Is GEO a real discipline or a rebranded SEO service?

Both claims are made and both are partly true. The technical foundation overlaps with SEO by roughly a third: crawlability, valid markup, pages that render. What has no SEO equivalent is the measurement — a fixed panel of buyer prompts run against multiple engines with answers logged verbatim — and the entity work that decides whether a model can verify you at all. If an agency sells GEO without describing how it measures answers, you are buying the rebrand.

Who is GEO for, and who should ignore it?

It matters most where buyers ask for a recommendation before they compare: B2B services, regulated categories, anything with a considered purchase. It matters least for impulse retail and for brands whose demand is entirely branded already. The honest test is to ask an engine the question your buyer would ask. If a competitor comes back, you have a problem worth measuring.

By running the prompts and reading the answers. A panel of category and branded prompts is run live through the engines, each answer stored with the sources it cited, and the same panel re-run on a fixed cadence so you get a series rather than a snapshot. Anything that reports a single score without the underlying answers is not measurement.

Does GEO replace SEO?

No, and treating it as a replacement is expensive. Many of the sources engines cite are themselves ranked pages, so rankings feed citations. The two run alongside each other with separate scoreboards: positions in Google, recall and citations in the engines.

What is generative engine optimization? Definition and numbers

Generative Engine Optimization (GEO) is the practice of making a brand visible, correctly described and cited inside the answers AI engines generate — ChatGPT, Claude, Gemini, Perplexity, Grok, DeepSeek — rather than inside a list of links. It is measured by running the prompts your buyers actually type and reading what the engines answer.

That is the whole definition. Everything below is why it turned into a separate discipline instead of a chapter in an SEO checklist.

the shift that created it

Search used to end with a page of links, and the job was to be one of them. Increasingly it ends with a written answer that names two or three companies. There is no page two, no scrolling past the fold, and no second chance: either the sentence contains your name or it does not.

That changes what optimisation means. Ranking is a competition for position among documents. Being named is a competition for inclusion in a sentence, and it is decided by whether a model can read you, verify you, and find something worth quoting.

the number that makes the case

We audited 100 companies between April and July 2026 and published the findings in full. Two figures from that dataset explain the entire discipline.

93% of brands were known to AI and described accurately when asked about themselves. 4.5% were named when a buyer asked who to pick in their category. Being known and being recommended turned out to be almost unrelated, and the gap between those two numbers is where GEO lives.

The average health score across the sample was 35 out of 100, 58% scored at or below that, and no site cleared 60. The practical reading is not that the market is incompetent but that the category answers are still unclaimed.

the three layers, in order

readable

Can the crawler fetch your pages and get your content back? Several AI crawlers execute no JavaScript, so a client-rendered site can return a valid response containing nothing. Half of the audited sites were unreadable to some degree and a third were a complete blackout. One site scored 100/100 for SEO in Lighthouse while GPTBot received 1,133 bytes of empty shell from the same URL.

verifiable

Can a model confirm you are a specific, real organisation? This is the entity layer: an organisation record with stable identifiers, references to profiles that exist, the same name and description everywhere a machine can read them, named humans on the content. Close to 100% of audited sites had no entity anchor at all.

quotable

Is there anything on the page a model can lift in one paragraph? Definitions, comparisons, prices, dated figures with sources. Anonymous content with no author and no dates appeared in 89% of audits, and it is the profile a model discounts when choosing whom to cite.

The order is not a preference. Entity work on a site the crawler cannot read is invisible; citation work on a brand the model cannot verify does not hold — the name appears in an answer and then drops out.

what GEO is not

Not prompt engineering. You are not optimising the question, you are optimising what is available to answer it.
Not buying mentions. In one audit the only readable corpus about a brand was its own paid promotion; the engines flagged the sources as promotional and issued a high-risk verdict. Volume without independent confirmation makes answers worse.
Not a guarantee of position. The systems are not ours and they change without notice. What can be promised is a protocol and a scoreboard.
Not separate from SEO. Many cited sources are ranked pages. Rankings feed citations; citations feed answers.

where to start

With a measurement, because every decision after it depends on which of the three layers is your bottleneck. The cheapest first step costs one command: request your own homepage with an AI crawler’s user-agent and read what comes back. If it is a shell, you have found your bottleneck without any help.

The full protocol is described on the audit page, the technical layer in detail on technical GEO, the vocabulary in the glossary, and the complete dataset behind every number here in 100 GEO audits: the report.

what is generative engine optimization?