I watched a founder open ChatGPT on his phone and ask it to describe his own protocol. The thing he’d been building for eighteen months, the one with the mainnet launch, the audited contracts, the six-figure TVL. The AI did not know it existed. What it knew was a fork from 2023 that had since been deprecated. It described the old tokenomics, the old governance model, the old team. It was polite, confident, and completely wrong.
He laughed. Uncomfortably. Then he refreshed the conversation and tried again, as if maybe the first answer was a fluke. It wasn’t. The second answer was the same, worded slightly differently, equally wrong. The project was invisible to the system that an increasing number of its potential users consult first.
This is not an edge case. If you run a Web3 project and you have never checked what an AI says about it, the odds are good that the answer is at best incomplete and at worst hallucinated. And the longer you leave it, the more that distorted representation becomes the one that compounds.
The problem is not that ChatGPT gets things wrong. The problem is that it gets things consistently wrong, and those wrong answers train future answers.
How do AI engines decide what to say about your project?
Generative engine optimisation, or GEO, is the practice of making your content the source that AI systems use to generate answers. It is distinct from SEO in a critical way. SEO competes for a position on a ranked list. GEO competes for inclusion in a generated response: a response that the AI constructs from scratch each time, drawing on patterns in the data it was trained on and the sources it can retrieve. As Semrush’s research puts it, “the key difference between SEO and GEO is that you aren’t competing to rank at the top of search results in GEO — you’re competing to be part of the final output.”
This distinction matters more for Web3 than for almost any other sector. Traditional search favours authority signals that take years to build: backlinks, domain age, editorial citations. Generative AI favours something slightly different: consistency of description across trusted sources, machine-readable structure, and the presence of specific, verifiable claims. A project can be technically excellent and still invisible to an AI if that information is scattered across Discord threads, unpublished whitepapers, and Medium posts written in different voices. I wrote about how this shift from SEO to GEO plays out in practice if you want the broader picture.
Moz’s analysis found that “ranking well in Google doesn’t mean you’ll be featured in generative answers. AI engines use different inputs and are trained to identify patterns, not positions.” In a client report cited by the same piece, 52% of all AI search traffic was coming from ChatGPT alone. That number will have shifted by now. ChatGPT serves 800+ million weekly users as of late 2025, but the direction is clear. AI search is not a niche channel. It is the channel.
What gets cited and what gets ignored?
The research on which content surfaces in AI responses tells a clear story. One study of 10,000 real-world queries found that pages containing quotes and statistics had 30% to 40% higher visibility in AI outputs compared to content without them. A 30% to 40% gap is not small: the difference between being described accurately and being described partially or not at all.
But there is a structural bias in how these systems surface information that Web3 founders need to understand. Google’s AI Overviews, for instance, cited direct retailers in only 4% of responses, favouring editorial sources that help users research before buying. BrightEdge’s 2026 data shows that ChatGPT took the opposite approach, linking directly to retailers at nine times the rate Google did. The distribution mechanic is not neutral. Different AI systems privilege different source types, and the absence of your project from one system does not predict its absence from another.
The common thread is this: generative AI does not verify. It pattern-matches. The sources it trusts are the ones that describe a thing the same way, repeatedly, across multiple platforms. If your project is described one way on your website, another way on CoinGecko, and not at all on your GitHub readme, the AI averages the available signals, or defaults to the one that appears most frequently. That is often the wrong one. There is more on how AI handles Web3 project descriptions specifically in a post I wrote after running this analysis on several client projects.
Why Web3 projects get hallucinated more than others
There is a specific reason this problem hits Web3 harder than, say, a SaaS company. Blockchain projects generate an enormous volume of publicly accessible information. On-chain data, governance proposals, forum discussions, audit reports, social posts, Dune dashboards: almost none of it is structured for the AI systems that are increasingly used to retrieve it.
The academic literature on hallucination is blunt about the underlying mechanics. One foundational paper describes how large language models “are prone to hallucination, generating plausible yet nonfactual content” and raises “significant concerns over the reliability of LLMs in real-world information retrieval systems.” When an AI cannot find a clear, consistent answer about a project, it does not say “I don’t know.” It constructs the most probable answer from the fragments it has, and it delivers it with the same confidence as a well-sourced one.
