What's the difference between SEO, AEO, and GEO?

You have probably seen a proposal with three line items on it — SEO, AEO, GEO — from three vendors, or from one vendor billing three times. You have probably also read a Google document saying all three are the same thing and you can stop worrying about it.

Both are sales positions. The useful version sits between: most of the craft is shared, and exactly three things are not. Name those three and you will neither overpay for a rebrand nor miss a real change in how buyers find you.

The definitions, without the pomp

• SEO — search engine optimization. Earning a ranked position on a results page. The user sees a list; the job is to be high on it.
• AEO — answer engine optimization. Being the extracted answer: featured snippets, People Also Ask, voice results, the direct-answer surfaces in chat interfaces. A list may still sit below you, but you have been lifted out of it.
• GEO — generative engine optimization. Being named, summarized, and recommended inside generated prose, where there may be no ranked list at all. Either your name is in the paragraph or it is not.

The third is where the real difference lives. A ranked list is a menu. A generated paragraph is a recommendation. They fail differently, so they have to be measured differently.

Google says it is all still SEO, and Google is mostly right

In May 2026 Google Search Central published guidance on optimizing for generative AI features. The framing sentence is not ambiguous.

From Google Search's perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO.

Google Search Central, “Optimizing your website for generative AI features on Google Search,” May 2026

Note the qualifier: from Google Search's perspective. Google benefits from a story in which nothing has changed and no new vendor category is warranted. That does not make the sentence wrong. It makes it a sentence with a motive.

Separate the framing from the claims, because the claims hold up. The document debunks five myths: no llms.txt or special AI markup is needed, content does not need chunking for machines, no AI-specific writing voice is required, manufactured brand mentions get filtered by spam systems, and “structured data isn't required for generative AI search.” Outside evidence supports all five. Ahrefs examined 137,210 domains using May 2026 data and found 97% of the 38,360 valid llms.txt files it located got zero requests.

Difference one: the win condition

SEO wins a position. AEO wins the extraction. GEO wins the recommendation. Three separate events, and a page can win the first while losing the other two indefinitely.

A February 2025 Ahrefs analysis of 1.4 million ChatGPT prompts compared cited pages against uncited ones. The clearest separator was semantic similarity between page title and prompt: 0.602 for cited pages against 0.484 for uncited. The same analysis found 88.46% of ChatGPT citations came from ordinary search results rather than Reddit, YouTube, or academic sources.

Read those numbers together. Ordinary search results are the ticket into the room — the 88.46%. Inside the room, position stops deciding and topical match takes over — the 0.602. Ranking is necessary and not sufficient. Treating rank as the outcome rather than the entry requirement stops describing what happens after retrieval, which is where the result now gets decided.

Difference two: the unit of competition

SEO competes for a query. You know what it is, a tracker can see it, and you can watch your position move.

GEO does not. Google documents a technique it calls query fan-out: rather than running your question once, the system issues multiple related sub-searches across subtopics and assembles one answer from the results. That is also why AI Overviews and AI Mode cite a broader, more varied set of links than a classic results page.

You never see those sub-queries. They are generated at answer time, they shift with phrasing and context, and no rank tracker reports them. The Ahrefs data hints at how much they matter: cited page titles matched fan-out-style sub-queries even more closely (0.656) than the original prompt (0.602). You are scored against questions nobody showed you.

The consequence is not mystical. Cover the neighborhood around a question — follow-ups, comparisons, objections, the is-it-worth-it variant — because some of those are the sub-queries. And accept that a rank report is now a partial description of where you stand.

Difference three: what you can measure

Rank is well behaved. It exists whether or not you look at it, it moves slowly enough to trend, and two tools checking the same keyword broadly agree.

Citation is not. There is no citation index to query. You observe it by sampling — running prompts, recording who gets named, repeating on a schedule — and what you sample is documented as volatile. Semrush tracked over 100 million citations across 230,000-plus prompts for 13 weeks in 2025, July 14 to October 12, and watched ChatGPT's Reddit citation share fall from roughly 60% to about 10%, Wikipedia from around 55% to under 20%. Google's AI Mode and Perplexity stayed comparatively stable.

