The question every crypto project is asking
ChatGPT doesn’t cite the most capable project in a category. It cites the one it knows most clearly. For most crypto projects, that’s a content and structured data problem, not a product problem.
Most teams notice this first when their project fails to appear in ChatGPT or Perplexity responses for their own category. The AI traffic conversion data makes clear why it matters: AI-referred visitors arrive having largely made their decision. The question is whether your website is the one they’ve been referred to, and most aren’t. Not because of problems with the product or service. But because the model does not have enough clean, structured information to surface them with confidence.
The citation decision gets made inside ChatGPT, Perplexity, or Google AI Overviews long before anyone visits your site. Most crypto projects aren’t even trying to influence this.
Two ways an AI knows something
Most assume “AI citation” is one thing. It isn’t. Two completely different systems are working behind the scenes and conflating them is why most “GEO advice” is only half-applicable.
When ChatGPT generates a recommendation, the core architecture behind these answers is parametric knowledge baked in during training, and retrieval-augmented generation (RAG) that pulls live information at the moment of the query. The two work differently and respond to different interventions, though in practice most platforms blend both. If your project wasn’t well covered in the training data before the model’s knowledge cutoff, there’s no quick fix and the model will get retrained on its own schedule, not yours. That’s largely outside your control. Retrieval is where the opportunity is.
Retrieval involves a live search, which means content published today can appear in a citation tomorrow regardless of the model training data age. Perplexity and Google AI Overviews blend retrieval with training data by default, whereas ChatGPT, Gemini and Claude use training data as default but will supplement with live searches for certain queries.
In practical terms, if your project is relatively new, then Perplexity is where fresh, well-structured content gets traction fastest. ChatGPT rewards the same content but on a slower timeline and has the most volume. Google AI Overviews have the most overlap with traditional search marketing and any foundational SEO you’ve already built will carry over more directly there than anywhere else.
The underlying requirement is the same across all AI models: well-structured, machine-readable content. The platform determines where that investment pays off first.
Why high Google ranking is not the same as AI citation
Most crypto teams getting involved with GEO for the first time assume AI citation vs SEO ranking work the same way: get to page one and the citations follow. The data shows it doesn’t play out that way, and the gap is bigger than most people expect.
- Only 12% of URLs cited by ChatGPT, Perplexity, and Copilot rank in Google’s top 10. This figure is an average across ChatGPT, Gemini, Copilot and Perplexity. [3]
- 90% of ChatGPT citations come from pages ranking in position 21 or lower. [5]
- 80% of LLM citations do not appear anywhere in Google’s top 100. [3]
The big difference is search engines rank pages, while AI answer engines extract answers. A page that ranks well has accumulated authority signals over time. A page that gets cited has a clear, direct, well-structured answer to the question being asked. Those are often not the same page.
For a crypto project that hasn’t dominated traditional search engine rankings, a GEO reframe can be a genuine opportunity. The citation game has different entry requirements: schema and content structure can outweigh years of backlink authority in AI citations.
The distinction between ranking and citation is covered in more depth in the GEO vs SEO guide, which sets out what carries over from traditional search and what needs to be treated as a separate problem.
Of all the projects we’ve worked with, the brands most surprised by this are usually the ones with years of Google optimisation behind them. They assume that traditional SEO work carries over automatically. It largely doesn’t. Google AI Overviews still line up closely with normal rankings. But for ChatGPT, Claude, and Gemini, the correlation is weak and it needs to be addressed as a separate problem. The exception worth noting is Perplexity, where 28.6% of cited URLs do rank in Google’s top 10, significantly higher than the 8% average across ChatGPT, Gemini and Copilot. Perplexity behaves more like a search engine than the other AI platforms.
What drives citation: the three actual signals
The citation mechanism matters. But what you do about it matters more. Three factors determine whether an AI model cites your project: entity salience, topical authority, and answer-first formatting. The three don’t just add up, they compound. Entity salience makes topical authority more legible to the model; answer-first formatting makes both extractable.