The Nielsen Norman Group illustrates the absurdity of this on their own 404 page, which reads: “Did an AI chat send you here? They sometimes get URLs wrong or hallucinate nonexistent NN/g articles.” A usability research firm that has been publishing for decades has to put a warning on its error page because AIs keep inventing articles that don’t exist. If NN/g cannot prevent hallucination through sheer reputation, a Web3 project with six months of content cannot either. The fix is not authority. It is architecture.
The wrong assumption most founders make
Most founders assume that the problem solves itself. They believe that if they build a good product, ship code, and grow their community, the AI will eventually figure it out. This is incorrect for two reasons.
First, generative AI does not crawl the chain. It does not verify that your TVL matches your Discord announcement. It reads the text that describes the project: the website copy, the blog posts, the CoinDesk article from the fork, the outdated Medium series from the last CMO. And it assembles a description from whatever it finds. If the most complete public description of your project is a Reddit thread from 2023, that thread becomes your canonical representation.
Second, the AI’s representation of your project is not static. It compounds. When a user asks a question and the AI gives an answer, that interaction becomes data. Future responses are shaped by past ones. A hallucinated tokenomics section in March becomes the model’s memory in April. By May, the hallucination is embedded in the distribution. The longer you leave it, the more effort it takes to dislodge.
Neil Patel’s breakdown of GEO puts it in practical terms: “instead of only seeking to appear at the top of traditional results, GEO aims to make your content the source that AIs use to generate direct answers to users.” The shift is from ranking to sourcing. You are not trying to be first. You are trying to be cited.
What actually fixes it
The fix has three layers, and only the first is technical.
Explainability. The AI needs to know, in plain language, what your project does, who it serves, and how it works. This sounds obvious, but most Web3 websites are written for people who already understand the problem. The homepage leads with the mechanism, not the purpose. AI systems are terrible at inferring purpose from mechanism. If your hero section says “a zk-rollup leveraging optimistic settlement with cross-domain composability,” the AI will repeat that phrase back, but it will not connect it to a user benefit. Write as though your audience is a smart, curious person who has never heard of your category. That is the AI audience.
Authority. Not domain authority in the SEO sense, but source authority for AI systems. This means being cited by publications that LLMs already trust. But there is a nuance here that the Moz piece captures well: “what matters is that your brand is consistently described the same way, across platforms, by people and publications that LLMs already trust. If you can shape that perception, AI engines will carry it forward.” The consistency is the signal. Not the individual citation, but the pattern of the same description appearing in multiple places the AI considers reliable.
AI-ready content architecture. This is the layer most projects skip entirely. Microsoft’s official guidelines for generative search, cited by Semrush, recommend making your catalogs machine-readable, structuring content to answer real questions, and establishing authority through credible sources and expertise signals. For a Web3 project, that means structured data on tokenomics, verified descriptions on ecosystem registries, consistent copy across all platforms, and a centralised source of truth that AI crawlers can find and trust. The AI visibility system builds across all three of these layers.
HubSpot’s research notes that “AI search users spend 6 minutes per session (vs. seconds on Google)” and that “conversions, by percentage, from LLMs are higher.” Nate Tower, president at Perrill, explains: “people chat with AI and see the software more as a friend, which is one reason why conversions from GEO are higher.” The stakes are not just reputational. If your users are asking an AI whether they should use your protocol, the quality of the AI’s answer determines whether they do.
Where to start
The first step is diagnosis, not content production. Before you can fix what the AI says about your project, you need to know what it currently says, what it gets wrong, and which sources are driving the errors. That is the AI visibility audit: a structured look at how AI systems describe your project, what they distort, and what signals are producing both the accurate and inaccurate parts of the response.
BrightEdge’s data shows that Google still commands over 90% of all search traffic, and AI search platforms combined account for less than 1% of total referrals. Those numbers will change. The same report notes that 90% of digital teams are increasing their SEO investment this year: the highest surge in five years. The market knows something is shifting.
Your project does not need more content. It needs better structure, clearer descriptions, and a deliberate strategy for how AI systems discover and reproduce information about it. The founder who opens ChatGPT and gets an accurate answer about his own protocol is not the one who published the most blog posts. He is the one who made himself un-hallucinable.
Someone has to own that work, because the AI is not going to fix itself.
See also
Leave a Reply