the citation monitor: find out which domains are actually winning your category in chatgpt.
the standard advice is "build authority and get mentioned in more places." true, useless, and impossible to act on. here's the version i actually use: find out exactly which domains ChatGPT cites when it answers your buyers' questions, then work on those specific surfaces instead of guessing.
This is measurable. It takes about an hour to set up and produces a list of named domains, not a vibe.
Citation monitoring is defined as repeatedly running your buyers' real questions through search-enabled AI systems and logging every source cited, in order to identify which domains decide the answers in your category. It differs from mention tracking, which records only whether your own name appeared.
why this is different from checking if you show up
Checking whether your name appears in an AI answer tells you your score. It doesn't tell you the game. Citation tracking asks a different question: when you are invisible, who is filling the space, and through which sources?
That distinction matters because the answer is frequently uncomfortable. Sara Soleymani published a citation-monitoring build in July 2026 that ran 15 buyer queries against ChatGPT's search-enabled model, three samples each, and logged roughly 300 citations. Her target brand's own website was cited exactly once across all 300. Once. When that brand appeared in ChatGPT's answers at all, it arrived through third-party sources, almost never through anything the company controlled.
That's the whole argument for measuring before optimizing. Publishing another comparison page on a blog nobody's citing changes nothing, and you'd never know that from a mention check alone.
the method, in five steps
1. Build your query set. 15 to 20 questions your buyers actually ask, with no brand names in them. Same set you'd use for any AI visibility tracking, which is deliberate: the two checks answer paired questions about the same landscape.
2. Ask with web search on, three times per query. This is the technical piece that matters. A plain chat completion answers from training data and cites nothing. Only a search-enabled model returns structured citations. Three samples per query, because a single run is a coin flip, not a measurement.
3. Log every cited URL. Query, sample number, domain, full URL, page title. If a sample returns no citations at all, log that too, otherwise your hit-rate math quietly inflates itself.
4. Classify each domain by type. Six buckets: your own site, a competitor's site, a review platform, a community (Reddit, forums), a publisher (media, blogs), or documentation. This classification is the whole point, because each type has a completely different playbook.
5. Run it weekly. Citation sources shift as models update their search behavior. A domain that dominates this month can fade next month. One run is a snapshot. Weekly runs are a monitoring system.
how to read what comes back
This is where the method earns its hour. Three ways to read the same data:
Read down a single query. Which domains own this specific answer? If a query that matters to your revenue returns citations to two review platforms and one Reddit thread, you now know the three surfaces to work on, and that none of them is your blog.
Read across the source types. What share of your category's citations go to review platforms versus communities versus publishers versus brand-owned pages? That distribution is your channel strategy, derived from data rather than instinct.
Read the trend line. Weekly runs show you whether the work you're doing is moving anything, which is the only way to tell "this isn't working" apart from "this is working and I can't see it yet."
the finding that should change your default assumption
Here's the one line worth remembering: the conventional GEO playbook may be pointing you at surfaces your category's AI answers barely read.
In Soleymani's run, the standard advice would have said invest in review platforms. The actual data showed G2 cited once and GetApp cited once across roughly 300 citations. Meanwhile, citations flowed to niche listicle blogs and product documentation. For that category, a review-platform program would have been a quarter of wasted effort, and only measurement would have caught it.
The reverse is equally possible in your category. That's the point. The playbook isn't wrong, it's just not universal, and the only way to know which version applies to you is to look.
what each source type actually tells you to do
| what the citations show | the play | who usually owns it |
|---|---|---|
| Review platforms cited heavily | Review volume, recency, and profile completeness on the specific platforms cited | customer marketing or CS |
| Community sources cited heavily | Genuine participation where the cited threads live, sometimes one specific thread | often nobody, which is why it gets missed |
| Publishers cited heavily | Earned media and inclusion in the specific roundups being cited | PR or comms |
| Your own site barely cited | Check whether the pages exist at all for those query intents before writing more | you |
That last row deserves a note. A page that exists but never gets cited is an optimization problem. A page that does not exist at all is a content spec, and the competitor and publisher pages being cited in its place are that spec's outline, already written for you.
the honest limit of this
The monitor tells you where the doors are. Opening them is a different kind of work, and frequently not content work at all. Review programs, community participation, and earned media each sit with a different function inside a company, which is exactly why AI visibility behaves more like a go-to-market program than an SEO task.
For a solo operator, that's actually good news: you own all of those functions already. There's no one to coordinate with. The constraint is your time, not organizational alignment, which means the measurement translates to action faster for you than for the enterprise teams paying five hundred a month for the same data.
get the template
I've packaged this into a free Citation Monitor Lite: a spreadsheet with the query structure and logging columns already built, plus a copy-paste Claude Code prompt that runs the whole thing for you if you'd rather not do it by hand.
If you'd rather have the full version run for you across every buying query in your category, with the ranked fix plan attached, that's what the audit is. start with the free scan →
Related reading in this series: the pilot study on who ai names when brands look for creators is this exact method run live in one category, and the 51-business ai visibility scan reached the same conclusion by a different route: the sources AI trusts are rarely the ones businesses own.
want to see your own citation picture first?
the free ai visibility scan runs the first part of this method against your site in about sixty seconds. results on the page, no email needed.
run the free scanfrequently asked questions
How do I find out which sources ChatGPT cites in my category?
Run your real buyer questions through a search-enabled AI model several times each, log every cited URL, and classify each domain by type. A plain chat answers from training data and cites nothing, so web search has to be enabled.
Why classify cited domains by type?
Because each type implies a completely different action. Review platforms mean a review program, communities mean genuine participation, publishers mean earned media, and your own site barely appearing means either a missing page or an unstructured one.
How often do AI citation sources change?
Frequently enough that a single snapshot is unreliable. One 2026 analysis re-collected citations across 30 queries over six weeks and found only 119 of 1,127 originally cited URLs still being cited at the end.
What if my own website barely appears in the citations?
Check first whether a page even exists for that query intent. A missing page is a content specification, and the competitor or publisher pages cited in its place are effectively its outline. A page that exists but is never cited is a structure problem instead.
How many queries do I need to track?
Ten to twenty real buyer questions is enough to see the pattern. Sort results by how many distinct queries each domain appears on rather than by total citations, because breadth across queries is what identifies a structurally important source.
written by elisabeth hitz, certified in anthropic's ai fluency program (framework & foundations, and ai capabilities & limitations), plus claude 101 and claude cowork. sources: sara soleymani, "which domains win your category in chatgpt, and what do you do about it?" (medium, july 2026), including the published n8n workflow and citation methodology · forrester's 2026 buyers' journey survey via machine relations research.