The AI-search landscape & how answers get sourced
Why this lesson
Section titled “Why this lesson”Ranking #1 no longer wins the answer. As of late 2025, when an AI Overview appears on a Google result, it cuts the click-through rate for the #1 organic page by roughly 58% — for every 100 clicks you used to get, Google now keeps 58. Meanwhile AI traffic is tiny but converts extraordinarily well: Ahrefs reports AI search was 0.5% of their traffic yet drove 12.1% of sign-ups. So the game has shifted from “get the click” to “be the source the AI quotes.” Before you can optimize for that, you have to understand the machine: which engines exist, where each one gets its facts, and what the actual evidence says earns a citation — as opposed to the hype. That’s this lesson. Get it right and every tactic in the rest of Level 3 will make sense instead of feeling like a random list of tricks.
The explainer
Section titled “The explainer”One practice, three acronyms
Section titled “One practice, three acronyms”AEO (answer engine optimization) is the practice of making your content visible and useful to AI systems that deliver direct answers — Google AI Overviews, ChatGPT, Perplexity, Gemini, Copilot. You’ll also hear GEO (generative engine optimization) and LLMO (large-language-model optimization). In 2026 these mean essentially the same job, and many practitioners argue the alphabet soup is pointless. A useful working split: an answer engine extracts and presents an answer; a generative engine synthesizes its own paragraph and cites sources as it writes. Either way, AEO doesn’t replace SEO — it layers on top of it. The fundamentals (quality content, authority, technical health) still decide whether an AI can find and trust you.
Two sources: training data and retrieval
Section titled “Two sources: training data and retrieval”Every AI answer draws on two very different wells. The first is training data — the frozen snapshot of text the model learned from. Ask “who is the CEO of Apple?” and ChatGPT answers “Tim Cook” instantly, from memory. But training data is static and months stale, so it can’t know about your product that launched last week. That’s where the second source comes in: real-time retrieval-augmented generation (RAG). When a question is fresh or specific, the engine goes out, searches the live web, pulls back pages, reads them, and generates an answer from what it found. Anchoring an answer to those retrieved documents is called grounding — it’s how an engine backs a claim with a real source instead of hallucinating. This is why classic SEO still matters: the skills that get you ranked are the same ones that get your page retrieved.
The engines and where each one pulls from
Section titled “The engines and where each one pulls from”They are not interchangeable — they source differently:
- Google AI Overviews / AI Mode — a generative summary on top of Google (~48% of queries in 2026). Uses query fan-out: it splits your question into many sub-queries, retrieves for each, then synthesizes and cites a handful of sources. About 76% of its citations are pages already in Google’s top 10.
- ChatGPT Search — retrieves via Bing’s index plus its own crawler (OAI-SearchBot), and leans heavily on Wikipedia (~48% of citations). Being indexed in Bing matters here.
- Perplexity — runs its own fast crawler (PerplexityBot), assigns citations during answer assembly, and skews toward Reddit (~47%).
- Gemini — Google’s model, using Google’s index plus its Knowledge Graph.
- Bing Copilot — Microsoft’s assistant riding the Bing index.
Visibility is a spectrum, and it’s probabilistic
Section titled “Visibility is a spectrum, and it’s probabilistic”Because retrieval and generation involve probabilities (plus a randomness “temperature”), citations are probabilistic, not fixed. Ask the same question five times and you might be cited three. That’s why we track AI visibility — a probability distribution — rather than a ranking. It comes in three flavors: cited and linked (best for traffic, but only ~28% of mentions carry a link), mentioned but not linked (word-of-mouth at scale; still trains the model to associate your brand with a topic), and not visible at all. In a 75,000-brand study, branded web mentions had the strongest correlation with AI visibility (0.664) — stronger than backlinks or domain rating.
The evidence, not the hype
Section titled “The evidence, not the hype”Much of the “AI ranking factors” content online is single-vendor and unreplicated — treat it skeptically. The one peer-reviewed anchor is the Princeton / Georgia Tech / Allen Institute paper “GEO: Generative Engine Optimization” (KDD 2024). Its replicated findings: swapping vague language for specific statistics lifted visibility inside AI answers by roughly 37–41%; adding cited sources and expert quotes lifted it 30–40%, with marked-up quotes reported around +41%. Those three moves — stats, quotes, sources — are the evidence-backed core of everything that follows.
- Pick your brand (or a practice site’s) and write down 5 real questions a customer might ask an AI — the kind with natural language, e.g. “what’s the best CRM for a two-person agency?”
- Run each question in three engines: Google (watch for the AI Overview), ChatGPT, and Perplexity.
- For every answer, record which of the three visibility types you hit: cited-and-linked, mentioned-but-not-linked, or not-visible. Note which competitors were named.
- For one answer, click into the cited sources and note where they came from — a top-ranking article, Wikipedia, Reddit, a YouTube video? Match it to the engine’s sourcing pattern above.
- Write one sentence: which visibility type matters most for your business, and why.
- Log it all in the Level 3 workbook — this baseline is what your capstone will improve.
Terms introduced
Section titled “Terms introduced”Check yourself
Where do AI search engines get the information behind an answer?
Why do we talk about "AI visibility" instead of "AI rankings"?
On average, how often does an AI mention of your brand actually include a clickable link?
You can move on when you can… name the major answer engines and where each one sources its answers, explain the difference between training data and real-time retrieval (grounding/RAG), and describe the three types of AI visibility — and why we measure visibility as a probability, not a ranking.
Go deeper
Section titled “Go deeper”- Princeton / Georgia Tech / Allen Institute — “GEO: Generative Engine Optimization” (KDD 2024): the peer-reviewed source for the stats-and-quotes findings. The evidence base for the whole level.
- Google Search Central — “Optimizing for AI features in Google Search”: the official, first-party guidance on how AI Overviews and AI Mode source content.
- Next: 3.2 · Citability — writing to be quoted — turn “be the source” into concrete, scorable writing moves.