SEO competes for a position on a page of ten results. AEO competes for inclusion in a single synthesized answer, where there is no second page and usually no more than five named options.
Almost every technical fundamental carries over unchanged. What does not carry over is the scoreboard: position, impressions and clicks describe a page of links, and an answer is not a page of links.
The unhelpful version of this comparison says SEO is dead and AEO replaces it. That is wrong in a way that costs money, because a team that believes it stops doing the work that gets them into the answer in the first place. The useful version is narrower: the inputs mostly stay, the output changes shape, and the measurement has to be rebuilt from scratch.
What stays exactly the same?
More than the category’s marketing would suggest. Google is explicit that its AI surfaces run on the same eligibility rules as Search: a page qualifies if it is indexed and can be shown with a snippet, and there are “no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary”.
If a crawler cannot fetch it, nothing downstream matters. Still true, now with more crawlers to think about.
Google’s AI features draw on the same index. A page that cannot rank cannot be summarised either.
Google’s guidance still asks whether “pages carry a byline, where one might be expected”. Unchanged and still under-done.
The helpful content questions are the same questions. Nothing about answers made thin content work.
One thing does get quietly harder, and it is worth flagging because it looks like a technical detail rather than a strategy question. There used to be one crawler to think about. There are now three classes of them, and they are configured separately.
OpenAI documents that a webmaster “can allow OAI-SearchBot in order to appear in search results while disallowing GPTBot”, and that the settings are independent of each other. Perplexity draws the same line between the crawler that surfaces and links sites and the fetcher that acts when a user asks. A robots.txt written in 2023 to keep content out of model training can, if it was written broadly, also be keeping you out of the answers. That is an SEO-shaped task with an AEO-shaped consequence.
The rules for resolving those directives have been formally specified since 2022. RFC 9309 defines that the most specific matching rule wins and that an allow beats an equivalent disallow, which is the opposite of what most people assume, and which is worth knowing before you edit a file that decides whether four engines can see you.
What actually changes?
Six things, and the first two cause the other four.
| SEO | AEO | |
|---|---|---|
| What you compete for | A position among ten results | Inclusion in one answer, usually among five or fewer named options |
| Where the evidence lives | Largely your own domain, plus links to it | Largely other people’s domains, summarised |
| Result stability | Near-deterministic, trackable daily | Varies between identical runs of the same prompt |
| Unit of measurement | Position, impressions, clicks | Citation rate, mention rate, share of the named set |
| Consolation prize | Position eight still exists and still gets clicks | There is none. You are named or you are not |
| Feedback loop | Analytics ties the visit to the query | An answer can influence a buyer and leave no trace at all |
The absence of a consolation prize is the change that matters most. Ranking eighth is a bad quarter. Not being in the answer is not being in the market.
Why can the same tactics stop working?
Because a tactic optimised for a result format can outlive the format. The clearest recent example is one most SEO teams spent real hours on.
Google removed the HowTo rich result in September 2023, noting it “is no longer shown in search results, on both desktop and mobile devices”. FAQ rich results were narrowed to authoritative government and health sites, then retired. Neither type appears in the current rich results gallery.
Question-shaped headings with self-contained answers are still exactly what a retriever wants, because retrieval works on passages. The rich result died. The reason for the format did not.
The lesson generalises. Markup that describes your page honestly keeps its value because structured data helps search understand the content, which is useful to any system reading the page. Markup pursued for a badge in a results layout is worth exactly as long as the layout.
Does keyword targeting still work?
Partly, and less than it did, for a mechanical reason rather than a philosophical one.
Classical retrieval scored on shared words. Modern retrieval scores on learned embeddings, and Karpukhin et al. showed dense retrieval beating a strong BM25 baseline by 9 to 19 points on top-20 passage accuracy. When a passage can be retrieved without containing the query’s words, repeating a phrase buys less than it used to.
What replaces it is not a trick. It is being the thing the question is about, described consistently enough that an engine can tell it is you.