Entity salience
Before an AI model can recommend your project, it needs to know what your project is. Not just the name, but a clear understanding of the entity: the category, the function, the problem it solves, and how it relates to adjacent entities in the space. Entity salience describes how well-established your project is as a named entity across the web, and how clearly and consistently it is defined across sources. Building entity authority signals through structured data is the most direct way to address this.
Schema markup gives explicit, structured declarations in your page code that tell the model what you are in a machine-readable format. Rather than leaving AI models to infer your category from body text, schema states it directly using the JSON-LD format recommended by Google. This format draws on Schema.org vocabulary, the standardised framework that both search engines and AI models use to parse entity relationships. When those relationships are declared clearly, the model can include the project in its knowledge graph associations for that topic. Schema can also provide E-E-A-T signals to Google: who you are, what you know, what your product or services are, and why you can be trusted.
Most crypto projects haven’t done even basic schema markup. The ones appearing consistently in AI answers usually have.
Answer-first formatting
According to a 2026 study by Kevin Indig of Growth Memo, 44.2% of LLM citations come from the first 30% of a page’s text. [7] A useful format for AI models is headers as questions, with a direct answer in the first sentence of each section. Content that buries the answer after three lengthy paragraphs of preamble gets passed over for content that leads with the answer. Most crypto content is written for people who already understand the context. The intro typically explains background, history and ecosystem, paragraphs two and three build the argument, and the actual answer lands in paragraph four. AI models read paragraph one and move on.
AI retrieval pipelines split pages into discrete chunks before scoring them for relevance. Each chunk is scored independently, which means a key claim split across a paragraph break, or buried in a long dense block, may score well on neither section it touches. Short, focused paragraphs where the main claim lands in the first sentence give the pipeline a cleaner, higher-scoring unit to work with.
Topical authority
A model learns topic associations by tracking which concepts and named entities appear together repeatedly across authoritative sources. Brand mentions across established publications in the crypto space build this co-occurrence signal faster than internal publishing alone, because the model reads the association from multiple independent sources rather than from a single site claiming its own relevance.
Semantic relevance builds through depth, not breadth. Publishing across fifteen loosely related topics dilutes the signal. Publishing deeply on the two or three subjects your project genuinely owns concentrates it. AI models then begin to associate your name with those topics reliably, which makes citation more likely with queries on those topics.
AI platforms run on Google more than they let on. Search is far from dead.
Research from February 2026, covering 1.3 million citations across 68,313 keywords, found that Google.com accounts for 17.42% of all AI Mode citations, more than YouTube, Facebook, Reddit, Amazon, Indeed, and Zillow combined. [8] That figure has tripled since June 2025, when it stood at 5.7%. When YouTube is included, Google-controlled properties account for roughly one in five AI Mode citations. The AI search dependency on Google’s index is significantly larger than the prevailing GEO narrative acknowledges.
Reddit is often claimed to be a prime source for AI citation. Its share across LLMs dropped roughly 50% between October 2025 and January 2026, falling from 2.02% to 1.01%, according to Conductor research. [9] Reddit does seem to have more concentrated authority when it does appear, moving from a broad-reach source to one for specific intent query types. For crypto projects, presence in recent Reddit community discussions still matters, particularly on Perplexity.
ChatGPT
Training data salience is the baseline. The more a brand appears in the training data, the more likely ChatGPT is to draw on it when answering. However, analysis of over 8,500 prompts found that approximately 31% trigger a live web search, and most users have no idea this is happening. [1]
Queries containing terms like “best”, “reviews”, “2025”, and “comparison” most reliably trigger a web search. Cited pages come disproportionately from position 21 and beyond in Google search for those terms, accounting for 90% of retrieved links. [3][5] ChatGPT’s public stance is that search relies on its own crawler, Microsoft Bing, and licensed publisher data. The reality appears to be broader: Search Engine Land reported in August 2025 that OpenAI had been using SerpAPI to extract Google results in parallel, particularly for real-time topics including news, sports, and finance. [14] Google filed a federal lawsuit against SerpAPI in December 2025, with OpenAI named as a primary customer alongside Meta and Perplexity. Optimising for Google therefore serves ChatGPT retrieval regardless of the Bing partnership.
