The 2026 AI Citation Visibility Study: How 50 Crypto Protocols Appear in ChatGPT, Perplexity, and Google AI Overviews

Key findings at a glance

Fig. 1 — AI Citation Visibility Score: top 25 protocols (out of 100 points)

Study methodology

The 2026 AI Citation Visibility Study scores 50 crypto protocols across a 100-point framework measuring AI answer presence, citation quality, website retrieval readiness, schema and entity clarity, and documentation quality. The full dataset, including raw audit data and scoring formulas, is in the downloadable spreadsheet.

Protocol selection

The 50-protocol sample spans seven categories, selected using publicly available rankings from DefiLlama, L2Beat, and CoinGecko. Protocols were chosen to represent a range of sizes, ages, and technical approaches within each category rather than to favour high-scoring outcomes.

Fig. 2 — Protocol sample distribution by category (n=50)

Prompt set

Forty prompts were tested across five types, written in natural user language rather than keyword-style queries. Category and comparison prompts were the primary source of citation data.

Testing conditions

Prompts were submitted to ChatGPT, Perplexity, and Google AI Overviews in May 2026 via API and browser automation from a US-based endpoint. ChatGPT returned no source URLs across any tested prompt. Perplexity citations total 69 real entries after removing 38 Cloudflare CAPTCHA and privacy-footer URLs scraped during automated sessions. Google AI Overviews citations total 947 entries. Full testing conditions are in the methodology tab of the downloadable dataset.

Scoring model

Limitations

Sample size. Fifty protocols is sufficient to identify patterns but not to make claims about the sector as a whole. The sample skews toward established, well-resourced protocols.

Single time period. All testing was conducted in May 2026. AI platform behaviour changes as models update. Findings reflect conditions at the time of testing.

Automated testing. Prompts were submitted via API and browser automation rather than manual entry. Results may differ from interactive sessions.

ChatGPT source opacity. ChatGPT returned no source URLs across any tested prompt. Citation quality for ChatGPT reflects mention data only.

Geographic scope. Testing used a US-based endpoint. Google AI Overviews results vary by geography. AIO findings reflect US-endpoint behaviour.

Causation. Statistical patterns reflect correlation within the sample. The study design does not establish causation.

AI presence findings: which protocols appeared and where

Across 40 prompts run on ChatGPT, Perplexity, and Google AI Overviews, 44 of the 50 audited protocols appeared in at least one AI-generated answer on at least one platform. Six protocols were invisible across all three: Pendle, Ethena, Sui, The Graph, Filecoin, and Arweave. Appearing at least once is a low bar. On measures of consistency and citation, the field narrows sharply.

Fig. 3 — Cross-platform protocol mentions: top 20 protocols by prompt count

Ethereum is not competing in the same category as any other protocol

Ethereum's total score of 81.9 is 18 points ahead of Aave in second place. That gap does not close when schema or website readiness is factored in. It reflects the depth and age of Ethereum's third-party content record: academic papers, developer documentation, media coverage, and forum posts accumulated over a decade. Ethereum functions as a different data type in this study. No finding about the broader sample should be read as if Ethereum is a representative data point.

Schema markup was not a reliable predictor of AI citation visibility

Thirteen of the 50 audited protocols have JSON-LD schema on their homepage. The two with the highest schema quality scores are Starknet (12/15) and Ledger (12/15). Starknet ranks 15th overall. Ledger ranks 14th. Both sit below ten protocols with no schema at all. Aave has no schema and ranks 2nd. Solana has no schema and ranks 5th.

Fig. 4 — JSON-LD schema adoption by category: 74% of audited protocols have no schema

Fig. 5 — Schema quality scores for the 13 protocols with JSON-LD

A note on the correlation between schema and mention rates

Protocols with schema in this study averaged 7.0 AI mentions across the prompt set. Protocols without schema averaged 4.6. That gap does not mean schema caused the higher mention rate.

