Getting mentioned by AI is fine. Getting cited as a source — with your link inside the answer — is another league. The first gives you a pat on the back; the second sends you traffic and authority. Most guides tell you to "create good content" and call it a day. Here's the concrete part: how you earn a citation in ChatGPT, Perplexity, and Gemini, no theatre.
Mention vs citation: why they're not the same
A mention is the model writing your name. A citation is the model listing you as the linked source of what it claims. They look alike and they're not: you can get name-dropped in passing and earn zero clicks, or be the source that holds the answer up and take both the credit and the visit.
The difference matters because the work is different. To get mentioned, being a known entity is enough. To get cited, you have to be the easiest piece to verify and reuse on that specific question. A citation is a mention with evidence.
How an LLM decides who to cite
A classic engine ranks links by backlinks and domain authority. A generative engine doesn't rank: it retrieves and synthesizes. Underneath sits a RAG (Retrieval-Augmented Generation) pipeline that picks passages on three things: semantic relevance, information gain, and entity coherence. Translation: it cites you if your content answers better, brings something the other sources don't, and makes clear who's saying it.
That shifts the lever. You don't optimize a page for a keyword: you optimize a passage to be the most citable answer to a question. And an uncomfortable stat: roughly 44.2% of citations come from the first 30% of the content. Bury the answer in paragraph twelve and you don't exist for the model.
The tactics that move the needle (with data)
The study that coined the term GEO (Princeton, presented at KDD 2024, across some 10,000 queries) measured what actually changes generative visibility. It's not magic: it's evidence.
- Put the answer up top. Front-load the direct claim in the first lines of every section. The model rewards what it can extract without digging.
- Cite data and sources. Adding statistics lifted visibility by 22% and adding citations by 37% in that study. AI cites those who cite.
- Bring your own data. A number that lives only on your site makes you a primary source: if the answer needs that figure, it needs you.
- Structure to be parsed. Clear headings, lists, one idea per block, a delimited FAQ. What's machine-readable gets cited more.
- Be an entity. Consistent name, clear description, schema. If the model doesn't know who you are, it won't risk citing you.
Let the crawler in (or you don't exist)
All of the above is moot if you lock the door. If your robots.txt doesn't allow the AI crawlers, that answer will never see you. The user-agents that matter today: OAI-SearchBot and GPTBot (OpenAI), PerplexityBot, ClaudeBot, and Google-Extended. Blocking one is giving up that engine's citations.
It's the most expensive mistake and the most invisible: people investing in GEO content with the gate shut. Before you optimize anything, check that you're letting in the one that's going to cite you.
Every engine cites differently
Optimizing for "AI" in the singular is the classic mistake. They don't read alike and they don't cite alike.
| Engine | How it cites | Where to focus |
|---|---|---|
| Perplexity | Retrieves the web live, cites on every answer | Recent, well-structured content: the fastest path to your first citation |
| ChatGPT | Answers more from training, cites less | Being a recognizable entity the model already "knows" |
| Gemini | Mixes Google’s index with synthesis | Solid SEO base + structured data |
Perplexity cites in around 13% of its answers; ChatGPT, far less — around 0.6% — but it drives most AI referral traffic. Practical takeaway: start with Perplexity to confirm you're retrievable, and build entity to break into ChatGPT, where the volume is.
What to measure to know it's working
If your only KPI is traffic, you won't see the citation coming. Measure three different things: presence (do you show up in the answer or not), citations (does it link you as a source or just name you), and share of voice (what slice of your category's citations is yours). Ask each engine about your category, log who gets cited, repeat weekly. What you don't measure here doesn't improve: in GEO there's no official dashboard — you build it yourself.