A service or product page outside the top 20 organic results in Google can still surface in ChatGPT responses if the content is well structured and matches query intent precisely. Citation pattern analysis found that Wikipedia accounts for 47.9% of ChatGPT citations, reflecting the model’s strong weighting toward established reference content. [18] For crypto projects without a Wikipedia presence, the credibility signals that substitute for it need to be present on the page or linked from it: peer-reviewed research, named studies, and authoritative third-party mentions.
ChatGPT citation behaviour is also highly variable. SparkToro’s January 2026 analysis found less than a 1-in-100 chance that ChatGPT gives the same list of brand recommendations if you ask the same question 100 times. [19] AI visibility on ChatGPT is a probability you influence across a large set of queries, not a position you hold.
Perplexity
Perplexity is a retrieval-first model with an extreme recency bias. Structured, current content outperforms older authoritative content, which inverts the traditional SEO logic where established pages with accumulated links tend to win. Perplexity is the most Google-aligned of the AI models, with 28.6% of its cited URLs landing in Google’s top 10, compared to around 8% for ChatGPT and Gemini. [3] Reddit represents 46.7% of Perplexity’s top citations, more than any other platform in its mix. [18] In my experience crypto brands are consistently surprised by how directly community presence translates to Perplexity citation. Perplexity searches on virtually every query by default. Claude is the opposite: it defaults to training data and only retrieves live content when triggered.
Claude
Claude operates differently from the other major platforms in one important respect: it searches less than any of them. A system prompt leaked in May 2025 confirmed that Claude defaults to internal training data and uses live web retrieval only when necessary. [15] For most stable, factual queries it will not search at all. The queries that reliably trigger a live search are those containing explicit recency signals (“2026”, “latest”, “current”), multi-part questions requiring synthesis across sources, and direct requests to search.
When Claude does retrieve, it uses Brave Search as its primary backend. Analysis by Profound found an 86.7% correlation between Brave Search top results and Claude citations – the highest and most predictable correlation of any major AI platform. [16] Ranking on Brave is a more direct path to Claude citation than any other single action. Check current rankings directly at search.brave.com.
For crypto projects, Claude’s low search frequency and Brave backend create a specific optimisation order. Training data representation comes first: the more consistently a protocol appears across quality sources before Claude’s training cutoff, the more likely it is to surface in responses that do not trigger retrieval at all. For newer projects without that footprint, the priority is Brave Search ranking combined with recency signals in headlines and opening paragraphs to push queries into live retrieval mode. Allow ClaudeBot in your robots.txt or Claude’s crawler cannot reach your content regardless of ranking.
Google AI Overviews and AI Mode
Google AI Overviews have the highest overlap with traditional SEO of any AI platform, though that overlap is shrinking. In mid-2025, 76% of AI Overview citations came from pages ranking in the Google top 10. By March 2026, Ahrefs put that figure at 38%, based on analysis of 863,000 keyword SERPs and 4 million AI Overview URLs. [17] The conventional SEO playbook carries over here more than on the LLM platforms, but relying on it alone is no longer sufficient. A page can rank on backlink authority, domain trust, and keyword match without being well structured for AI extraction. That page may earn the ranking but not the citation. A page ranking 9th can displace a page ranking 2nd in AI Overview citations if the content is better structured for extraction.
Semantic completeness correlates at 0.87 with Google AI Overview citation selection, the strongest predictor of any signal measured, based on analysis of 15,847 AI Overview results across 63 industry verticals. [10] A page is semantically complete if it fully answers the question without the reader needing to go anywhere else. A page that leaves key terms undefined, assumes background knowledge, or references other pages for the important detail forces the AI to either patch the answer together from multiple sources or skip the page entirely. A page that defines its terms, answers directly in the opening, and covers the obvious follow-up questions gives the AI something it can lift cleanly and cite with confidence.
By February 2026, 59% of AI Mode citations pointed to traditional organic search result pages, meaning AI Mode now mainly uses Google’s own results, not just local listings. Only 53% of AI Mode sidebar sources match the top 10 organic results. [6] AI Mode draws nearly half its sidebar sources from outside the top 10, so the optimisation target on Google’s platforms now includes ranking well enough to appear when AI Mode cites Google’s own SERPs.
