What is GEO? The definition, the evidence, and what it cannot prove

· by boneLau

Short answer: GEO — generative engine optimization — is the work of getting a brand into the answers AI assistants synthesize, instead of into a list of links. The term was formalized in a 2024 paper by Aggarwal et al. (KDD 2024), which reported visibility gains of "up to 40%". Almost every summary of that paper drops the condition attached to the number: the content had to already be in the retrieved context. That condition is the whole game.

I work on a tool that measures this weekly, so assume the usual bias. The numbers below are either from that tool's runs or from papers and vendor studies I read in October 2026, and I name which is which.


The definition, and where it comes from

The academic definition is narrower than the marketing one:

"GEO is a novel paradigm to aid content creators in improving their content visibility in generative engine responses through a flexible black-box optimization framework for optimizing and defining visibility metrics." — Aggarwal et al., GEO: Generative Engine Optimization, KDD 2024

Two things matter in that sentence. First, "generative engine" means a system that synthesizes an answer from multiple sources rather than returning a ranked list. Second, "visibility" is defined by what appears inside the answer — a mention, a citation, a position in the list of names the model produces.

In practice, people use GEO, AEO (answer engine optimization), LLMO and "AI SEO" more or less interchangeably. The distinctions vendors draw between them are mostly positioning, not measurement.

What the +40% figure actually measured

The paper tested nine kinds of content edits against a benchmark of queries. Adding quotations from sources, adding statistics and citing credible sources performed best. Keyword stuffing did nothing. Gains were largest for pages that already ranked low — citing sources reportedly raised visibility by 115% for a page ranked fifth.

Then, in 2026, a survey of 45 studies reached a blunter conclusion: no reviewed GEO technique showed a stable, longitudinal, cross-platform causal effect. The 40% was a relative gain on one metric, in one fixed configuration, conditional on the source already being present.

Both things are true. Content edits can change what a model says when the model has already retrieved your page. They cannot make it retrieve your page in the first place.

What GEO is not

  • Not SEO with a new name. Google's own guidance on AI features says there are no additional requirements, which some people read as "GEO is just SEO". That is half right: the same crawlable, well-linked pages are the substrate. But the output being optimized is different — an answer, not a rank.
  • Not paid placement. There is no ad slot inside an AI answer. What looks like placement is a third-party page the model retrieved.
  • Not llms.txt. It is cheap, harmless and occasionally useful for documentation. It is not distribution, and there is no evidence it produces citations on its own.

SEO vs GEO: what you're competing for

SEOGEO
What you compete fora position in a listinclusion inside an answer
What success looks likea clicka mention or a citation, often with no click
Unit that gets useda pagea passage
What decides itranking signalswhat the model retrieved and chose to lift
What you measureimpressions, rankings, sessionsmention rate, citation rate, accuracy

The two are not separate systems. AI engines retrieve from a search index, so a page that cannot be crawled or ranked usually cannot be cited. But the overlap is looser than people assume: an Ahrefs study of 1M AI Overviews put the share of cited pages that also ranked in Google's top 10 at 76% in mid-2025; a later update over 863K keywords put it at 38%. Different methodology, same direction — the citation layer is drifting away from the ranking layer.

Four claims we can defend with data

  1. Third-party pages outweigh your own pages. A 2025 study of AI search found a systematic bias toward earned media over brand-owned content. We see the same thing from the other side: our own site scored 0 mentions across 18 unbranded questions in the week we stopped putting our name in the prompt.
  2. A definition that can be lifted verbatim beats a better-written page. Models quote self-contained sentences. If your first paragraph needs three other paragraphs to make sense, it will not survive extraction.
  3. Branded and unbranded are two different metrics. "Does the model describe us correctly when asked about us" and "does a buyer who has never heard of us ever see us" differ by 80+ points for a young brand. A single GEO score blurs which one it answered.
  4. Change is monthly, not weekly. Run-to-run variance is roughly ±10 percentage points at small samples. A St. Gallen study needed 33–94 repeats per prompt before rankings stabilised. Treat any single-week movement as noise unless it repeats.

How to measure GEO without fooling yourself

Four rules, in order of how often they are broken:

  1. Write down the denominator. Mentions divided by all answers, by answers that name any brand, or by answers where the prompt already contains your brand — three different numbers from the same week.
  2. Run each question at least three times. One run is an anecdote.
  3. Separate mentions from citations. Being named and being the source cited are different outcomes and move at different speeds.
  4. Put an error bar on the number. If a report gives you 43% with no interval, you cannot tell 43% from 38%.

The measurement setup we use is described here, including the parts we cannot measure.

What GEO cannot do

It cannot guarantee inclusion — model versions change and retrieval shifts. It cannot prove that a specific action caused a specific change, because the variance is larger than most effects. And it cannot be verified by one before/after pair, no matter how clean the chart looks.

That is not a reason to ignore it. It is a reason to measure it the way you would measure anything noisy: fixed prompts, repeated runs, a denominator you can state, and a four-week window before you believe anything.

If you only do one thing

Write one page that states what you do, who it is for, and what it is not — in sentences short enough to quote. Then get one third-party page to say the same thing in its own words. That combination is the only one we have seen move an unbranded answer, and it is what most GEO advice is a longer way of saying.

Methodology

  • Definitions quoted from Aggarwal et al., GEO: Generative Engine Optimization (arXiv:2311.09735, KDD 2024).
  • Citation/ranking overlap: Ahrefs studies of 1M AI Overviews (July 2025) and 863K keywords (later update), as reported by Ahrefs and Search Engine Journal.
  • Earned-media bias: arXiv:2509.08919, Generative Engine Optimization: How to Dominate AI Search (2025).
  • Stability estimates: St. Gallen study on answer repeatability; Peec AI's published estimate of three months to a year to confirm a change.
  • Our own numbers: weekly runs on brandscope.dev in October 2026, five core surfaces (ChatGPT, Claude, Gemini, Perplexity, Google AI Overviews), 52 answers per run.
  • We sell AI visibility monitoring. That is a conflict of interest; it is also why the measurement rules above are strict.

Want your own baseline? Run one brand through the live question set — about 30 seconds, no account needed. Check your brand →

← All posts