North Signal

GEO 101

AI visibility, in plain words

No jargon, no hype. What's proven, what isn't, and how to think about it — grounded in the controlled research (full detail on the methodology page).

What is GEO?

Generative Engine Optimization is the practice of getting your brand mentioned, cited, and recommended inside AI answers — ChatGPT, Google AI Overviews, Perplexity, Gemini, Copilot. It's the successor question to SEO: not "do we rank?" but "when a buyer asks AI, are we the answer?"

Why does it matter now?

A growing share of buying research starts as a conversation with an AI instead of a list of blue links. When the AI answers, it names two or three options. If you're not one of them, you were never in the running — and you can't see it happening in your analytics.

What actually works? (the proven stuff)

Three things have real evidence behind them. First: evidence density — concrete statistics, direct quotes from named people, and citations to credible sources make your pages the easiest thing for an AI to quote (this is causally proven in controlled research, with gains up to ~40%). Second: corroboration — AI recommends brands that many independent sources agree about, which is why reviews, communities, and 'best X' lists matter more than your own website (~75% of citations point at third-party pages). Third: clarity — clean, direct, self-contained writing gets retrieved and paraphrased more.

What's overhyped?

Visual formatting has no measured effect — AI reads your content, not your design. Schema markup is hygiene, not a growth lever. Keyword stuffing actively backfires. 'Authoritative tone' by itself does nothing — it's evidence-backed specificity that works. And any tool selling you a single visibility score is selling noise: AI answers change every time you ask.

Why do you keep saying 'confidence intervals'?

Ask ChatGPT the same question ten times and you'll get different answers — sometimes you're mentioned, sometimes not. So 'your visibility is 34%' is meaningless without knowing how many times we asked and how much it varies. We show every number as a range for exactly this reason. It's more honest, and it stops you from chasing changes that are just randomness.

What's a controlled experiment, and why should I care?

It's the only way to know a fix worked. We split your prompts into two groups, you ship your fix, and we compare movement in the treated group against the untouched control group. If both moved, it was the weather (a model update, randomness). If only the treated group moved — significantly — your fix caused it. Marketing tools almost never do this; science always does.

How do I start?

Get a free baseline. Paste your site on the homepage; you'll see where you stand on the questions your buyers actually ask, and the first three things to fix. Everything after that is the loop: fix, measure, prove, repeat.