AI citation systems: training vs retrieval
| Goal | System type | What affects it | What doesn’t |
|---|---|---|---|
| Get cited in ChatGPT answers (no browsing) | Training-time | Content quality and authority before training cutoff, robots.txt, GPTBot access | Schema markup, content published after training cutoff |
| Get cited in Perplexity | Retrieval-time | Live crawler access, content structure, page speed, semantic chunking | Historical training data |
| Get cited in ChatGPT browsing responses | Retrieval-time | Page accessibility, structured content, crawlability at query time | GPTBot access, training data |
| Get into Google AI Overviews | Hybrid | Google-Extended crawler access AND training data quality | Either factor in isolation |
| Get cited in Claude responses | Training-time (primarily) | Content authority, ClaudeBot crawler access, cross-domain entity mentions, Brave Search ranking, recency signals in headlines | Real-time page changes |
The table above describes the technical picture. There is a structural and legal dimension behind it that is worth understanding.
The regulatory dimension
Search engines and AI platforms extract value from publishers by scraping their content, training on it, summarising it, then serving that content back to users in a form that often removes the incentive to visit the original source. Zero-click search has been a concern for publishers since AI Overviews launched; the regulatory complaints now reflect how far that concern has escalated. The publisher gets the cost of production. The AI platform gets the traffic.
In February 2026, the European Publishers Council filed a formal antitrust complaint against Google with the European Commission, naming both AI Overviews and AI Mode. The complaint argued that Google was extracting commercial value from publishers’ content to power AI-generated answers while simultaneously diverting users away from the original sources, without authorisation and without compensation. [11]
Chegg, the American educational technology company, reported a 49% decline in non-subscriber traffic in 2024, attributing the drop directly to AI Overviews answering the educational queries that previously drove visits to its platform. [12] Wikipedia recorded a 50% surge in bandwidth consumption from AI bots scraping content for training and citation, while human pageviews fell. [13]
Regardless of how the legal landscape settles, brands building structured, schema-marked content with clear authorship signals right now will be better positioned. If the EU investigation produces remedies that constrain how Google can use publisher content, the value of well-structured content will increase even more.
What this means in practice for a crypto project
AI search optimisation for Web3 projects follows the same logic regardless of platform: give the model something it can identify, categorise, and cite with confidence. The projects already appearing in AI answers tend to share four characteristics.
Every optimisation below assumes AI crawlers can reach your site. Check your robots.txt allows the crawlers for the platforms you are targeting: GPTBot and OAI-SearchBot for ChatGPT, PerplexityBot for Perplexity, ClaudeBot for Claude, and Google-Extended for Google AI Overviews. Blocking any of these, intentionally or through a blanket disallow, removes that platform as a retrieval option regardless of content quality or ranking.
1. Declare the entity explicitly with schema markup. AI platforms use entity resolution to classify sources and assess authority. Google AI Overviews and AI Mode reward content already well-understood by Google’s systems. ChatGPT’s query fan-out retrieval and Perplexity’s recency-weighted indexing both favour sources where the entity is unambiguous. Schema markup that explicitly declares what a project is, what category it belongs to, and what problems it solves is the fastest path to improving entity salience.
2. Build content depth on the 2–3 topics the project actually owns. Ahrefs data shows that organic keyword breadth correlates more strongly with AI visibility than backlinks. A Search Engine Land analysis of 800 domains confirmed it: domains covering an entire topic area comprehensively outperform those optimising for a handful of high-volume terms. For a crypto project, breadth across unrelated topics dilutes topical signals. Depth concentrates them.
Analysis of 15,000 prompts found that ChatGPT generated two or more follow-up queries on 89.6% of searches, expanding the total query surface well beyond the original prompt. 32.9% of cited pages were discovered only through those fan-out queries, not the starting search, and 95% of those fan-out queries had zero monthly search volume by traditional metrics. [2] Content covering the supporting questions around a core topic creates citation paths that standard keyword research will not surface.
Content not updated in over three months is three times more likely to lose citation visibility. [1] Owning 2–3 topics deeply and keeping that content current is a more defensible position than spreading across ten topics with irregular updates.
