How to get cited by ChatGPT: the playbook for crypto projects

Most crypto teams approach AI visibility the way they approached SEO in 2019: rewrite the homepage, add schema, build links, wait. That work is not wasted, but it is aimed at the wrong surface. An answer engine does not rank your site against other sites. It retrieves a handful of pages that look like they answer the question, reads them, and writes a paragraph. If your project is not on those pages, nothing you do to your own site puts it in the paragraph. What follows is five steps in the order they pay off, each one attached to a figure from a corpus of 39,948 citation events, and then a dated log of us running them on citeos.io between 24 August and 16 September 2026, including the two places they have not worked yet.

How does ChatGPT decide which crypto brands to name?

It retrieves, then it writes. Ask an engine "best crypto exchange for beginners" or "top AEO tools for crypto" and it pulls a small set of pages it judges relevant, reads them, and assembles an answer out of what they say. Every major engine will show you that set. Perplexity numbers its sources. Gemini attaches grounding annotations. ChatGPT lists the pages it browsed. That source list is the whole game, because a brand that appears on none of those pages cannot appear in the answer no matter how good its website is.

Which pages get retrieved depends on the question. A category question pulls third-party comparison articles, listicles and roundups. A brand question pulls registries, directories and reference pages that say who you are. A job-to-be-done question pulls guides and forum threads. These are different page populations, so they need different work, which is why the steps below are not interchangeable.

Here is where the citations in our corpus actually came from, across 72 audited crypto brands between April and August 2026.

Two things in that chart change how a crypto marketing budget should be spent. The first is that Tier-1 crypto press is almost irrelevant to AI citations: CoinDesk, Cointelegraph, The Block and Decrypt together account for 143 events, 0.36% of the total, while a basket of 21 mid-tier outlets accounts for 907 events at 2.27%, cited 6.34 times more often. CryptoSlate alone carried 312. That does not make a Tier-1 placement worthless, it makes it a credibility and audience buy rather than a citation buy, and it should be priced as one. The second is that YouTube is the single largest source domain in the entire corpus at 4.49%, and it appears for all 72 brands. Not one of them, all of them. If nobody has made a video that mentions your project, you are missing the most consistently cited surface in crypto.

The largest slice, the long tail at 79.09%, is ordinary pages: blog posts, comparison articles, vertical roundups, docs, aggregator pages. No single one of them matters. Collectively they are four fifths of everything, and that is the population the next three steps are aimed at.

Step 1: make the engine certain who you are

Do this before anything else, because if the engine does not know which company you are, every other signal gets attached to somebody else. Crypto is unusually bad for this. Token names collide with ordinary English words, protocol names collide with unrelated companies, and half the projects in the space share a name with a startup in another sector.

The fix is not more schema on your own site. It is presence on the registries the engine already trusts, saying the same thing your site says. In practice that means claiming and completing, with an identical name and an identical one-sentence description:

• Crunchbase and LinkedIn for the company entity, with a claimed vanity URL rather than a numeric one.
• SourceForge and Slashdot if you ship software. In our corpus SourceForge is cited for 55 of 72 audited crypto brands and Slashdot for 54, which makes them two of the highest-coverage single sources available to a crypto project, and both are free.
• G2, Capterra, SaaSHub and AlternativeTo for products with buyers who compare.
• CoinGecko and CoinMarketCap if you have a token. CoinGecko is cited for 55 of 72 brands.

Then put every one of those profile URLs into the sameAs array of your site's Organization markup, so the graph closes in both directions. This is a day of work and it is the highest-yield day in the whole playbook.

A listing on a registry the engine already reads outranks anything you can put in your own head tag.

The reason it works is worth stating plainly: your own structured data is evidence you make about yourself, and an engine weighs it accordingly. A third-party registry is evidence somebody else made about you. When the two agree, the entity resolves. When only your own site asserts it, it often does not. In our own log this step moved a wrong answer to a right one in two days.

Step 2: get onto the pages the engine already reads

This is the step crypto teams skip, and it is the one that carries 92.22% of all citations. The mechanic behind it is the strongest single result in our corpus, and it is almost embarrassingly simple:

Corpus figures from the published methodology, rebuilt 25 August 2026 from 105 audited domains. Every figure travels with its sample.

A 9.5% coefficient of variation across 72 businesses that have almost nothing in common is the kind of regularity you can plan against. It also reframes the work. The question stops being "how do I optimise for ChatGPT" and becomes "how many pages that a buyer might read currently mention my project by name". For most crypto projects at launch the answer is somewhere between two and ten, and all of them are owned or paid.

