↳ marketingengineer / essay
GEO: how AI answers pick their sources
Generative engines don’t rank pages, they cite passages. Notes from three months of testing what gets a brand quoted by ChatGPT, Perplexity and AI Overviews.
The search box is splitting in two. One half is still ten blue links. The other half writes you an answer and cites three sources — and if you’re not one of them, you don’t exist for that query.
Getting cited is a different game from ranking, and most of what I believed from years of SEO turned out to be half-right at best. These are my notes from three months of structured testing.
Engines cite passages, not pages
Classic SEO optimises a page to own a keyword. Generative engines retrieve chunks — a paragraph, a table, a definition — and assemble an answer from them. The unit of competition shrank from the page to the passage.
The practical consequence: a page that answers one question completely in one extractable block beats a page that answers ten questions diffusely. We restructured our highest-intent pages so every H2 is a question and the first paragraph under it is a self-contained answer — no “as mentioned above”, no pronouns referring to other sections. Citations from AI Overviews on those pages roughly tripled within six weeks.
What moved the needle
Across the test set, in order of observed impact:
- Being the primary source for a number. Engines love citing concrete figures. Publishing original data — even a small benchmark — earns citations far outside its keyword neighbourhood.
- Extractable structure. Tables, definition-style openers, and step lists get lifted verbatim. Walls of brand prose do not.
- Third-party corroboration. Mentions on review sites, forums and comparison pages seem to function as the trust layer. Reddit threads we never touched cited us more than our own category pages.
- Freshness signals on the source page. Updated-date honesty, not fake-fresh.
What did not measurably move: schema markup beyond the basics, keyword density of any kind, and — painfully — domain authority as classically understood.
Measure it or it didn’t happen
There’s no Search Console for AI answers yet, so we built a panel: 120 high-intent prompts, run weekly through the major engines via API, with citations logged to a sheet. It costs almost nothing and it turns GEO from vibes into a metric you can move.
Run the panel before you change anything. Otherwise you’re optimising for screenshots.