How do I get cited by ChatGPT?
You get cited by being findable on the live web at the moment someone asks, not by being famous enough to be in the model's training data. Modern assistants search the web before answering, so citations come from pages that are indexed, structured, and clearly answer the question in the words a buyer uses — plus third-party sources that corroborate you. We measured this on one of our own sites: asked seven buyer questions, web-connected ChatGPT and Gemini named it in 20 of 27 checks.
There is a lot of advice about getting mentioned by AI assistants, and most of it is written by people who have never measured whether their advice worked. This post is the opposite problem: we measured, got a confident answer, and the answer was wrong. What we found when we fixed the measurement is the most useful thing we know about AEO.
Why most AI visibility numbers are measuring the wrong thing
For five months we probed four AI assistants with buyer-intent questions about our own properties and recorded whether each one named the business. The result looked unambiguous. Perplexity cited them. ChatGPT, Claude and Gemini never did — not once, across roughly 4,700 questions.
That is the kind of finding that gets turned into a chart. It is also completely false, and the reason is embarrassingly simple: we were calling those three assistants through their APIs as bare models, with no web search attached. A bare model can only answer from what it memorised during training. So the question we were actually asking was not “does ChatGPT recommend this business?” It was “is this business famous enough to have been in the training data?”
For a student hostel in Indore, or a spice exporter in Gujarat, the answer to that second question is no, permanently, no matter how good their SEO is. Perplexity was the only engine in our test that searches the web by construction. It was not one engine out of four disagreeing with the others. It was one connected engine against three disconnected ones, and the zeros were an artefact of our own setup.
A bare model tells you whether you are famous. A web-connected assistant tells you whether you are findable. Only the second one is a marketing problem you can solve.
What the number looks like when you measure it properly
On 19 September 2026 we re-ran the same seven buyer questions against ChatGPT and Gemini with web search enabled, for radianceresidency.com — a student hostel we build and run. Questions like “best girls hostel near Medicaps University Indore” and “affordable hostel near Medicaps Indore”: what a parent or student would actually type, never the brand name.
| Assistant | Named the business | Questions asked |
|---|---|---|
| ChatGPT (web search on) | 15 | 21 |
| Gemini (web search on) | 5 | 6 |
| Combined | 20 | 27 |
Zero, then 20 of 27. Nothing about the website changed between those two measurements. Only the question we were asking changed.
If you are paying for an AI-visibility tool, this is the first thing to ask about it: does it query the assistants the way a customer does, with search enabled, or does it query a bare model? If it cannot tell you, its zeros mean nothing.
Ask the question a buyer would type, not your brand name
The second way these measurements go wrong is subtler. Searching an assistant for your own brand name and finding yourself proves nothing — of course it finds you, you gave it the answer. A citation only counts when the assistant had to choose you from a field.
- Useless: “Radiance Residency reviews” — the brand is in the question.
- Useful: “best girls hostel near Medicaps University Indore” — the assistant has to pick.
- Useful: “affordable PG near Medicaps for students” — same intent, different words. Run both; assistants are inconsistent.
Assistants also vary between runs. We have asked the identical question minutes apart and been named once and not the other time. That is not a bug in the measurement, it is how these systems work — which is exactly why a single answer proves nothing and a rate across many questions means something. Any citation figure without a denominator beside it is unreadable.
So what actually earns the citation?
Given that assistants are searching the live web and then summarising what they find, the work divides into three unglamorous parts.
1. Be indexed, which is not the same as being live
An assistant cannot cite a page its search index has never fetched. We watch this constantly on our own sites: our third property has 29 of 188 known pages indexed, most of the rest sitting at “discovered, currently not indexed”, and until that changes more reach simply spreads the same clicks thinner. Check coverage in Search Console before you write another word.
2. Answer the question in the words that were asked
Assistants extract passages. A page with a question-shaped heading and a direct two-sentence answer underneath it is far easier to quote than the same information spread across five paragraphs of preamble. Put the answer first and the argument after — the opposite of how most agencies write.
3. Be corroborated somewhere other than your own site
This is the part people skip because it is the least controllable. Assistants weight third-party sources heavily, and a claim that appears only on your own domain is the weakest form of evidence available. Directory profiles, listicles and professional-network pages do disproportionate work here. We take our own medicine on this and publish what it costs us: our structured data ships a `sameAs` field that is currently empty, because we do not yet have the third-party profiles to put in it.
What none of this can promise you
No one can guarantee you a position in an AI answer. The assistants change their retrieval and ranking without notice, they disagree with each other, and they disagree with themselves between runs. Anyone offering you a guaranteed AI citation is either not measuring or not telling you how.
What you can have is a number, a denominator, and a direction of travel. We publish ours on the work page, including the engine we currently cannot measure at all — because reporting an unmeasured engine as a zero is precisely the mistake that cost us five months of wrong data.