GEO Is Not a Ranking. It Is Three Events You Can Measure.
Classic SEO has a scoreboard. A query, a position, an impression, a click. Generative answers do not. The page can be fetched, quoted, ignored, or paraphrased with no link, and none of those show up as a rank.
I treat GEO as measurement first and copy second. If I cannot say which event I am looking at, I do not call the work optimization.
Three events, not one metric
When I audit a page for answer engines I separate three things.
Discovered. A crawler fetched the URL. The evidence is a bot hit in the logs, or a fetch I can reproduce. A sitemap entry is not a fetch.
Retrieved. An assistant pulled that page into a live answer. The page was in the context, not merely somewhere in a training corpus. A model that "knows" your product from memory has not retrieved you.
Cited. The answer names the source, or a person arrives from an AI product. A fluent mention with no source is not a citation. Liu, Zhang and Liang frame the related failure mode as verifiability: a statement should be supported by the source it points at, and the source should actually support the statement.[1]
Collapsing these into "AI visibility" is how teams congratulate themselves. A crawler visit is not a citation. A citation in a lab prompt is not a session in analytics.
What the GEO paper actually showed
Aggarwal and colleagues formalized this shift at KDD 2024. A generative engine retrieves sources and writes one answer with inline attributions. Visibility is no longer a rank on a list. They define impression from how much of the answer is tied to a source, and where that source sits.[2]
On their GEO-bench, adding citations, quotations, or statistics improved visibility. Keyword stuffing did not. They report relative gains on the order of 30 to 40 percent for the stronger methods on position-adjusted word count, and up to about 40 percent in the abstract. On a Perplexity file-upload setup, quotation addition was their best position-adjusted result, about 22 percent over the unmodified baseline.
Read the setup before you borrow the number. The main engine was a constructed pipeline: top five Google results, then gpt-3.5-turbo, five samples. The Perplexity run used uploaded source text, not the open web, and a 200-query slice. The paper is explicit that methods will have to change as engines change, and that they did not measure the effect on classic rankings.[2]
So I do not tell a client "GEO adds 40 percent." I tell them the paper found that cited, quoted, and quantified passages were easier for that engine to keep, and that stuffing keywords was a waste. Then I measure their own pages.
Rankings still matter. They answer a different question.
Search Console impressions and average position still tell you whether Google showed a link. They do not tell you whether an assistant used the page. Structured data does not close that gap either. Google's own documentation says structured data is a clue that can enable rich results, and that the markup must describe content the user can see. It is not a citation from an answer engine.[3]
I keep the SEO scoreboard. I add a second one:
- Did an AI crawler fetch the URL, and when?
- If I ask the question the page claims to answer, does a live assistant retrieve this URL or only paraphrase the topic?
- If it names a source, is it this page, and does the sentence it supports actually appear here?
Anything I did not observe stays unmeasured. I do not write a zero for a day the logs failed, and I do not mark a check as passed because the HTML "looks fine."
What I refuse to optimize
I do not spoof Googlebot. A fetch with a browser user agent is not Googlebot access.
I do not hide text for models. If a claim is worth being cited, a person should be able to read it.
I do not treat a single chat transcript as a rate. One session is an anecdote. The GEO paper averaged multiple samples for a reason. When I spot-check an answer, I repeat the question, paraphrase it, and say what I did not measure.
The useful question is narrow. For this URL, for this question, which of the three events happened, and what on the page made the retrieved passage easy to keep?