"AEO vs SEO" is usually argued with vibes. We tested it with data. In August 2026 we took 45 real queries a site ranks for on Google, all at position 12 or better, and measured what AI engines actually did with those pages. The result is the clearest picture we have of where SEO stops and AEO begins.
The numbers: across those 45 queries, a live web-search index surfaced the site on 27 (60 percent), a browsing AI model cited it on 7 (15.6 percent), and in a separate 9-query check, a non-browsing model answering from training memory named it 0 times. Ranking bought retrieval. It did not buy the citation. The full dataset, methodology, and per-query table are published in the ranking vs AI citation study.
Is AEO actually different from SEO?
Yes, because getting cited by AI is two gates, not one, and each gate rewards different work.
| Dimension | SEO (gate 1: retrieval) | AEO (gate 2: the citation) |
|---|---|---|
| The question it answers | Can a search index surface your page as a candidate? | Does the model quote or name your page in the answer? |
| What moves it | Rankings, crawlability, links, technical health | Answer phrasing, self-contained facts, format, mention density |
| Unit of optimization | The keyword and the SERP | The question and the quotable passage |
| Measured by | Rank trackers, Search Console | Share-of-voice checks, citation tracking |
| What our study showed | 60% of queries surfaced the ranking pages | 15.6% cited when browsing; 0/9 recalled without browsing |
The gates are sequential. Fail gate 1 and gate 2 never happens: a model cannot cite a page its retrieval layer never saw. But passing gate 1 is where most SEO advice stops, and the drop from 60 percent to 15.6 percent is the part ranking cannot fix.
Does ranking on Google get you cited by AI?
Not in any predictable way. If ranking drove citation, the cited queries in our study would cluster at the best positions. They did not. The seven cited pages ranked at positions 2.0, 3.1, 5.1, 5.3, 5.4, 6.8, and 8.7. Meanwhile pages ranking at 2.2, 2.7, 3.1, and 3.2 went uncited. One page the browsing model cited was not even surfaced by the live web index for that query; the model found it another way.
The pattern inside the misses was more useful than the hit rate. The pages that lost tended to phrase their key numbers as a vendor's price rather than a neutral market fact. To a model choosing one source to quote, "our sprint costs $X" reads like a pitch; "a typical fixed-scope build runs $15,000 to $60,000 in 2026" reads like an answer. Same information, different liftability.
What content formats actually get cited by AI?
Working from what won citations in our study and what the browsing models quoted instead of us, the pattern is consistent: models lift finished answers, not raw material. Concretely:
- A direct answer in the first hundred words. The quotable version of the page's core fact, stated up front as a neutral market figure with real numbers, before any product framing.
- Comparison and roundup formats. "X vs Y" pages, "best tools compared" tables, and cost breakdowns give a model exactly the structure an answer needs. This is why comparison content dominates AI citations across the queries we tested, and, transparently, why this page is shaped the way it is.
- Question-shaped headings. H2s that match how buyers actually phrase the question give retrieval something to match and give the model a labeled passage to lift.
- Self-contained stat sentences. A sentence that survives being quoted alone ("across 45 queries, browsing AI cited the ranking site 15.6 percent of the time") is liftable. A sentence that depends on the paragraph above it is not.
None of this is a trick. It is writing the answer the model is looking for, then being the most citable version of it in the retrieved pool.
Why do non-browsing models never mention you?
Because parametric memory is earned over years, not ranked into. A model with no web access answers from what it absorbed in training, which reflects how often and how widely a source is mentioned across the web. One strong page, however well it ranks, does not register. Our 0-of-9 no-browsing result came from a site ranking top-12 on Google for every tested query.
This is the recall gap, and it is closed by mention density, not on-page work: genuine roundup inclusions, real reviews, community threads where the buying question gets answered and your name comes up. Slow, authentic, and cumulative. Manufactured links do not build it and can hurt the ranking gate you already won.
How do you know if any of this is working?
Measure the second gate directly instead of inferring it from rankings. Ask the engines your buyers' questions from clean sessions, or run a structured check: the free AI Visibility Check puts buyer-intent questions in your space to four engines live and reports your share of voice, the questions where you are invisible, and the competitors named instead of you. If you want scheduled tracking across engines, the paid tracker landscape is compared here, and if the number comes back low and you want it moved rather than monitored, that is what Continuum does.
One honest caveat to carry with you: this is one site, 45 queries, one pass, in a category where engine behavior shifts monthly. Treat the 60/15.6/0 funnel as a mechanism worth testing on your own queries, not a law. The mechanism, retrieval first, lift second, recall last, is the part we would bet on.