasings, required disclaimers, jurisdiction-specific copy requirements. Average 2–4 days for the review pass.
The process produces somewhere between 4 and 8 long-form pieces per month per writer-editor pair. We staff for that capacity. Writers who think they can produce 12 long-form AEO pieces per month at the level we ship are usually skipping editorial; we have not seen this volume sustained without quality drop in 18 months of running this workflow.
The output is content that earns citations from ChatGPT, Perplexity, Gemini and Google AI Overviews on commercial crypto queries — verifiably, on the 22-prompt monitor we run weekly. If the workflow is interesting to compare against your current process, the discovery call is free, 30 minutes, named lead.