The Complete AEO
& GEO Guide

AI answer engines retrieve, then generate. They cite the sentence with the highest information gain, not the page that “ranks.” Fan-out sub-queries are the real unit of AI search.

AI answer engines work in two steps: retrieve, then generate. Google still ranks documents against a query; ChatGPT, Perplexity, Gemini, and Copilot extract passages and stitch them into an answer. That distinction is why a page that ranks well on Google can still be invisible to an answer engine, and vice versa. Large-sample citation analysis finds 44.2% of citations come from the first 30% of text, and 78.4% of question-shaped citations come from H2 headings. That is not a coincidence — it is how the models chunk and lift. It is also why this complete AEO guide opens with a definition and uses question H2s throughout.

Aggarwal et al. (KDD 2024) (opens in a new tab) framed GEO as optimizing content for generative engines. The operator takeaway is narrower than the paper: entities, numbers, and mid-paragraph facts get pulled more often than throat-clearing intros. 53% of citations come from the middle of a paragraph, not the first sentence. If your opening sentence is a hook and your evidence is buried three paragraphs down, the model is more likely to cite the buried evidence than your hook — so write accordingly.

Fan-out is the hidden SERP. One user prompt becomes many sub-queries under the hood — the engine breaks the question into pieces, retrieves for each, then synthesizes. You do not “rank for AEO.” You get retrieved for the sub-questions the engine actually runs. Map those sub-questions before you optimize a headline, or you are writing for a query the model never asked.

Strategy is not one-size-fits-all. Citation overlap across major AI surfaces is about 13.7% in industry studies — treat ChatGPT, Perplexity, Google AI Overviews, Gemini, Copilot, and Claude as different retrieval systems.

Citation overlap across major AI surfaces is about 13.7% in industry studies. That number is the single most important fact in this guide, because it means a strategy that works on ChatGPT can quietly fail on Perplexity and you will not know unless you measure them separately. Treat ChatGPT, Perplexity, Google AI Overviews, Gemini, Copilot, and Claude as different retrieval systems that happen to share a model shape, not as one channel. The table below is the operator view of what each surface rewards and the assumption that gets teams in trouble.

Read the table as a reminder to measure before you optimize. If you run one “AI SEO” report and call it done, you are averaging over platforms that disagree 86% of the time. The practical consequence is that a win on ChatGPT can look like a flat month overall if Perplexity and AIO moved against you. Forbes, HubSpot, and CXL already own definitional SERPs for “answer engine optimization,” and a16z’s GEO-over-SEO essay (opens in a new tab) plus Google’s own AI-search guidance own the GEO head term. The gap this complete AEO guide fills is the practitioner how: one workflow, platform notes, measurement, and the brand-mention lever that the head-term pages skip.

Use the ski ramp: answer in the first 150 words, question H2s, entity-dense paragraphs, a closing summary. 44.2% of citations come from the first 30% of text.

The ski ramp is the content shape that matches how models retrieve. Put the dictionary definition and the headline stat in the lede — the first 150 words. Phrase H2s as the question a buyer or an LLM would ask, then follow each H2 with a direct answer and the proof. Target roughly 15–20% entity density — brand names, paper names, product names, dates — without stuffing. The point is not keyword repetition; it is giving the model named, verifiable things it can lift verbatim.

Why this shape works: 44.2% of citations come from the first 30% of text, so a slow intro is a citation tax you pay on every prompt. Question H2s match the question-shaped sub-queries the engine fans out into, which makes your section more likely to be the retrieved passage. Entity density gives the model something specific to cite instead of paraphrasing you into anonymity.

The Signal Tower Framework is the refresh version of skyscraper for this era. You do not out-word-count the SERP; you add a signal nobody else has — original data, a worked example, a named framework — then make that signal extractable. Information gain is the only durable AEO strategy once formatting is table stakes. Anyone can write question H2s; only you can run the study.

If GPTBot, ClaudeBot, or PerplexityBot cannot fetch the HTML, you are invisible. Server-render the answer. Ship llms.txt. Do not leave Cloudflare Bot Fight Mode on “block AI” by accident.

Before any of the content work matters, the crawlers have to be able to read your page. If GPTBot, ClaudeBot, PerplexityBot, or Google-Extended cannot fetch your HTML, you are invisible to that platform — no amount of ski-ramp structure will fix it. The checklist below is the minimum technical floor. Most of it is verifiable in a week of log review.

• llms.txt and llms-full.txt: a curated map for models, not a sitemap dump. Point at your canonical guides and methodology pages, not every tag page.
• SSR or prerender for the passage you want cited. Client-only shells are a citation tax — the model sees an empty div, not your answer.
• Allow GPTBot, ClaudeBot, PerplexityBot, Google-Extended per your policy — then verify in logs that they actually hit the pages you care about.
• Cloudflare Bot Fight Mode and similar: audit before you celebrate lower bot traffic. A “bot blocked” spike can quietly be your AI crawler audience.

Pair this section with the technical SEO guides hub for crawl, canonicals, and sitemaps. AEO does not replace indexation; it adds a second crawler audience on top of the Googlebot one. A page that cannot be indexed cannot be extracted.

CITE is Capture → Inspect → Target → Edit. It is the operating loop for answer engine optimization when you are done defining terms.

