GEO AI: How AI Engines Decide Which Sources to Cite

AI engines decide which sources to cite through two distinct mechanisms working together: retrieval, which selects live web pages at the moment you ask, and training data, which shapes what the model already considers authoritative. Understanding the difference is what separates guessing at GEO from engineering it. This article opens the black box.

Most people picture an AI assistant reading the whole web in real time. That is not how it works. A citation is the output of a pipeline with clear, observable steps, and each step is a place where your page either qualifies as a source or gets dropped.

The two engines behind every citation

When ChatGPT, Perplexity, Gemini, or Google AI Overviews answer a question, two different systems decide what gets cited. The first is the model's training data: the frozen snapshot of text the model learned from, which encodes which domains and ideas it treats as trustworthy. The second is retrieval: a live search that fires at query time and pulls a handful of current pages into the context window.

These two engines reward different things. Training data favors sources that were already prominent and frequently referenced when the model was trained, so authority compounds slowly over months and years. Retrieval favors pages that are crawlable right now, match the query precisely, and state their answer clearly enough to be quoted. You influence the first through reputation and the second through structure.

The practical consequence: a page published last week cannot be in the training data, but it can absolutely be retrieved and cited today. Retrieval is the lever you can pull immediately. That is why GEO work concentrates there first.

How the retrieval step actually selects a source

Retrieval, often called RAG (Retrieval Augmented Generation), runs a search, ranks the results, and feeds the top few into the model as context. The model then synthesizes an answer and attaches citations to the passages it leaned on. A page only gets cited if it survives every stage of that funnel.

The signals that decide whether your page survives retrieval and earns the citation:

• Crawler access: If GPTBot, Google-Extended, ClaudeBot, or PerplexityBot are blocked in robots.txt, your page cannot enter the candidate set at all. This is the silent disqualifier most sites never check.
• Query match: Retrieval is semantic, not keyword-literal. A page that answers the specific question, in the user's framing, ranks above a page that merely mentions the topic somewhere.
• Extractable answer: The model prefers passages it can lift cleanly. A self-contained answer in the first 150 words, with no dependency on prior context, is far easier to quote than an answer buried mid-article.
• Structured signals: Schema markup (FAQPage, HowTo, Article) and clear question-based headings tell the retrieval layer what each block answers, making the right passage easy to locate and attribute.
• Corroboration: Engines lean toward claims that agree with other sources and with the model's training data. A page that contradicts the consensus without strong signals is less likely to be the cited source.

Why citation is not the same as ranking

It is tempting to assume that ranking first on Google means getting cited by AI. The overlap is real but partial. Both reward crawlable, authoritative, well-structured content, so a strong SEO page starts ahead. But the selection criteria diverge in ways that matter.

Google ranking is a relative contest: you beat the other results for a slot. AI citation is a fit test: your passage either answers the exact sub-question the model is composing, or it does not get pulled, no matter how high you rank. The signals split like this:

• What wins a Google ranking: backlinks, position history, click-through rate, overall page authority
• What wins an AI citation: a directly extractable answer, query-precise framing, schema clarity, crawler access

This is why a page can rank on page two of Google yet get cited by Perplexity, and why a number-one result can be ignored by ChatGPT. The engines are answering different questions about your page, so optimizing only for rank leaves citations on the table.

How to engineer for citation

Once you understand the mechanism, the tactics follow directly. These five moves target the exact points where retrieval decides to keep or drop your page:

1. Unblock AI crawlers first. Check robots.txt for GPTBot, Google-Extended, ClaudeBot, and PerplexityBot. Blocked crawlers mean zero citations, full stop.
2. Front-load a self-contained answer. State the direct answer in the first 150 words so the model can extract it without needing the rest of the page.
3. Match the question's real framing. Use headings that mirror how people actually ask, since retrieval is semantic and rewards precise fit over keyword stuffing.
4. Add structured signals. FAQPage and Article schema label each block so the retrieval layer knows which passage answers which question.
5. Earn corroboration over time. Backlinks and consistent, accurate claims feed both retrieval ranking and the next generation of training data.

The limits of engineering a citation

You cannot structure your way past weak content. Retrieval can surface a thin page, but the model often declines to cite a passage it does not trust, and post-training filters increasingly demote low-quality sources. The mechanism amplifies credibility, it does not invent it.

You also cannot control the training-data half directly. That layer moves on the timescale of model releases and reflects your reputation across the whole web. The honest takeaway: optimize retrieval now for fast wins, and build genuine authority so the next training run treats you as a default source.

Turning the mechanism into a plan

Knowing how engines cite sources is only useful if you act on the pages that already have a chance. The fastest path is to find your pages with existing impressions and traffic, then apply these retrieval signals to them first, since they are already proven and crawled.

If you have GA4 and Search Console connected, that prioritization is already possible from your own data. They Will Know Me reads it and builds a 30/60/90-day plan that targets exactly these citation signals, page by page, for 9.99 euros a month. It connects in 60 seconds and generates your first report immediately.