The 13 schema-adopting protocols tend to be among the larger and more established names in the sample: Ethereum, Chainlink, Polygon, Ledger, Safe, Starknet, MetaMask. These protocols carry more accumulated third-party coverage and would likely score higher on AI presence regardless of schema. The Graph is the clearest counter-case: it has JSON-LD schema and registered zero AI presence across all three platforms.

The protocols that appear in AI-generated answers at high rates have something schema cannot produce: an accumulated record of third-party coverage that AI training data and retrieval systems already contain. Schema may help AI systems parse and classify a page. It does not create the content record that determines whether that page gets cited. The more defensible claim for schema in this context is entity resolution: helping AI systems confirm what a protocol is and how it relates to other entities in the space.

Website readiness vs schema quality: four protocol profiles

Plotting readiness scores against schema quality reveals four distinct groups in the sample.

Protocol pages are losing the citation layer to third parties

When AI systems cited sources in this study, they almost never cited official protocol pages.

Fig. 6 — Who controls the AI narrative: domain citation breakdown by platform

The ChatGPT finding deserves attention. It does not mean ChatGPT is not drawing on sources. It means protocol teams have no visibility into which sources are shaping its answers. A response naming Aave as a leading DeFi protocol could be drawing from Aave's own documentation, a CoinGecko aggregator page, a Reddit thread from 2021, or all three simultaneously. There is no way to know from the output.

YouTube and Reddit are the top two cited sources in Google AI Overviews for crypto queries

For a sample of crypto-specific prompts, Google AI Overviews cited YouTube content more than twice as often as it cited official protocol documentation. Reddit at 94 citations reflects the weight Google places on community discussion in this category. For protocol teams whose strategy has focused on official documentation and blog content, the gap is significant. The platforms where their audience discusses the protocol may carry more weight in AI-generated answers than the pages those teams have spent the most time building.

Perplexity's domain profile is different. CoinGecko leads with 6 citations (9% of the Perplexity total), followed by OpenZeppelin and Chainlink's own documentation at 4 citations each. Exchange learning hubs (BingX and KuCoin) also appear at 4 citations each. Two platforms, two distinct retrieval profiles, neither of which favours official protocol pages.

Fig. 7 — Top 15 cited domains: Perplexity (left) and Google AIO (right)

Platform-level differences are significant

The three platforms behave differently enough that cross-platform presence is a more reliable signal than single-platform performance.

ChatGPT produced the broadest mention coverage but provided no citation URLs. Protocol teams have no way of auditing which sources ChatGPT uses. Its mention behaviour appeared to draw heavily on training data, giving older, more established protocols a structural advantage.

Perplexity cited sources consistently, drawing primarily from aggregators and developer documentation. Its 69 real citation records produced a clean domain profile: reference sources rather than media or community content.

Google AI Overviews had the lowest mention rate by category across most protocol types. Only 25% of Layer 1 protocols appeared in AIO results across the prompt set. For Infrastructure protocols, the figure was 33%. AIO was more selective and more likely to rely on video and community content when it did surface a protocol.

Safe (Gnosis Safe) produced the sharpest platform divergence in the dataset: zero Perplexity mentions, one ChatGPT prompt appearance, and a Google AIO score of 3/10, placing it 10th overall. Its schema quality (10.5/15) and website readiness (16.2/20) suggest Google's retrieval placed more weight on structured signals for this protocol than Perplexity did.

Category findings: who leads and who lags

Fig. 8 — Protocol mention rate by category across all platforms

Layer 2 and Staking protocols achieved 100% mention rates on Perplexity. These figures require context: appearing once in a single prompt at position 8 is categorically different from appearing across multiple prompts as a primary recommendation.

Layer 1 protocols had the lowest Google AIO mention rate at 25%, despite having the highest average website readiness score in the study. Ethereum and Solana drove most of the Layer 1 AI presence. Avalanche, Cardano, Polkadot, and Sui collectively contributed very little.