CITE is the operating loop for answer engine optimization once you are past definitions and structure. Capture a prompt set — and “a prompt set” means 60 to 100 runs per prompt per platform if you want a metric, not an anecdote. A single ChatGPT screenshot is not a measurement; it is a story. Inspect mention frequency, the citation URL, and sentiment. Target the page or third-party source that would most change the answer. Edit for extractability and freshness, then recapture the same prompt set and compare.

The reason CITE is a loop and not a checklist is that citations move. A page that earned the cite in January can lose it in March because a competitor published fresher data or because the model retrained. Recapture after every material edit, and recapture on a calendar even when you did not edit — because the rest of the web did.

Brand mentions correlate with AI visibility at r=0.664 versus r=0.218 for backlinks. Brands are about 6.5x more likely to be cited via third-party sources than their own domain.

In replicated studies across roughly 75,000 brands, brand mentions correlate with AI visibility at r=0.664 — over three times stronger than backlinks (r=0.218). Brands are about 6.5x more likely to be cited via third-party sources than their own domain. Read those two numbers together and the strategy writes itself: you cannot AEO your way to visibility by editing your own pages alone, because the models are trained on what the rest of the web says about you.

That is why vendor-only AEO underperforms. Earned roundups, review sites, Reddit threads with real answers, podcasts, and transcripts are the citation pool for commercial queries. Unlinked mentions still train co-occurrence — the model learns “brand X belongs to category Y” even without a hyperlink. Digital PR is now an AEO channel, not a nice-to-have. If you are not earning mentions on pages the model already reads, you are invisible to it regardless of how good your own pages are.

Build a mention ecosystem: one definitive page on your site (this complete AEO guide, a comparison, a methodology) plus a distribution plan that puts your name on the pages models actually retrieve. SEO & AEO knowledge transfer is the internal corpus your team works from; the public web is what the models retrieve. You need both.

A single query is not a metric. Run 60–100 times per prompt per platform. Track mention frequency, citation quality, and share of voice on a frozen prompt set.

A single query is not a metric. AI citations are volatile — on the order of 45% change between consecutive observations in some studies. If you run a prompt once on Tuesday and once on Friday and panic at the difference, you are reading noise. The fix is to run 60 to 100 times per prompt per platform on a frozen prompt set, then track mention frequency, citation quality (URL plus sentiment, not just yes/no), and share of voice over time.

Measure quarterly patterns, not weekly swings. A platform that missed you this week may cite you next week for the same prompt with no change on your end. Do not rebuild the site because Tuesday’s Perplexity screenshot missed you. Do rebuild the page if a frozen prompt set has trended down for two consecutive quarters — that is signal, not noise.

Content not updated quarterly is about 3x more likely to lose AI citations. Freshness is a signal; a 3-month cliff as a hard rule is a myth, but stale stats still get dropped.

Content not updated quarterly is about 3x more likely to lose AI citations. The “3-month cliff” is not a hard rule — there is no calendar line where a page suddenly drops — but stale stats get dropped, and a page that has not been touched in a year is competing against pages that were. Freshness is a signal the models weight, and it is also a signal to the humans who link to you.

Decay shows up as traffic, rankings, and citation share moving together or diverging. Traditional content-decay tools track the first two and miss the third, which means a page can be losing AI citations while its Google traffic looks flat. That is the silent failure mode: the dashboard says fine, the models say gone. Recapture the prompt set after every material edit, and audit citation share on a calendar even when you did not edit.

Build the CITE loop if you have prompts, pages, and a writer who can add information gain. Buy measurement if you cannot run 60–100 samples. Partner when the gap is distribution — agencies who already live in the citation pool.

Most teams do not need to pick one of build, buy, or partner — they need all three at different stages. Build the CITE loop yourself if you have prompts, pages, and a writer who can add information gain. Buy measurement if you cannot run 60 to 100 samples per prompt in-house; a tool that does it once a quarter beats a team that means to but never does. Partner when the gap is distribution — agencies who already live in the citation pool can earn mentions faster than you can build the relationships.

The checklist below is the ungated AI visibility audit. Run it on your priority URLs before you commission a tool or an agency. If half the items are missing, the cheapest win is fixing them — not buying a dashboard that measures around them.

01/

What is answer engine optimization?

Answer Engine Optimization (AEO) is the practice of structuring content and building online presence so AI-powered search engines accurately extract, cite, and recommend your brand in generated responses. The job is twofold: make your own pages extractable, and make sure the rest of the web says something accurate about you.

02/

Is GEO different from AEO?

GEO (generative engine optimization) names the same job aimed at generative engines. Treat AEO and GEO as one practice with platform-specific retrieval. Splitting them into two content strategies is how teams duplicate work and produce two mediocre pages instead of one good one.

03/

Does AEO replace SEO?

No. Crawl, index, and classic rankings still decide whether a passage exists for the model to extract. AEO adds extractability, brand mentions, and prompt-set measurement on top of that foundation. A page that cannot be indexed cannot be cited.

04/

What is LLM optimization?

LLM optimization (LLMO) is the overlapping label for making content and entities easy for large language models to retrieve and cite. Use AEO and GEO as the public-facing terms; reach for LLMO when you specifically mean model-side retrieval and training co-occurrence.

• Cite mechanics first: retrieval, fan-out, first 30% of text, question H2s.
• Platform overlap is low. Do not run one generic “AI SEO” report.
• Brand mentions beat backlinks for AI visibility in the large-sample work. Distribute.
• Measure prompt sets, not screenshots. Refresh quarterly or watch citations decay.