GEO and AIO: SEO for AI Search

By the end of this level you will know why an AI system cites one page and quietly absorbs another, and you will be able to write passages built to be retrieved, read and attributed by Google AI Overviews, Google AI Mode, ChatGPT, Perplexity and Gemini. This is not a bag of tricks. It is the citation selection mechanism, taught with the evidence that actually exists and with the gaps marked where they are.

The mechanism fits in one sentence: interchangeable content gets absorbed, distinctive content gets cited. When a page repeats what fifty other pages already say, the model can compose its answer without leaning on any one of them, so it names none of them. Advanced Web Ranking (July 2026) found that 82% of uncited passages were pure restatement of consensus. What survives the filter is what the model cannot assert on its own: a dated figure, a number only you hold, an entity named without ambiguity, a position that contradicts the consensus with evidence behind it.

That asymmetry has been measured twice, by different teams using different methods. The GEO paper (KDD ’24) found that citing sources raised AI visibility by +30%, adding quotations by +41%, adding statistics by +40% and defining technical terms by +20%. Advanced Web Ranking (July 2026) found that fresh dates appeared in 80% of cited passages against 53% of uncited ones, and that 96% of cited passages named entities at first mention against 82% of uncited ones. Two independent lines of evidence point at the same writing habits.

One warning before you start. This is the most volatile level in the course: the systems studied here change behaviour without notice and without a public changelog, and a large share of what gets published about them has no primary source at all. This level works only from what can be verified, flags contested claims as contested, and says plainly when no number exists.

What this level covers

• Telling GEO, AIO, AEO and LLMO apart instead of repeating the industry’s vocabulary confusion, and knowing which of those labels describes real work.
• Explaining how Google AI Overviews and Google AI Mode choose the sources they cite, and why treating them as one system produces false conclusions.
• Writing self-contained passages that survive chunking and still make sense away from the page that contains them.
• Deciding on evidence whether to block GPTBot, ClaudeBot or PerplexityBot, knowing what each choice costs and what does not come back afterwards.
• Measuring what can be measured and saying out loud what cannot: Google Search Console’s generative AI report gives impressions only, with no clicks, no queries and no CTR.
• Recognising an unsourced claim about AI search and discarding it, even when the whole industry is repeating it that week.

The absorption pattern: why consensus leaves no trace

A generative AI system cites when it needs a source and absorbs when it does not. Google AI Overviews, ChatGPT and Perplexity compose an answer from several retrieved passages; when a passage says something the model can already assert unaided, the idea enters the answer and the source falls away on the journey. Advanced Web Ranking (July 2026) measured that 82% of uncited passages were pure restatement of consensus.

The implication is uncomfortable for anyone writing general guides: the better you summarise what is already known, the less reason there is to name you. A flawless explainer on what internal linking is makes perfect absorption material. Your own experiment measuring what internal linking did to one specific site, with a date and a sample size attached, is citation material.

Writing to be cited means handing the model something it cannot generate alone: a figure only you hold, a recent and explicit date, a precise definition of a technical term, an entity named so it cannot be confused with another. Defining technical terms measured +20% AI visibility in the GEO paper (KDD ’24), and it is the cheapest habit on the list to adopt.

Retrieval and reading are two separate filters

Being retrieved and being read are different events, and a page can pass the first and fail the second without anyone noticing. ChatGPT’s retrieval layer caches pages as HTML converted to Markdown, with roughly a 30-minute freshness window. What the model processes is not your page as a browser renders it. It is the output of that conversion.

One technical consequence follows directly. Semantic HTML served from the server converts cleanly to Markdown; a thicket of nested containers produced by a visual page builder does not. Real headings, real tables, real lists and real paragraphs come through the conversion with their structure intact, and that structure is what makes an individual passage possible to isolate.

The second consequence is editorial. The passage, not the article, is the unit that gets retrieved: a chunk starts at a heading and ends before the next one. When the opening sentence of a section begins with a pronoun pointing back at the previous paragraph, that chunk reaches the model with no subject in it. This is why the full entity name repeats in every section even when reading the whole article makes it feel redundant. Advanced Web Ranking (July 2026) found that 96% of cited passages named entities at first mention, against 82% of uncited ones.