Infrastructure is the weakest category across all three platforms. Chainlink is the only Infrastructure protocol that appears consistently. The Graph, Filecoin, and Arweave registered zero presence across all three platforms despite The Graph scoring 17.5/20 for website readiness.

Strong website readiness, weak AI visibility: the structural gap

Several protocols score highly for website readiness but have minimal AI citation presence. This pattern appears across categories and contradicts the assumption that technical content infrastructure translates directly to AI visibility.

Fig. 9 — Website retrieval readiness by category

Fig. 10 — Average website audit component scores by category

Full study results: all 50 protocols

Fig. 11 — All 50 protocols: website readiness, schema, and documentation scores

The six protocols invisible across all three platforms

Six protocols registered zero AI presence across ChatGPT, Perplexity, and Google AI Overviews. Their structural scores vary considerably. This is not a group united by poor infrastructure.

Fig. 12 — Least visible protocols: platform presence across all three AI systems

Sui has a readiness score of 16.2. Ethena scores 15.0. The Graph scores 17.5. None of them appeared in a single AI-generated answer across the full prompt set. Pendle and Filecoin both have schema. Neither appeared anywhere.

What these six share is not poor infrastructure but limited third-party content coverage relative to the category-level and comparison queries used in this study. The prompts were category-level and comparison-based rather than branded. A prompt asking about Layer 1 networks reliably surfaces Ethereum, Solana, and Avalanche because those protocols have been written about extensively in exactly those terms. Sui is a Layer 1 network that has not yet accumulated that depth of comparative coverage.

What the data means

The central finding of this study is not that AI citation visibility is hard to achieve. It is that for most crypto protocols, it is largely unmanaged.

Many protocols in this sample have well-structured sites, indexable documentation, and reasonable schema implementations. What they have not done, in most cases, is build the kind of distributed, third-party content record that AI retrieval systems treat as the primary signal for what a protocol is and why it matters.

The protocols that rank highest share one characteristic more than any other: they have been written about extensively, by people who are not on their payroll, over a long enough period that the coverage has accumulated across formats and platforms. AI systems retrieve from that accumulated record.

What the structural layer actually does

This study does not show schema as a driver of AI citation visibility. A protocol with strong readiness and indexable documentation is easier for AI systems to parse and classify. Beyond that, the data does not show structural scores producing AI visibility in protocols that lack third-party coverage. Cardano scores 18.8 out of 20 for website readiness. Avalanche scores 17.5. Both have near-zero AI presence across all three platforms. The structural layer was not the constraint. The content corpus was.

The third-party source problem is structural, not incidental

Perplexity cited official protocol pages 1% of the time. Google AI Overviews cited them 4% of the time. Protocol teams that assume their official pages are the authoritative source for AI systems are working from an assumption the data does not support. The authoritative source, in the AI's view, is often CoinGecko, or a QuickNode developer guide, or a Reddit thread from three years ago.

That is not a problem that more schema solves. It requires building the kind of content that third parties cite, share, and link to: original research, developer-focused explainers that third-party tools reference, data that journalists quote.

The 2026 baseline

This is the first year of what will be an annual study. The 2026 findings represent the current state of AI citation visibility in crypto: a sector where a small number of well-established protocols dominate AI-generated answers, where official pages are rarely cited, and where the gap between structural readiness and actual citation is wide enough to suggest that most teams are solving the wrong problem. The study will be repeated in 2027 with the same scoring model and prompt set, making year-on-year comparison possible.

AI citation visibility in crypto is not a solved problem for most protocols, and it is not going to be solved by the current generation of GEO advice. Structured data, page speed, heading structure, and schema validation are hygiene, not strategy. The protocols that appear in AI-generated answers have earned the source layer that makes citation possible: years of third-party coverage, developer adoption, media attention, and community discussion that AI systems can retrieve and synthesise. That layer is built through content other people find worth citing, not through content the protocol publishes about itself. Most protocol teams have spent more time on the latter. The data shows it.

Frequently asked questions