Where to go, in the order the corpus supports it:

• Category roundups and comparison articles in your vertical. Someone has already written "best crypto card", "top perp DEX", "safest crypto wallet". Those are the pages engines retrieve for the questions your buyers ask. Getting added to an existing one is cheaper than ranking a new one.
• YouTube. 4.49% of all citations and present for every single audited brand. A review, a walkthrough, an interview, a comparison. It does not have to be your channel.
• Reddit and community threads. Reddit is 2.19% of citations and appears for 71 of 72 brands. Participate honestly under your own affiliation; a thread that reads as astroturf is worse than nothing.
• Mid-tier crypto media over Tier-1. 2.27% against 0.36%, as above.
• Aggregators and directories beyond the entity set in step 1: software aggregators regenerate their listicles continuously, which makes them persistent rather than one-off.

None of this requires a link. The corpus counts mentions, and an unlinked mention on a page an engine reads does the same work as a linked one. That is genuinely different from SEO and it is the part most agencies have not updated.

Step 3: publish one page that is about your category, with you in it

Own-domain pages are only 7.78% of citations, so this step is smaller than the last one. It is also the only part of the system you fully control, and there is a right shape for it. The corpus measures three ways a page can treat a brand, and the differences are large:

A page where your brand is the subject earns nearly twice what a page earns when your brand is just one name in a ranked list. That points somewhere counterintuitive: the page to write is not "why our exchange is the best". It is a page that answers the buyer's actual question across the whole field, with your project honestly placed inside it. That page is about the category, and about you, at the same time. Engines retrieve it because it answers the question, and they name you because you are in it.

Listicles are 33% of our full corpus and 48% among the most-cited pages we read by hand, so the format works. Two rules make the difference between a page engines cite and a page they ignore:

1. Publish the rubric above the scores. State what you are measuring and how it is weighted before a single number appears. A reader can then disagree with your weights instead of your conclusion, and so can an engine.
2. Put yourself where you honestly land, and name where you lose. A vendor list that ranks itself first and concedes nothing reads as an advertisement. An engine assembling an answer out of listicles has plenty of other listicles to choose from.

Then leave it alone. Ours took 19 days to turn into first-place answers on two engines, which is the lag to plan around.

Step 4: pull the source list every week, because it changes

Every engine tells you which pages it read. Pull that list for each of your prompts and treat it as the work order: if ChatGPT is assembling your category out of four articles you are not in, the job is those four articles, not another blog post on your own site.

The trap is treating that list as stable. It is not. We ran the same prompt on ChatGPT four days apart and the source set turned over almost completely:

Two runs of the same prompt through the same route, four days apart. One source of eight survived. Source: citeOS measurement log, 15 and 19 September 2026.

One page out of eight persisted. Any outreach plan built around "get into these four articles" would have been obsolete before the first reply came back. The lesson is to target the class of page rather than the named instances: software aggregators that regenerate listicles automatically, vertical roundups that get updated, and reference pages that stay put. Chasing this week's four is whack-a-mole, and the coverage law says breadth is what pays anyway.

Step 5: measure weekly, and claim narrowly

Pick four prompts a buyer would actually type, not four keywords. One category question, one comparison, one job-to-be-done, one brand check. Run them on the same engines on the same day each week and record three states per cell: is the brand named, is a URL you own cited, or is it absent. The AI visibility tracking method has the template.

Then be careful about what you say in public. Engine answers vary between runs, and a result you screenshot on Tuesday may be gone on Friday. "Named first by Perplexity on 15 September" is a fact that stays true. "AI ranks us first" is a claim the next run can turn into a lie. Name the engine, name the date, and name the surface: Google's AI Overview and its AI Mode tab gave opposite answers on the same query on the same day in our own log, so "Google says" is not precise enough to be honest.

Case study: 30 days of running this on our own domain

We ran the playbook on citeos.io and logged it. On 24 August 2026 the site was a few months old with 41 live pages, an on-page score of 95 out of 100, structured data everywhere and zero referring domains. Asked "what is citeOS and who is it for", ChatGPT answered with a different company: Citeos, a VINCI Energies street-lighting brand in France. Asked "top AEO tools for crypto", neither ChatGPT nor Perplexity named us. Zero of four prompts carried a citation to our domain on any engine. That is the normal starting position, and it is the one most crypto projects never leave.