What you can measure and what you cannot

Instrumentation for AI search is markedly worse than for classic search, and it is worth knowing that before promising anyone a monthly report. Google Search Console’s generative AI report gives impressions and nothing else: no clicks, no queries, no CTR. Any dashboard offering all three for Google AI Overviews is estimating them from its own model rather than measuring them.

Outside Google Search Console there is referral analytics. ChatGPT and Perplexity link the sources they cite, and those clicks land in your analytics with their originating domain attached, so attributed visits are real even when the volume is small. What never shows up is the majority case: the answer that resolves the user’s question without a click on anything.

The honest method while instrumentation stays this thin is your own dated sampling. Fix a list of representative prompts, run them with identical wording across Google AI Mode, ChatGPT, Perplexity and Gemini, record which domains appear as citations, and repeat on a stable cadence. That produces a first-party time series with a date and a sample size on it, which is exactly the kind of data other people later have to cite.

The four measured effects from the GEO paper

The GEO paper (KDD ’24) remains the strongest public experimental evidence on which writing changes raise the odds of a generative engine citing a page. It measured four specific interventions on the text itself, all of them executable by a writer without touching site code. The table sets out each effect and what it means at the keyboard.

Change to the text	Measured effect on AI visibility	What it means when writing
Citing sources	+30%	Link the primary source at first mention and close with a dated sources block. It doubles as your defence when a figure ages badly.
Adding quotations	+41%	The largest of the four effects. A literal, attributed sentence from a named source in every major section, not a paraphrase of one.
Adding statistics	+40%	A specific number carrying its date and its sample size. A percentage with no origin and no date fails this test and is indefensible anyway.
Defining technical terms	+20%	Use the precise term and define it on first appearance. Simplifying the vocabulary does not help; explaining it does.

Two cautions on reading this table. All four percentages come from one experiment over one set of queries, so treat them as reliable directions rather than per-page guarantees. And none of the four substitutes for having something to say: applied to a restatement of consensus, they still produce a restatement of consensus.

The 14 lessons in this level

1. GEO, AIO, AEO, LLMO: What They Mean and Which Ones Matter
2. AI Overviews: How Citations Are Selected (The 93.67% That Explains It)
3. AI Mode Is Not AI Overviews: 13.7% Overlap and What It Means
4. How ChatGPT Search Really Works: The ‘Labrador’ Index Exposed
5. Retrieved Isn’t Enough: Getting Read (74% vs 7%)
6. The Absorption Pattern: Why Consensus Content Never Gets Cited
7. The GEO Paper: +41% for Quotations, +40% for Statistics
8. Writing Retrievable Chunks: Optimizing at Passage Level
9. Entities and Disambiguation: How a Model Recognizes You
10. Does Schema Help AI Citations? The Evidence Says No
11. llms.txt: Autopsy of a Standard Nobody Reads (97%)
12. AI Crawlers: GPTBot, ClaudeBot, PerplexityBot and the Cost of Blocking
13. Cloudflare, Pay-Per-Crawl and the Economics of Content
14. Reddit, YouTube and LinkedIn: The 59% of Citations You Don’t Control

How to work through this level

Take the lessons in order, because the first four build the vocabulary and the mental model the other ten rely on. Lessons 1 to 5 explain how the systems behave, lessons 6 to 9 are the writing lessons and the ones that change your daily work most, and lessons 10 to 14 cover infrastructure and platform decisions.

Work each lesson against a real page of yours rather than a worked example. Pick a published article, find its passages, rewrite the opening sentence of each section so it stands alone, add a dated figure or an attributed quotation to every section, and check that each entity is named at first mention inside that section.

Set up your own prompt sampling in week one, however small. Ten fixed prompts, run with identical wording across Google AI Mode, ChatGPT, Perplexity and Gemini once a month, are worth more than any third-party study, because they cover your market and they carry your date.

Then apply the expiry rule. Research into AI search has a short shelf life. Any measurement, screenshot or claim about how Google AI Overviews, Google AI Mode, ChatGPT, Perplexity or Gemini behave that is more than a quarter old gets treated as a lead worth re-checking, never as a fact. That includes what you read in this course: every lesson is dated so you can apply the same rule to it.

This level sits inside a longer path that starts with fundamentals and technical SEO and ends with measurement and strategy. If you arrived here directly, the full index with every level and its recommended order is on the free SEO course at doctor-seo.net, and it is worth having the technical level settled before optimising to be cited: a page that crawls badly does not get retrieved either.