AEO (Answer Engine Optimization) is the practice of making your content likely to be selected and cited when AI systems generate answers to questions. GEO (Generative Engine Optimization) is the research term for the same problem — coined in the 2023 Aggarwal et al. paper (arXiv:2311.09735) — now used in the industry as the umbrella name for AI-visibility work. Everything on this page is the condensed basics: the definitions, the pipeline an answer goes through, the five levers the work acts on, and what the evidence actually settles. Below that: the 12-part series in reading order, one line per part.

The pipeline in four steps

Every generated answer you’re trying to get into is produced the same way, across surfaces:

  1. Understand the query. The question may be rewritten or fanned out into related sub-queries — Google documents its “query fan-out” technique of “issuing multiple related searches across subtopics and data sources” (Google Search Central, https://developers.google.com/search/docs/appearance/ai-features).
  2. Retrieve. The engine draws on the model’s training knowledge, on live web search, or both. ChatGPT chooses to search the web based on the question (or on your manual toggle); Perplexity searches the live web in real time by default; in Profound’s ~700,000-conversation ChatGPT study only ~18% of conversations triggered a web search at all.
  3. Select. A subset of retrieved candidates is chosen to ground the answer. Retrieved-more-often candidates are more likely to be cited (observational finding, arXiv:2610.15382), but the selection logic itself is unpublished, and external proxies recover only about a quarter of the URLs that actually get cited.
  4. Generate with citations. Prose is composed from the grounding, citations attach to claims, and claims bind to entities — your named product or company, resolved from consistent naming and cross-web corroboration, not from a single URL.

The full treatment, with per-surface verification, is how AI answers are actually generated.

The five levers

All AEO/GEO work reduces to five levers. A gap in any one is invisible to a rank tracker, which is why the failure is usually found by asking the questions in the surfaces themselves rather than by auditing tooling:

  1. Extractability. The passage must be liftable: specific, self-contained, current. Marketing copy that earns a click cannot be quoted into an answer. What a gap looks like: your pricing or comparison page has the fact, but paraphrased, buried, or behind a form.
  2. Entity clarity. Your product and company must resolve as one unambiguous entity — consistent naming, Organization/Product structured data, stable URLs, facts that agree across your own pages. What a gap looks like: the product name is in the logo but not the text, and your category label differs between the homepage and the docs.
  3. Corroboration. Answers draw from several sources — cited ChatGPT conversations average ~6 unique citations (Profound, February 2026). Your domain is one slot at best; the others are review, community, and comparison presences you can legitimately build. What a gap looks like: you have no current profile on the review platform your category uses, and no independent comparison page names you.
  4. Retrievability. The fetchers that ground live answers — distinct crawler identities, plain-HTTP fetches, their own robots.txt rules — must actually reach your pages and get clean content. What a gap looks like: your top page returns a JavaScript shell to a plain HTTP client, or robots.txt blocks a crawler you didn’t recognize and now can’t see in any report.
  5. Measurement. Cited, mentioned, or absent — per money question, per surface, per month, against a named competitor set. No diff, no discipline. What a gap looks like: you last asked your category question in ChatGPT two quarters ago.

The evidence base, honestly labeled: the founding study (Aggarwal et al., 2023) is a single experiment showing passage-level changes can lift response visibility by up to 40% in its benchmark — a real direction, an upper bound, one study. The 2026 ChatGPT citations study (arXiv:2610.15382) is observational production evidence, not causal. Vendor data (Ahrefs’ most-cited domains, Profound’s conversation analysis) is vendor-run and vendor-labeled. None of it supports promises; all of it supports the five levers.

What is settled, what isn’t

Settled: the pipeline (retrieve → select → generate with citations) is how the surfaces work, and it is documented by the surfaces themselves; citation presence is measurable per question and per surface; ranking #1 on Google does not transfer to the answer layer — the CTR data behind that divergence is in you rank #1, AI answers still skip you.

Not settled: the exact selection logic inside any engine; the size of live retrieval pools; whether citation drives traffic or is merely correlated with stronger brands (Seer’s +35% cited-brand CTR premium is observational, via the CTR studies comparison); and any long-term “AI traffic” forecast. Anyone selling certainty on the second list is selling a promise.

The 12-part series, in reading order

The series has three layers: definitions, infrastructure, and content & corroboration. Reading order within each layer is suggested, not mandatory — every post stands alone. Each post opens with a self-contained definition or summary, so you can jump to the layer you need: if you’re selling, start at the definitions; if you’re fixing, start at infrastructure; if you’re writing, start at content & corroboration.

Layer 1 — Definitions

  1. What is AEO? — the commercial definition, the term’s history, and what AEO is not.
  2. What is GEO? — the research origin (arXiv:2311.09735), what the study did and didn’t show, and the acronym problem.
  3. AEO vs SEO vs GEO — the comparison table, where the terms overlap, and which to use when.
  4. How AI answers are actually generated — the four-stage pipeline, verified per surface: ChatGPT, Perplexity, Google.

Layer 2 — Infrastructure

  1. How AI engines crawl your site — training crawlers vs answer crawlers, the fetch path, and what robots.txt silence costs.
  2. What is llms.txt — the proposed machine-readable index for LLM agents: what the spec says, what it can’t do.
  3. AEO and GEO glossary — 30 terms defined precisely enough to be quoted.
  4. Your first 20 minutes of AEO — a manual audit, no tools required.

Layer 3 — Content & corroboration

  1. What makes content AI-citable — the extractability mechanics: specificity, self-containment, in-passage entities, currency.
  2. Entity clarity for AI answers — why citations bind to entities, and how to be the right one.
  3. Why AI answers triangulate — what third-party sources do for your citations, and which presences are legitimate.
  4. AI visibility metrics, defined — the five numbers that matter and what each cannot tell you.

The measurement and fix process behind all of it — query map, surfaces, monthly diff, raw captures — is documented in our methodology.

Where to start

If you have ten minutes: read this page, then post 4. If you have an hour: posts 1–4 and 8. If you’re deciding whether to invest at all, skip the reading entirely and run the measurement first — the diagnosis is worth more than the theory. Once you have the snapshot, the rest of the series tells you what to do about whatever it found, and the standing version of that run — query map, four surfaces, monthly diff, raw captures, so a snapshot becomes a tracked number — is documented in the methodology.