GEO vs SEO: what actually changed when answers replaced links.
SEO competes for a position in a list of links. GEO competes for words inside a single synthesized answer. That one sentence explains every tactical difference between the two disciplines.
What transfers from SEO
More than purists admit. Retrievability is still ranking: AI assistants read the top search results, so technical SEO — indexing, site speed, canonical hygiene, internal linking — remains the entry ticket. A page that ranks nowhere gets cited nowhere. Authority also transfers: the third-party mentions that built PageRank are the same corroboration signals generative engines cross-check.
What breaks
Keyword-centric content. The GEO-bench study tested keyword stuffing directly and found it ineffective for generative engines — in some configurations, counterproductive.[1] A language model doesn't count keyword occurrences; it evaluates whether a passage is a clean, evidenced, liftable answer. Ten variations of "best GEO agency Paris" scattered through a page signal nothing to a model except low quality.
The click model also breaks. SEO could win with a compelling title and a mediocre page, because the ranking, not the content, earned the visit. In a generated answer there is no click to win — the content itself must be quoted, or the visit never happens.
What's genuinely new
Three things. First, passage-level competition: models cite paragraphs, not domains, so a single well-built paragraph can outperform an entire authoritative site. Second, machine-facing formats: JSON-LD entity markup and llms.txt files speak to the retrieval pipeline directly. Third, probabilistic outcomes: the same question asked twice can produce different citations, which makes measurement — citation rate over many runs, not a single ranking check — the only honest scoreboard.
What to do first
In order: verify AI crawlers aren't blocked; get your money pages indexed and ranking for the underlying searches; rewrite the top of each priority page as a self-contained, evidenced answer; add the structured data; then invest in the slowest, highest-leverage layer — being mentioned where assistants already look. Our full sequence is on the method page.