AI visibility tracking for crypto brands: how to track your mentions in ChatGPT and Perplexity, week by week

If you have only ever checked once, start with the 20-minute AI visibility check. It tells you where you stand today. This page is the next step: the same prompts every week, so that when an answer changes you know it changed, roughly when, and what page did it. We run this on citeOS itself, and the log below includes the week one of our own answers flipped.

What AI visibility tracking actually measures

Three states, per answer, per engine. Named means the engine wrote your brand into the answer. Cited means it listed a page on your domain as a source, whether or not it named you. Absent means neither. The three are independent: an engine can name you without citing you, which is the common case, and cite a page of yours without naming you, which happened to us on Perplexity for three weeks.

Tracking these per engine, never pooled, is the whole discipline. Engines read different pages and move at different speeds. Perplexity is retrieval-first and shows its sources, so it moves within days. Gemini grounds on Google's index, so anything Google already ranks shows up there first. ChatGPT blends training data with a short list of trusted domains and is the slowest to change. An average across the three hides every one of those facts.

The second number is why the log needs a column for other people's domains. In our corpus, 92.22% of citations landed on pages the brand did not own. The page that moves your answer is almost never yours, so the log has to record whose it is.

Build the fixed prompt set, then leave it alone

Twenty prompts is the size we use for a paid audit. Four is enough to start, and four is what we track on our own brand. Whatever the count, the rule is the same: write them the way a buyer types, fix them, and never edit them. Adding a prompt later is fine. Rewording one destroys comparability, which is the only thing the log is for.

Four families cover a crypto brand:

1. Category

   "best crypto exchange for beginners", "top AEO tools for crypto", "safest self-custody wallet". The question a buyer asks before they know your name. This is where being absent costs the most.

2. Comparison

   "Coinbase vs Kraken for a first buy", "citeOS vs Crawlux". Asked by buyers who already have a shortlist. Engines answer these from comparison pages, so the source column matters most here.

3. Trust

   "is Bitget safe", "is this exchange licensed in the US". The prompt your deposits ride on. Engines read help centres, security pages and regulators for these, and they read Reddit.

4. Brand

   "what is citeOS and who is it for". Asked by people who have already heard of you. The only family where a wrong answer is worse than no answer, and the one that exposes entity collisions.

For citeOS the fixed set is four prompts, one per family, chosen on 24 August 2026 and unchanged since: top AEO tools for crypto, best AEO agency for crypto projects, crypto AI visibility tool, and what is citeOS and who is it for. They are quoted here so anyone can re-run them.

What to log per run

One row per prompt per engine per run. Eight columns. Anything more and the log stops getting filled in.

Two of the columns do the work. Other domains cited is the list of pages the engine read to produce the answer. When you are absent, it is the list of pages you need to be on. When you are named, it is the list of pages that can drop you. Rivals named is your competitive set as the engine sees it, which is often not the set you would have written down.

Run each prompt in a fresh session with memory off. Engines sample rather than look up, so a single reading can wobble; if one cell matters, run it three times and record the majority. And read the answer body for position, not the citation list. Being cited at number three while a rival is recommended in the first sentence is not a top-three result.

How to read movement: our own log, three weeks

Here is what a flip looks like. These are the readings on two of our four prompts across three runs, with the source that carried each change.

Read the top row first. On 24 August ChatGPT answered "what is citeOS" with a French street-lighting contractor that shares the name. On 26 August a Crunchbase profile for citeOS went live. On 27 August ChatGPT answered correctly, described the product accurately, and cited that profile three times. It cited nothing on our own site. Four months of structured data on citeos.io had not moved the answer. One directory listing did, in about two days.

Now the bottom two rows. On 24 August Perplexity did not name citeOS for "top AEO tools for crypto". On 27 August we published a ranked page on that question. On 15 September Perplexity opened its answer with citeOS, and Gemini gave it its own section, and in both cases the only page of ours they cited was that one. Nineteen days from publish to first place on two engines, on one page.

Neither of those readings would mean anything as a single check. They mean something because the prompt, the engine and the market were identical each time, so the only variable left is what changed on the web between the dates. That is the entire argument for tracking over checking.

A check answers "am I in the answer today". Tracking answers "what put me there, and what would take me out".

When the answer is wrong about you

The brand prompt is where tracking earns its keep, because a wrong answer looks fine until you read it. Ours was a name collision: a much older company called Citeos, and a newer product at citeos.ai in our own category, both sharing our name. Engines merged the three. The fix has four steps, and the log is what tells you whether each one worked.

• You cannot edit the engine. Feedback buttons are noted, not acted on. What you can edit is the set of pages the engine reads about you, which the other-domains column has already listed for you.
• Put one canonical description everywhere. Same name, same one-line description, same category, on your site, your directory profiles and your social profiles. Where a name collides, add a disambiguation line that names the other entity and says you are not it. We did, and engines now volunteer the distinction unprompted.
• Directories carry more entity weight than your own schema. Crunchbase fixed ChatGPT for us. SourceForge fixed Perplexity three weeks later. Both are free, both are in the corpus as sources for most of the 72 brands, and both were cited first in the corrected answers.
• Re-run the same prompt weekly until it holds. A correct answer one week can revert. Ours has held on ChatGPT since 27 August and on Perplexity since 15 September, and the log is how we know.

Spreadsheet or tool

Four prompts on three engines is twelve rows a week. A spreadsheet is the right tool, and the template above is the sheet. It stays the right tool until one of three things happens: the prompt set grows to twenty, you need Google AI Mode alongside the other engines, or you need every run sampled several times so a single wobble does not read as a change.

At that point the job is twenty prompts, five engines, weekly, majority-merged across samples, with the source list stored per answer so you can diff it. That is what a citeOS audit is: 20 buyer prompts, 5 engines, 100 observation points per run, weights published on the methodology page. Whether you run it by hand or through us, the method does not change. The prompts stay fixed, the dates move, and the difference between two dates is the finding.

AI visibility tracking, answered