AEO vs SEO in 2026: which one actually drives traffic now

In 2024, "do SEO" meant rank for keywords. In 2026, "do SEO" means also be cited by ChatGPT, Perplexity, Google AI Overview, and Copilot — because for a growing share of searches, the user never sees the blue links at all.

The mechanical difference

SEO optimises for selection: out of ten blue links, does yours get clicked. AEO optimises for inclusion: out of everything the engine retrieved, does your text end up inside the generated answer.

That sounds like a small distinction and it changes almost every tactic downstream, because the unit of competition moves. In classic search the unit is the page — it ranks, or it does not. In an answer engine the unit is the passage. The engine retrieves chunks, assembles an answer from several of them, and attributes each to a source. A page can be retrieved and still contribute nothing, because none of its paragraphs said anything liftable.

This is why a well-ranked page can be invisible in AI answers while a worse-ranked one gets cited constantly. The second one had a paragraph that answered the question on its own.

How the retrieval actually works

Roughly the same shape across engines, and knowing it explains most of the advice:

1. The query is rewritten — often into several sub-queries. A single question becomes three or four retrievals.
2. Candidates are retrieved, usually from a live search index plus the model's own store, and chunked.
3. Chunks are ranked for relevance to the rewritten query, not to the original phrasing.
4. The answer is generated from the top chunks, with citations attached to the passages actually used.

Two consequences follow directly.

Your competition is per-passage. You are not competing with the other nine results. You are competing with every paragraph the engine retrieved, including several from the same page.

Context is not carried. A chunk arrives without the heading three sections up that defined the term. A paragraph that reads perfectly in place and means nothing in isolation will lose to a clumsier one that stands alone.

What actually changed

Three things, in order of impact:

1. Google AI Overview hit ~60% of queries (April 2026 measurement, Sistrix). The first thing the user sees is a generated answer with citations — usually 3-5 sources lifted from the SERP.
2. Perplexity passed 30M monthly active users and quietly became the default search for builders. Citations there compound — being cited once means you keep getting cited because the engine prefers sources it has already deemed authoritative.
3. llms.txt jumped from 0.4% to 11% of top-10K sites in 12 months. The de-facto standard for "tell AI what your site is about" is no longer optional in competitive niches.

What to actually do this quarter

Five concrete moves, in order of ROI:

1. Add a 40-60 word direct-answer block after your first H2. Cited verbatim 4.6x more than long-form intros.
2. Add FAQPage JSON-LD to your top 10 pages. 1.8x lift in Copilot citations alone.
3. Drop a /llms.txt describing what your site is, who runs it, and where the canonical content lives. Picked up on next crawl.
4. Embed 3+ statistical anchors per 300 words. "60% of queries", "2.1x lift", "11% adoption" — concrete numbers are what AI engines copy.
5. Use definition-lead sentences. "[X] is a [category] that [diff]." Picked up as the entity definition for that page.

Writing a passage that survives extraction

The five moves above are the checklist. The underlying rule is one sentence: every paragraph should make sense to someone who read only that paragraph.

In practice that means four habits.

Resolve pronouns and demonstratives. "This makes it faster" is meaningless out of context. "Caching the access token makes each request faster" survives the trip.

Put the claim first, the qualification second. An engine truncating a chunk keeps the beginning. A paragraph that spends two sentences building to its point often gets cut before reaching it.

Attach the number to the claim. "Adoption grew sharply" is unusable. "Adoption grew from 0.4% to 11% of the top 10,000 sites in twelve months" is a quotable sentence, and it carries its own evidence.

Answer the question in its own words. If the query is "how do you calculate CAC payback", a heading that says "The payback question" is worse than one that says "How to calculate CAC payback" — not for keyword reasons, but because the chunk ranker is matching against a rewritten query that looks like the question.

The identity surface

Beyond the prose, a small set of files tells an agent what your site is. All of them sit at fixed paths, all are cheap to produce, and most sites ship none:

llms-full.txt is the one most worth the effort and the one most often skipped. A crawler that has to render fifteen HTML pages to find your definitions will usually not bother; one plain-text fetch that contains every definition, every direct answer and every FAQ makes citing you nearly free. If your articles already carry structured front-matter — a definition sentence, a direct answer, statistical anchors, FAQ pairs — generating it is mechanical.

The per-bot policy deserves a real decision rather than a default. The bots that drive cited-answer traffic (OAI-SearchBot, ChatGPT-User, Claude-SearchBot, PerplexityBot, Google-Extended) are not the same population as the bots that scrape for training with no attribution path. Allowing the first set and denying the second is a coherent position; blanket-allowing everything or blanket-denying everything usually is not.

Where SEO and AEO diverge

Backlinks still help SEO; they barely help AEO. The new currency is entity clarity — does the engine understand what you are, who you serve, and what you say with certainty? FAQ schema, llms.txt, and statistical anchors all push that signal up.

Three more divergences worth knowing:

Freshness reads differently. A dated blog post loses to an updated reference page on the same topic, because the engine is choosing a source to trust rather than a result to show. Evergreen pages with a visible dateModified and no datePublished sidestep the "this is from last year" read entirely.

Position matters less, presence matters more. Ranking fourth instead of first costs a lot of clicks and almost no citations. Being on page two costs everything in both.

Ambiguity is fatal in a way it never was for SEO. Google will happily rank a page whose subject it half-understands. An answer engine that cannot tell what entity your page is about will not cite it, because a citation is an assertion it has to stand behind.

The measurement problem, stated honestly

The uncomfortable part: you largely cannot measure this yet.

Search Console reports impressions and clicks from Google search, and an AI Overview citation that produced no click is not a row you can isolate. Perplexity and ChatGPT report nothing to you at all. Referral traffic from answer engines exists but undercounts badly, because the most valuable outcome — the user got their answer and saw your name — generates no referral.

Three partial substitutes, in increasing order of effort:

1. Ask the engines directly. Query the terms you care about in ChatGPT, Perplexity and Google, and record who gets cited. Manual, but it is ground truth, and a monthly snapshot beats a dashboard nobody can build.
2. Watch branded search volume. If citations are working, people who saw your name go on to search it. Branded volume is a lagging proxy but a real one.
3. Track non-brand impressions on the pages you optimised. Not the same as citations, but AI Overview inclusion and organic impressions correlate, since both draw from the same retrieval.

Anyone selling you a precise AEO attribution number is selling you a model, not a measurement. That does not make the work not worth doing; it makes the case for doing the cheap parts first, since you cannot yet A/B your way to the answer.

The cost either way

The cost of the shift is close to zero new tools: front-matter fields, a plain-text file or two, some discipline about paragraph structure. Most of it is writing habits, not infrastructure.

The cost of not shifting is your competitor's direct-answer block appearing inside the AI answer while your better-researched page sits on page two of the blue links nobody scrolls to.

Want a worked example? See what-is-mcp-server — definition lead, direct-answer block right after the first H2, statistical anchors, FAQPage schema. The full AEO surface in one post. The guides are the same structure applied to evergreen reference pages rather than dated ones.