GEO vs SEO vs AEO: what actually changed

Three names for one shift

Search Engine Optimization, Answer Engine Optimization and Generative Engine Optimization are usually presented as rivals. They are better read as a timeline. Each term was coined when the system on the other end of the query changed, and each inherits almost everything from the one before it.

The short version: SEO optimises a page to rank in a list, AEO optimises a passage to be extracted as an answer, and GEO optimises a source to be cited inside an answer that a language model has written. This guide places each term, separates what carries over from what does not, and ends with a decision guide for where to spend effort.

Where each term came from

SEO: the page and the ranking

Search Engine Optimization dates from the mid-1990s and has a stable meaning: make a URL discoverable, indexable and competitive for a query so that it ranks high in a list of results. Its signals are well documented — crawlability, relevance, links, page experience — and Google publishes most of the guidance that defines the field on Search Central. The reader is a human who scans ten links and picks one.

AEO: the passage and the extraction

Answer Engine Optimization is an industry term, not an academic one, and it has no single origin. SEO practitioners were using it before generative AI to describe the work of winning featured snippets, "People also ask" boxes and voice-assistant answers — all systems that return one passage instead of a list. After ChatGPT's release in late 2022 the term was widely adopted for AI assistants as well. The glossary entry gives the working definition this site uses, and What is Answer Engine Optimization? walks through the tactics.

GEO: the source and the citation

Generative Engine Optimization has a precise origin. The paper GEO: Generative Engine Optimization by Pranjal Aggarwal, Vishvak Murahari and colleagues was posted to arXiv on 16 November 2023 and later presented at KDD 2024. It defines a generative engine as a system that synthesises an answer from multiple retrieved sources with a large language model, proposes a benchmark called GEO-bench, and measures which edits to a source increase its share of the generated answer.

Its most repeated finding is that edits which added citations, quotations from authorities and statistics improved visibility, while stuffing in extra keywords did not help. That result is the practical dividing line between GEO and SEO: evidence outperforms keyword coverage when the reader is a model.

What stays the same

The newer terms attract a lot of claims that the fundamentals no longer apply. The vendors say otherwise.

Google's documentation on AI features states that "there are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary", and that no new machine-readable files or special schema.org types are needed. The same Googlebot that indexes for Search feeds AI Overviews and AI Mode, and the same noindex, nosnippet, data-nosnippet and max-snippet controls limit what those features can show.

Three things therefore carry over unchanged:

1. Crawlability and indexing. A page that cannot be fetched, or that is blocked for the relevant user agent, cannot be retrieved. OpenAI, Anthropic and Perplexity each publish the crawler names their search features use, and each says those crawlers respect robots.txt.
2. Quality and E-E-A-T. Google's guidance on helpful content says E-E-A-T "isn't a specific ranking factor", but that its systems use a mix of signals to identify experience, expertise, authoritativeness and trust. Answer engines lean on the subset of those signals a machine can check without reading for quality: a named author, visible dates, and links to primary sources.
3. Semantic HTML and clean rendering. Server-rendered HTML with a logical heading outline was good practice for Googlebot and is now the difference between a passage that chunks cleanly and one that does not.

If a page does not rank at all, AEO and GEO tactics have nothing to work with. That is the single most useful thing to know about the three terms.

What actually changed

Three things did change, and they are structural rather than cosmetic.

The unit is the passage

Classic ranking scores a URL. An answer engine splits the page into chunks — usually by heading and paragraph, sometimes by a fixed token window — and scores each chunk against the question. A passage that opens mid-argument, or that refers to "it" without naming the entity, loses at this stage regardless of how strong the page is. Every chunk has to survive on its own.

The reward is the citation

A ranking produces a click. A generative answer produces a citation: a footnote or inline link to the source the model used. The citation may or may not produce a click, and some engines send far less referral traffic than others. That changes what you measure — share of citations for a set of questions, not only sessions from organic search — and it changes what you write for, because the model is choosing which passage to attribute, not which page to list.

The pipeline is retrieval, not ranking

Answer engines are built on retrieval-augmented generation: fetch candidate passages, ground the model in them, generate, attribute. Each stage rewards something different — being fetchable, chunking cleanly, matching the question's wording and entities, containing verifiable evidence, and being the answer rather than discussing it. SEO tactics map onto the first stage only; AEO and GEO tactics map onto the rest.

The pipeline also changed the crawler landscape. OpenAI's crawler documentation lists OAI-SearchBot for ChatGPT's search features, ChatGPT-User for user-initiated fetches and GPTBot for model training. Anthropic's support documentation lists Claude-SearchBot, Claude-User and ClaudeBot with the same three roles, and Perplexity documents PerplexityBot for search and Perplexity-User for user actions. Google uses Googlebot for AI features; Google-Extended is a separate robots.txt token that limits AI training and grounding in some of its other systems. Being cited and being trained on are now separate decisions, and robots.txt can express them separately:

# Allow search and user-initiated fetches; decline model training.
User-agent: OAI-SearchBot
User-agent: ChatGPT-User
User-agent: Claude-SearchBot
User-agent: Claude-User
User-agent: PerplexityBot
Allow: /

User-agent: GPTBot
User-agent: ClaudeBot
User-agent: Google-Extended
Disallow: /

Two caveats from the vendor pages. OpenAI and Perplexity both note that their user-initiated fetchers (ChatGPT-User, Perplexity-User) act on behalf of a person and may not apply robots.txt rules. And Google-Extended does not affect AI Overviews or AI Mode; those follow the ordinary Googlebot rules.

Side by side

Which one to prioritise

The terms stack, so the order is fixed even though the effort split is not. Work through them in sequence and stop at the first gap.

1. Start with SEO if the page is not indexed or does not rank for its core query. No answer engine will retrieve a passage from a page it cannot fetch or does not trust. Fix crawl access, rendering, canonicals and thin content first.
2. Move to AEO once the page ranks but is paraphrased rather than quoted. Put a 40–70 word direct answer at the top, give each question its own URL and heading, name entities in full, and emit Article, FAQPage and Person markup from the same fields as the visible text.
3. Move to GEO once passages are extracted but rarely cited. Add what the GEO paper found generative engines reward: outbound citations to primary sources, quotations from named authorities and concrete statistics with a source. Check crawler access for the search-mode user agents above.
4. Measure at every stage in the unit that stage rewards. Rank for SEO, snippet or answer ownership for AEO, and citation share across a fixed question set for GEO.

The names matter less than the sequence. A team that ships crawlable pages, answer-first passages and cited evidence is doing SEO, AEO and GEO at once, whatever it calls the work.