Step 1 took two days. A Crunchbase profile went live on 26 August and into the sameAs markup on 26 pages the same day. On the 27 August run ChatGPT answered the brand question with citeos.io, described the product correctly, and volunteered a disambiguation nobody asked for: "there is also a completely different company called Citeos, a VINCI Energies brand". The whole answer rested on one source cited three times, the Crunchbase profile. Not our homepage, not our methodology page. Four months of on-site JSON-LD had not moved that answer; one directory profile did. Perplexity followed the same route on a different registry: a SourceForge listing went live on 2 September, and on 15 September Perplexity's brand answer flipped to citeOS with the SourceForge page as its first citation.

Step 3 took 19 days. On 27 August we published Crypto AEO tools in 2026, compared: six tools scored out of 50 on a rubric published before the scoring, citeOS first at 39, and its own entry naming three places rival tools beat it. Nothing else on the site changed for that query. On 15 September, Perplexity opened its answer with "for crypto-focused AEO, the strongest specialized option in the results is citeOS because it is built specifically for crypto/Web3 and includes a ranked database of crypto publications that AI engines cite", and that page was the only citeos.io URL among its 19 sources. Gemini the same day gave citeOS its own "Crypto-Specific AEO Tools" section with one entry, grounded on the same page. On 16 September, Google's AI Overview in a clean browser listed citeOS first. Both engines quoted the page's own framing back at us, which is what step 3 is for.

Where it has not worked

A case study that only reports wins is not a case study. Three of these are still open:

• ChatGPT's category answer. Still does not name us. It reads generic AEO listicles we are in none of, and as the table in step 4 shows, that set rotates weekly. This is step 2 work and we have not done enough of it.
• Google AI Mode. Absent in the US and in India. Its answers rest on two pages that do not list us, and citeos.io is outside the organic top 10 for the query, so the Overview is drawing on a wider set than AI Mode is. We say "Google AI Overview" and never "Google".
• "Best AEO agency for crypto projects". Zero on every engine. Perplexity names Victoria Olsina, ColdChain, ICODA, MarketAcross and Coinbound, from a Reddit thread and three listicles. citeOS is a product, not an agency, so this is a different brand's problem and this page does not pretend otherwise.

Ten crypto-native rivals also entered the category answers during the window, several on six to thirty referring domains, and at least four are newer than we are. The category is being written right now and nobody has it locked, which is the actual reason to start this quarter rather than next.

What the log cannot tell you is the rate. Three changes with a cause on one side and a date on the other are enough to say the mechanism works. They are not enough to say how often, or how long it holds. That is what the weekly re-run is for.

How the case study was measured

• Four fixed prompts, sent verbatim

  "top AEO tools for crypto"; "best AEO agency for crypto projects"; "crypto AI visibility tool"; "what is citeOS and who is it for". No location in any prompt, no system prompt, nothing prepended. The set has not changed since 24 August 2026.

• Engines and routes

  ChatGPT through the DataForSEO scraper with web search forced and the country set to the United States. Perplexity through its API, model sonar, web search on. Gemini 2.5 Flash with Google grounding, added on 15 September, so it has no baseline. Google AI Overview checked by hand in a clean browser on 15 and 16 September, screenshot on file. Google AI Mode through the MrScraper endpoint on 16 September, countries United States and India.

• Runs

  24 August (baseline), 27 August, 15 September, 19 September, plus the Google check on 16 September. One sample per cell, so these are dated observations rather than rates. Two states read from each answer: whether citeOS is named, and whether a citeos.io URL is in the sources. Gemini returns source domains through a redirect, so we record domains for it rather than URLs.

• Site-side changes in the window

  Crunchbase profile live 26 August. Category page live 27 August. G2 profile live 27 August. SourceForge and Trustpilot listings live 2 September. Entity markup corrected across the site 27 August. No links were built and no press ran.

• The corpus behind the playbook

  39,948 citation events across 72 audited crypto brands, 17,276 pages and 4,395 source domains, April to August 2026. Confidence intervals on the page-treatment figures are bootstrapped. The Tier-2 outlet basket is a judgement call over 21 outlets and we will show it on request. Full weights and formulas are on the methodology page.

• What we do not claim

  That AI ranks citeOS first: it is three engines on one prompt on one day each. That citeOS is the only crypto AEO tool: Crawlux, DABLOCK, Qvery, Visoryn and XanLens exist and ChatGPT names them. That Google names us: the AI Overview does and AI Mode does not, and we say which.

Common questions