SEO vs GEO: What Actually Changes in AI Search?
Most of what works in AI search is ordinary SEO. This article shows where the two overlap, where retrieval and citation add new steps, and what is still unproven.
The question behind "SEO vs GEO" is usually a practical one: does a site owner who already does technical SEO well need to do anything different for AI answer engines? The short answer is that the foundations are identical, the retrieval step introduces some new considerations, and the citation step is where the evidence runs out. This article walks the pipeline from crawl to citation and marks, at each stage, what is standard, what is emerging, and what is guesswork.
Two products, one supply chain
A conventional search engine crawls pages, indexes them, and for a query returns a ranked list of documents. The user chooses which to open. The engine's job ends at the click.
An AI answer engine crawls pages, indexes them, and for a prompt retrieves a handful of passages, feeds them to a language model, and returns generated prose. Some providers attach citations; some attach links; some name sources without linking. The user may never click anything.
The upstream half of both pipelines is the same infrastructure. Google says this outright: to be eligible to appear as a supporting link in AI Overviews or AI Mode, a page must be indexed and eligible to be shown in Google Search with a snippet, and there are no additional requirements or special optimisations. OpenAI documents a dedicated search crawler, OAI-SearchBot, and states that sites opted out of it will not be shown in ChatGPT search answers. Anthropic and Perplexity document equivalent search crawlers. In every case, the page has to be crawlable and indexable before anything generative happens.
The downstream half differs. Ranking a document list and selecting passages to ground a generated answer are different operations, and it is the second one that GEO tries to influence.
The pipeline, stage by stage
| Stage | Conventional search | AI answer engine | What changes for the site owner |
|---|---|---|---|
| Crawl | Googlebot, Bingbot | Provider search crawlers plus user-triggered fetchers | More user agents to make policy for; same robots.txt mechanism |
| Index | Document-level index | Document index plus passage-level retrieval index | Passages become the unit of selection |
| Query understanding | Keyword and intent analysis | Prompt interpretation, often expanded into several sub-queries | One prompt may trigger several searches |
| Selection | Ranking documents | Retrieving passages, then the model deciding what to use | Being in the top ten is no longer the target; being the best passage for a sub-question is |
| Presentation | Title, URL, snippet | Generated prose with optional citations | Attribution is at the provider's discretion |
| User action | Click | Read, possibly click a citation | Traffic and visibility decouple |
The table understates one thing: the model's choice at the selection stage is not deterministic. Two identical prompts to the same model minutes apart can retrieve the same passages and cite different ones. That is not a bug to be optimised away; it is a property of the system that measurement has to account for.
What remains standard SEO
Everything a technical SEO audit already checks is a precondition for AI visibility, and there is first-party documentation for each item.
Crawlability. Robots.txt is governed by RFC 9309. A crawler may treat a 4xx response as "no rules" but must treat a 5xx as full disallow, so a broken robots.txt is a crawl blocker. The provider crawler docs from OpenAI, Anthropic and Perplexity all state that their search crawlers honour robots.txt. ESTABLISHED STANDARD and OFFICIAL PROVIDER GUIDANCE.
Indexability and canonicalisation. A noindex page or one whose canonical points elsewhere cannot be a supporting link in Google's AI features, by Google's own eligibility rule. OFFICIAL PROVIDER GUIDANCE.
Snippet controls. Google says that nosnippet, data-nosnippet, max-snippet and noindex limit what its AI features show from a page. These are the same directives that have governed Search snippets for years. OFFICIAL PROVIDER GUIDANCE.
Content quality and originality. Google's helpful-content guidance asks whether a page offers original information, reporting, research or analysis, and whether it is self-evident who wrote it. Its spam policies prohibit scaled content abuse, defined as generating many pages primarily to manipulate rankings rather than help users. Nothing about AI features relaxes this. OFFICIAL PROVIDER GUIDANCE.
Structured data. Schema.org markup is a published standard for describing entities on a page. Google states that no special structured data is required to appear in AI features, so the correct framing is that schema helps machines understand what the page is about, and is a well-documented part of Search, not an AI-specific lever. ESTABLISHED STANDARD for the markup.
Internal linking, headings, titles, descriptions. All of these govern whether a crawler discovers and understands the page and are unchanged. ESTABLISHED STANDARD in the sense that they are settled practice documented by every search engine.