That has a consequence people find uncomfortable. In SEO you could win a query you had no business winning, by covering a topic thoroughly enough that the algorithm rewarded the page regardless of who published it. In AEO that route is much narrower for buying questions, because the engine is not asking which page is best on this topic. It is asking which companies belong in the answer, and that judgement is formed from what the wider corpus says about you rather than from any single page you wrote.
You can still win informational questions with excellent content, which is most of what this blog is doing. What you cannot do is write your way into a recommendation list.
Does any of this apply outside Google?
The technical fundamentals do, and the specifics do not, which is why treating “AI search” as one thing goes wrong.
Google’s surfaces run on the Google index with Google’s eligibility rules, so classical SEO work is close to a complete description of what qualifies you. The assistant surfaces are separate stacks with their own crawlers and their own corpora, and there is no equivalent of Search Console to tell you how you are doing in them. The only way to know is to ask them, repeatedly, and record what comes back.
That asymmetry is the practical reason AEO needs its own measurement rather than an extra column in an SEO report. One half of the landscape publishes diagnostics about you. The other half will only answer questions.
What happens to the page itself?
One structural change is worth more than every other on-page adjustment combined, and it follows directly from how retrieval works.
Search engines index pages. Retrievers score passages. That means the unit competing for inclusion is not your article, it is the two hundred words under one of your headings, judged alone and without the three sections that came before it.
“As we saw above, this approach has three drawbacks.” Pulled out on its own it is meaningless, so it is rarely pulled out, and when it is, it cannot be quoted.
A question-shaped heading, then an answer that makes sense to somebody who arrived at that paragraph and nothing else. Self-contained is the whole requirement.
Notice this is also just good writing, and it was good writing before any of this. Google’s snippet documentation has always described snippets as generated automatically from page content, which rewards the same self-contained structure. The format did not become valuable because of AI. AI raised the stakes on getting it right.
How does the work itself change?
Crawl access, indexability, structure, markup. Roughly unchanged, with more crawler classes to configure.
Carries overReview platforms, comparison pages, community presence. Was a link-building activity, is now the primary input.
Much heavierStill worth doing, still the thing teams over-invest in. It amplifies evidence rather than creating it.
Same effort, lower shareWhat breaks in your reporting?
Two things break, and neither has a workaround.
Attribution. An answer can name you, persuade someone, and send no click. That visit arrives later as direct traffic with nothing tying it back. Any vendor claiming to close that gap is inferring rather than measuring.
Position. There is no position. Being named third of four is not third place, because a different run of the same prompt names you first, or not at all. Reporting a rank against an answer is reporting a number the format does not contain.
Deterministic, comparable week to week, attributable to a page. None of those properties survive in an answer.
A rate over a fixed prompt set and a fixed run count. Reproducible, and it degrades gracefully when the engines change under you.
So should you stop doing SEO?
No, and the question is usually a budget argument wearing a strategy costume.
On Google’s surfaces the entry requirement for AI features is ordinary search eligibility, so stopping SEO removes you from the AI results too. On the assistant surfaces, the retrieval is different but the raw material is the same open web. A site that is fast, crawlable, well structured and genuinely useful is the input to both.
What should change is the proportion, the measurement, and the expectation. Less of the budget on your own domain, a scoreboard rebuilt around rates instead of positions, and an acceptance that some influence will never show up in analytics.
Who should own it?
Usually the same team, and the argument for splitting it is almost always an org-chart argument rather than a technical one.
The overlap is large enough that separating them duplicates work: both need crawl access, both need clean structure, both need entity consistency, both need somebody who can read a log file. What genuinely differs is the third-party evidence layer, which looks less like content production and more like the work a PR or partnerships function already does.
The practical failure mode is a team that treats AEO as a content brief. It is not one. The single highest-value activity in it happens on domains you do not own, and no editorial calendar reaches there.
What do people ask most about the difference?
If you want the vocabulary straight before you rebuild any reporting, the glossary defines the sixty terms this field actually uses, and marks the four that decide everything else.