Classic SEO optimizes for ten blue links. GEO — generative engine optimization — optimizes for being quoted accurately when someone asks ChatGPT, Perplexity, or Claude a question your data could answer. The two overlap, but the second has barely-written-down best practices. Here's what we actually changed.
What changes when your reader is a model
An LLM doesn't skim a hero image or click a tab. It ingests text, and it rewards three things: facts stated plainly, clear provenance (who says so, and as of when), and structure it can lift without mangling. A figure trapped in a <canvas> or a chart image is invisible; the same figure in a sentence with a date and a source is quotable. The working rule: write so a careful stranger could quote you and still be right.
Let the crawlers in
You can't be cited if the AI crawlers can't read you. Two files do most of the work. A robots.txt that explicitly admits them:
User-agent: GPTBot
Allow: /
User-agent: ClaudeBot
Allow: /
User-agent: PerplexityBot
Allow: /
User-agent: Google-Extended
Allow: /
And an llms.txt at the root — a plain-Markdown map of what the site is and its key pages, so a model gets an honest summary instead of inferring from nav soup:
# Furiosa Studio — data hub
> Free, sourced datasets across astronomy, football and history.
## Datasets
- [Next eclipse by location](https://next-eclipse.com)
- [Top scorers, 30+ leagues](https://top-scorers.com)
Mark up the facts
Schema.org Dataset + JSON-LD turns a page of numbers into something an engine can attribute:
{
"@context": "https://schema.org",
"@type": "Dataset",
"name": "2026 total solar eclipse — visibility by location",
"creator": { "@type": "Organization", "name": "Furiosa Studio" },
"dateModified": "2026-06-17",
"license": "https://creativecommons.org/licenses/by/4.0/"
}
The creator, dateModified, and license fields are exactly what an engine uses to decide whether a figure is safe to quote and whom to credit.
Endpoints as a citation surface
Prose is for humans; JSON is for machines. Every dataset also has a public, stable JSON endpoint (moons, eclipses, pandemics, top scorers). A model — or a developer building on you — pulls the raw figure instead of scraping a rendered table, so it quotes the number you published, not a hallucinated approximation. Stable URLs plus a documented shape are a clean citation surface.
Hub and spoke
The leverage move: one umbrella hub (furiosadata.com) links every data site, and every site links back to the hub. To a search or generative engine this reads as one coherent entity rather than eight unrelated pages — so authority earned by any one spoke lifts all of them. It's the internal-linking logic that's always worked, applied at the portfolio level.
What we'd do differently
GEO is new and the feedback loop is slow: you don't watch a ranking move day to day, you watch whether engines start quoting you over months. Two honest lessons. First, structured data only helps if the prose already states the fact clearly — schema is a label, not a substitute for saying the thing. Second, llms.txt is still a convention, not a standard, so treat it as cheap insurance, not a guarantee anyone reads it.
This is how we built furiosadata.com, the Furiosa Studio data hub that catalogs and cross-links the whole portfolio.
Built by Furiosa Studio.
Top comments (0)