If a site fails any of these, the GEO conversation is premature. The SEO audit exists to find these failures first.
What retrieval changes
Retrieval is the step where an AI system, given a prompt, pulls candidate passages from an index. This is the part of the pipeline that most resembles something new, and it is where the EMERGING PRACTICE tier lives.
The passage, not the page
A retrieval index generally stores chunks of a page rather than the whole page, and scores each chunk against the query. The consequences follow from that design, not from any provider announcement:
- A section whose heading names the question and whose body answers it completely is more likely to be retrieved intact than an answer spread across three sections.
- A passage that depends on context from earlier in the page ("as shown above", "this approach") is weaker in isolation than one that stands alone.
- Extremely long sections dilute the match; extremely short ones carry too little to be selected.
None of this is documented by a provider as a rule for their product. It is inference from how retrieval systems are built, corroborated by practitioners, and unconfirmed. That is the definition of EMERGING PRACTICE. Passage-level content structure covers the specifics.
Query fan-out
Google describes its AI features as using "query fan-out", issuing several related searches to assemble an answer, and says this surfaces a wider and more diverse set of links than a single search would. The practical reading is that a page can be retrieved for a sub-question the user never typed. Pages that answer specific, narrow questions well have a route into answers that a broad page on the same topic does not. This is OFFICIAL PROVIDER GUIDANCE for the mechanism and EMERGING PRACTICE for the tactical inference.
Entities
A retrieval system that cannot tell whether "Acme" on your page is the same Acme as the one in the question will not confidently connect them. Stating the entity plainly in text, using one canonical name, and describing it consistently on an about page and in Organization markup is what entity clarity means. The markup part is standard; the effect on retrieval is plausible and unconfirmed.
What citation changes
Citation is the step where the model, having used a passage, attributes it. This is where the evidence is thinnest, because no provider documents how their model decides what to cite.
What is documented is the shape of the output. OpenAI's web search tool returns url_citation annotations on the response, each with a URL, a title and the character range of the text it supports, and OpenAI states that inline citations must be made clearly visible and clickable when displayed to end users. That tells you what a citation is, structurally, in one provider's API. It does not tell you how to earn one.
Some inferences are reasonable at the EMERGING PRACTICE tier:
- A passage that contains a specific, verifiable claim (a number, a date, a definition) gives the model something to attribute. Vague prose gives it nothing that needs a source.
- A page that is the origin of a fact is a more natural citation than a page that repeats it.
- A page whose author and organisation are clearly stated is easier for a model to describe as a source, and easier for a user to trust when it does.
And some claims are not supportable at any tier above HYPOTHESIS:
- That a particular phrasing pattern, sentence length or "AI-friendly" writing style causes citation.
- That any file or tag addressed to language models is read at citation time.
- That being cited once means being cited reliably.
The distinctions have to survive: retrieved is not cited, and cited is not recommended. A model can name your page as a source for a fact while recommending a competitor in the same answer.
Side by side: the same task in each discipline
| Task | SEO framing | GEO framing | Tier of the GEO framing |
|---|---|---|---|
Fix a noindex on a key page |
Restores indexing | Restores eligibility for AI features | OFFICIAL PROVIDER GUIDANCE |
| Write a title tag | Click-through from a result list | Largely irrelevant to a generated answer; still governs indexing | OFFICIAL PROVIDER GUIDANCE (SEO side) |
| Add an FAQ section | Possible rich result (now limited) | Question-shaped passages for retrieval | EMERGING PRACTICE |
| Publish original data | Links and E-E-A-T | A fact worth attributing | EMERGING PRACTICE |
Add Organization JSON-LD |
Knowledge panel eligibility | Entity disambiguation | ESTABLISHED STANDARD (markup) |
Publish llms.txt |
No effect | Unconfirmed; no documented consumer | EXPERIMENTAL |
Block GPTBot |
No effect on Search | Opts out of OpenAI training; does not affect ChatGPT search | OFFICIAL PROVIDER GUIDANCE |
Block Google-Extended |
Google states no effect on Search inclusion or ranking | Opts out of Gemini training and grounding | OFFICIAL PROVIDER GUIDANCE |
The last two rows are worth stating carefully because they are often confused. OpenAI documents GPTBot for training and OAI-SearchBot for search, as separate tokens with separate consequences. Google documents Google-Extended as a token governing whether crawled content may be used for training Gemini and for grounding in Gemini Apps and Vertex AI, and states that it does not affect a site's inclusion in Google Search nor act as a ranking signal. Both are OFFICIAL PROVIDER GUIDANCE. Neither statement says anything about AI Overviews specifically, and this article does not either.