One condition applies to all of this: the content needs to contain something worth citing. Content produced by running a keyword through a SERP analysis tool and reproducing the structural consensus of top-ranking pages (what we call compiled content) gives AI models nothing unique to extract. The citation moment only exists if the article contains something that cannot be found on the other pages covering the same topic.
3. Structure every piece for extraction. Pages with greater title-to-query alignment and higher Flesch Reading Ease scores appear more frequently among cited pages. [2] Mid-authority sites with domain authority between 20 and 80 accounted for 63.6% of all citations, a larger share than the highest-authority domains. High-authority sites were retrieved more often but cited at a lower rate than every other tier. [2] Have the answer in the first paragraph, use headers as questions, and keep the key claim prominent. A page that answers what a project does and who it is for in the opening paragraph will outperform a longer, better-written page that makes the reader work for it.
4. Build credibility signals beyond your own site. AI engines do not evaluate your content in isolation. SE Ranking’s November 2025 study found that domains with profiles on review platforms including Trustpilot, G2, and Capterra have 3x higher ChatGPT citation rates than those without. Domains with significant Reddit and Quora presence have roughly 4x higher citation rates. [20] For crypto projects, the equivalent signals are protocol listings on DeFiLlama and CoinGecko, community threads on relevant subreddits, and mentions in independent research or audit reports. These are not supplementary to content quality. They are part of what the model uses to decide whether a source is worth citing.
Adding authoritative outbound citations to your top pages is the on-site complement to this. The Princeton/Georgia Tech GEO study found that adding 5 to 8 authoritative outbound links produced a 115% visibility increase for mid-ranked sites in AI-generated responses. [21] Citing named studies, reports, and primary sources signals to the retrieval system that the page has done its research, which in turn makes it easier to cite with confidence.
The commercial case for prioritising this: Seer Interactive’s analysis found that brands cited in AI Overviews earn 35% more organic clicks and 91% more paid clicks than non-cited competitors on the same queries. [22] A citation at position 5 is now worth more commercially than a ranking at position 2 on the same query.
FAQ
Does Google ranking affect AI citation?
It depends on the platform. Google AI Overviews show the strongest correlation with traditional rankings, though that correlation has weakened: 76% of AI Overview citations came from top-10 pages in mid-2025, falling to 38% by March 2026. [17] For ChatGPT and Gemini the correlation is weaker still. 80% of LLM citations do not appear anywhere in Google’s top 100, which means a well-structured page at position 35 can outperform a poorly structured page at position 3 for AI citation purposes. [3]
What is entity salience and how do you build it?
Entity salience is how clearly and consistently an AI model understands what your project is: its category, function, and relationship to other entities in the space. It is built primarily through schema markup, which gives the model a structured declaration of the entity rather than leaving it to infer from body text. Brand mentions across authoritative publications in the space strengthen the signal further by showing the model that independent sources associate your name with the same topic.
Which AI platform is easiest to get cited in quickly?
Perplexity is retrieval-first with a strong recency bias, which means well-structured content published today can appear in citations within days. ChatGPT relies more heavily on training data and responds on a slower timeline. Google AI Overviews benefit from existing SEO investment, so they are the most accessible if a site already ranks in the top 10 for relevant queries.
Does SEO still matter for AI search?
Yes, but the relationship varies by platform and is changing quickly. For Google AI Overviews, the overlap with top-10 organic results dropped from 76% to 38% between mid-2025 and March 2026, so traditional SEO investment carries over less directly than it did a year ago. [17] For ChatGPT and Gemini, the correlation is weak enough that they need to be treated as a separate problem. The projects appearing most consistently in AI answers are not always the ones with the strongest SEO. They are the ones with the clearest entity signals and the most structured content. SEO and AI search optimisation are complementary, not interchangeable.
Summary
The citation decision is made before anyone visits your site. For most crypto projects, the infrastructure to influence that decision is either absent or incomplete. Schema, topical depth, extraction-ready structure, and content that actually contains something worth citing are not optional additions to a content plan. They are the plan.
Want to know where your project stands?
An AI visibility review covers your existing content against what’s being cited in your category and produces a clear report on the gap and what closes it.
Get in touch