Nothing those sites did caused that; a vendor changed a source policy. A metric that can move fifty points on someone else's product decision needs a different cadence and tooling than rank tracking: more prompts, more engines, more repetitions, less confidence in any single reading. Single-prompt wins reverse often enough that a Semrush and Growth Memo study of 1,094 categories, January to June 2026, set the bar for real ownership at four of five related prompts. It is why GetFound3 scores the three separately.

What is actually shared, which is most of it

Now the correction. Strip away the three scoreboards and the work underneath is close to identical.

• Crawl access. If the crawler is blocked in robots.txt, all three fail at the same line of the same file.
• Indexing. Google states that to appear as a supporting link in AI Overviews or AI Mode, a page must be indexed and eligible for Google Search.
• Speed and rendering. A page a crawler cannot render is a page nothing can quote.
• Information architecture. Clear titles, readable URLs, one topic per page. Ahrefs found natural-language slugs cited 89.78% of the time against 81.11% for opaque ones — advice SEO has given since 2005.
• Writing something true and specific. There is no citation to extract from a sentence that could have run anywhere.

Roughly the same craft, three scoreboards. Anyone selling a wholly new methodology is selling the scoreboard as though it were the craft.

Why the split matters commercially

This would be academic if clicks were holding steady. They are not.

The first is a SparkToro and Similarweb clickstream study, which notes its methodology shifted between measurement years — trust the trend more than the decimal. The rest is Chartbeat data reported by Axios in March 2026. The shape is obvious: the value of being the answer rises precisely as the value of the click falls. If sessions from organic search are your only scoreboard, you are measuring a shrinking share of the outcome, and you will read a real gain in influence as a loss.

What a team actually does differently for each

For SEO

• Keep the technical work unchanged: sitemaps, internal linking, redirect hygiene, Search Console coverage, titles matching intent.
• Add Bing Webmaster Tools if you never have. Several answer engines retrieve from indexes other than Google's, and Bing is the cheapest to be present in.
• Weight rank reporting toward queries that still produce clicks — transactional and navigational — over informational ones that now resolve without a visit.

For AEO

• One question per page, the question as the H1, a self-contained 40-to-60-word answer in the first paragraph. If it only makes sense after two paragraphs of setup, it cannot be lifted.
• Structure for extraction: question-shaped H2s, tables for comparisons, ordered lists for sequences. Those are shapes an extractor can take whole.
• Audit the snippets and People Also Ask entries you own or keep losing — unlike citations, those are visible in an ordinary results check. Note that Google removed FAQ rich results entirely on 7 May 2026.

For GEO

• Build a prompt panel: 20 to 40 real buying questions, run monthly across the engines your buyers use, recording whether you are named and linked.
• Score mention share rather than position — what fraction of answers name you, across related prompts rather than one lucky win.
• Write the adjacent questions, comparisons, and objections, because that is what fan-out retrieves against.
• Work on third-party surfaces. One analysis found most AI citations come from earned media rather than brand-owned pages — directional rather than precise, but consistent across studies.
• Watch referrals carrying utm_source=chatgpt.com. It is your only first-party evidence of being cited, and it undercounts.

When the distinction matters, and when it is vendor theater

The unhedged version. It is theater when someone sells GEO as a new technical discipline with its own file formats, markup, and writing voice — Google is right about those five myths. It is theater when the deliverable is a rewritten robots.txt and an llms.txt file that almost nothing ever requests. It is theater when the pitch depends on you not knowing that 88.46% of ChatGPT citations came out of ordinary search results.

It matters when the question is what you aim at and what you count. A team holding only a rank report cannot see whether it is being recommended, cannot see the sub-queries it is losing, and keeps optimizing a number that describes less of the outcome every quarter. No amount of “it's still SEO” closes that gap. One craft, three scoreboards, diverging fast enough that you need all three on the wall.

1. Open your robots.txt. Confirm Googlebot, bingbot, OAI-SearchBot, PerplexityBot, and Claude-SearchBot are allowed.
2. List your ten highest-intent buying questions. Check whether each has a page whose title is that question.
3. On those pages, move a self-contained 40-to-60-word answer into the first paragraph.
4. Run those ten questions as prompts in ChatGPT, Gemini, and Perplexity. Who gets named? That sheet is your GEO baseline.
5. Add a referral segment for utm_source=chatgpt.com.
6. Report three numbers side by side each month: rank, extraction, mention share.