How measurement differs
In SEO, a rank tracker reports a position for a keyword, and although positions vary by location, device and personalisation, the result list is an ordered thing and the position is real.
In AI search, there is no ordered list. The honest measurement is a set of controlled observations: a fixed prompt, sent to a named model under recorded conditions, several times, with each response stored and checked for mentions and citations. From those, a mention rate and a citation rate can be computed with a confidence interval, and repeated runs of the same prompt reveal how volatile the answer is. What cannot be computed is a rank, and any product that reports one for a generative answer has abandoned the distinction between an observation and an invention.
A further caveat: an API response is not what a person sees in a chat product. The consumer product has its own retrieval stack, orchestration and personalisation. Measuring through the API is the reproducible option; it is not the same as watching a user's screen. The methodology page sets out exactly what SiteRank AI records and why.
What not to assume
- That doing GEO lets you skip any part of technical SEO. The eligibility rule for Google's AI features is the Search eligibility rule.
- That a page visited by a search crawler is in the retrieval index, or that one in the index will be retrieved for a given prompt.
- That any practice at EMERGING PRACTICE is a signal, a factor or a requirement. It is a plausible practice with no confirmation.
- That the volatility of generated answers is a measurement error. It is the thing being measured.
- That a chat product's behaviour matches its API.
Key takeaways
- The crawl-and-index half of AI search is conventional SEO, documented by the providers, and non-negotiable.
- Retrieval introduces the passage as the unit of selection and query fan-out as a route in. Both support EMERGING PRACTICE tactics, not requirements.
- Citation is undocumented by every provider. Reasonable inferences exist; guarantees do not.
- Measure AI visibility as observed rates with intervals, never as a position.
Official sources & further reading
- AI features and your website — Google Search Central
- Creating helpful, reliable, people-first content — Google Search Central
- Spam policies for Google web search — Google Search Central
- Google's common crawlers, including Google-Extended — Google Search Central
- Overview of OpenAI crawlers — OpenAI
- Web search tool — OpenAI
- Does Anthropic crawl data from the web, and how can site owners block the crawler? — Anthropic
- Perplexity crawlers — Perplexity
- RFC 9309: Robots Exclusion Protocol — IETF
Related reading
Frequently asked questions
My technical SEO is already in good shape. What is left to do?
Three things, in descending order of certainty: set a deliberate policy for each provider's crawlers, structure pages so a single section answers a single question without leaning on its neighbours, and start measuring what the answer engines say about you. Only the first has provider documentation behind it. The other two sit at EMERGING PRACTICE and are worth doing largely because they also make pages easier for a person to read — they are not worth rebuilding a site over. Measurement is the part most sites skip, and it is the only way to find out whether the other two did anything: see AI visibility monitoring.
Do AI answer engines use Google's index?
Google's own AI features do, in the sense that its eligibility rule for them is the ordinary Search eligibility rule. The other providers described here document their own search crawlers, which is why opting out of OAI-SearchBot has a consequence that opting out of Googlebot does not, and none of them documents how its index is built or how passages are chosen from it. Treat "indexed by Google" as a precondition for Google's surfaces and as no information whatsoever about anyone else's.
Should I rewrite pages into a question-and-answer format for AI?
Write sections that resolve one question completely and can stand alone, which is a structural change rather than a stylistic one. Mechanically converting a page into a list of questions is the old keyword-stuffing instinct in new clothes, and Google's spam policies on scaled content abuse do not become lenient because the intended reader is a machine. The passage argument is about self-containment: a heading that names the topic, a body that answers it without "as shown above". If the rewrite reads worse to a person, it is not the change the retrieval argument asked for.
If two identical prompts give different answers, how can I tell whether a change worked?
By sampling, not by asking once. One answer before and one answer after tells you nothing, because the variation between two runs of the same prompt can easily exceed anything your change did. The workable method is a fixed prompt set, repeated under recorded conditions, reported as rates with confidence intervals, and compared only while the model and the prompt wording hold still — the mechanics are in how LLM visibility monitoring actually works. Even then the finding is association: a rate that moved alongside your change, not evidence that your change moved it.