Search & AI

GEO isn't SEO with a new name. Here's the actual difference.

Most agencies added "GEO" to the pricing page and changed nothing about the work. That is a problem, because the two disciplines optimise for different outcomes.

SEO competes for a position. GEO competes to be the source.

An SEO win is a rank. Position three on a query with volume, and a percentage of that volume clicks through. The mechanism is a list, and your job is to be high on it.

A GEO win is a citation. A model reads your page, decides it is the most reliable statement of a fact, and reproduces that statement in an answer — usually with a small link nobody clicks. You get no traffic. You get something arguably better: you become the thing the buyer now believes.

Three things that change in practice

1. Passage-level clarity beats page-level authority. A model does not lift your page. It lifts a paragraph. So the unit of work drops from "the page ranks" to "this specific paragraph is the cleanest available statement of this specific fact." Long preambles, throat-clearing and SEO-padding actively hurt you here — they make the quotable bit harder to isolate.

2. Being contradicted is worse than being absent. In classic SEO, a competitor outranking you costs you clicks. In GEO, a competitor's page being cleaner than yours means the model repeats their framing of the category to your buyer. You are not losing traffic; you are losing the definition.

3. Off-site mentions carry more weight than links. Models are trained and grounded on text, not on a link graph. A detailed, specific mention of your brand in a forum thread, a review site or a comparison article can influence an answer even with no link at all. This is closer to old-fashioned PR than to link building — and most SEO teams are not staffed for it.

How we audit it

We run a fixed set of buying-intent prompts across ChatGPT, Gemini, Perplexity and Google's AI Overviews — the questions a real buyer types before a purchase. We record which brands get named, which sources get cited, and what claim is attached to each. Then we work backwards: what would have to be true on the open web for the model to name you instead?

That gap is the brief. It is usually not more blog posts.

The uncomfortable part

Nobody has a reliable, audited attribution model for AI citations yet. Anyone selling you one is selling you a dashboard, not a result. What we can measure is citation share on a fixed prompt set over time, and whether it moves when we change things. That is a weaker signal than a rankings report — and it is still the best signal available.