How AI search engines decide which sources to cite

If you've read our explainer on what GEO actually is, you'll know the goal is no longer "rank in Google" but "get cited by the AI." This post goes one level deeper into the mechanics — because understanding how citation works is the only way to influence it.

Citation is not search. It's a different machine.

When you type a question into Google, the search engine returns a list. You scan, click, decide. Easy mental model.

When you ask the same question to ChatGPT or Perplexity or Google AI Overviews, something fundamentally different happens. The AI doesn't return a list. It returns an answer — one piece of synthesised text — and then footnotes that answer with three to five sources. The sources are no longer destinations. They are evidence the model used to build the answer.

That single change rewires the whole game. In list-based search, ranking matters because rank determines clicks. In citation-based search, only three to five sources get shown — and being source #6 is no different from being source #6,000. There is no second page of citations.

There is no second page of citations.

The four stages of an AI citation

Every major AI search engine — ChatGPT, Perplexity, Google AI Overviews, Gemini, Bing Copilot — uses some variation of the same four-stage pipeline. The implementations differ. The shape is the same.

The trap most businesses fall into is optimising for stage one. They focus on being found — building backlinks, climbing rankings, getting indexed — and then assume the AI will do the rest. But being found only gets you into the candidate pool. The citation is decided two stages later, by whether the model can extract a clean answer from your page.

The five signals that decide your fate

Across the four stages, five signals do most of the work. Get these right and you're in the conversation. Get them wrong and you're invisible regardless of how good your business actually is.

Notice what's missing from that list: backlinks. They still matter as a trust input, but they no longer dominate. A page with no backlinks but a perfectly extractable answer can outcite a page with hundreds of backlinks but no clear answer structure.

How the five major platforms differ

The pipeline is shared. The implementation isn't. Each platform draws on different source pools and applies different weights to the signals above.

What this means for your website

The implications are uncomfortable for businesses that have invested heavily in traditional SEO. Some of what was true a few years ago is no longer true. Specifically:

• Long doesn't beat clear. A 2,500-word pillar page built to rank can lose to a 600-word page built to be extracted. Word count is no longer a positive signal — it's neutral at best.
• Keyword density is irrelevant. The AI is reading for meaning, not for term frequency. Repeating your target phrase eight times in 400 words now hurts more than it helps.
• Hedged language costs you. "Many experts suggest that perhaps" is unextractable. "This is X" is extractable. Confidence wins.
• Your entity matters more than your domain. A consistent, verifiable business identity across Google Business Profile, LinkedIn, schema markup, and third-party mentions can outweigh raw domain authority.
• Blocking AI bots is self-sabotage. Some businesses are still blocking GPTBot and ClaudeBot by default out of caution. That decision has a direct cost: invisibility in the search engines that are eating the share of voice.

The shortcut question

If you only have time for one diagnostic, ask this:

If yes, you're already in the game. If no — and most websites are in the "no" camp — that's the single biggest piece of low-hanging fruit in GEO. Restructuring the opening of key pages to lead with the answer, before context and storytelling, typically lifts citation rates within weeks.