What Is Generative Engine Optimization? Google Says It's Just SEO

Where the term came from: the KDD 2024 paper

GEO was coined in GEO: Generative Engine Optimization by Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande, presented at ACM SIGKDD 2024. The authors' framing is worth noting: they wrote it as a creator-side response to a problem, arguing that the black-box nature of generative engines leaves content creators with little control over when and how their content is shown.

The setup was a benchmark called GEO-bench: roughly 10,000 queries drawn from nine datasets across about 25 domains, split into training, validation and test sets, weighted towards informational intent. For each query, the researchers took the set of source documents an engine would work from, rewrote one of them using a candidate optimisation method, and measured whether that source's presence in the generated answer improved.

Visibility was scored two ways. Position-adjusted word count measures how much of the answer your source accounts for, weighted by where it appears. Subjective impression uses a model to rate qualities like relevance and influence. Both are proxies for prominence inside an answer. Neither is clicks, traffic or revenue — a distinction that gets lost every time the headline number is quoted.

What the paper actually found

Nine methods were tested. The results are more interesting than the headline, because the losers are as instructive as the winners.

Two findings deserve to be stated on their own, because they are routinely inverted in the trade press.

First: the winning tactics are all forms of raising evidentiary quality. Adding real numbers, quoting credible sources, attributing claims, writing clearly. There is no formatting trick in the top four. The paper's own summary is that including citations, quotations and statistics can significantly boost source visibility.

Second: the one recognisably "SEO hack" tested performed worse than doing nothing. Keyword stuffing lost ground against the unmodified baseline, and lost more in the paper's live validation on Perplexity than in the benchmark. A generative model is not counting term frequency; padding the page dilutes the passages that would otherwise be liftable.

The authors also found that effectiveness varied by domain — statistics help most on data-led topics, quotations on historical and cultural ones — and concluded that domain-specific optimisation is necessary. That caveat almost never survives into the summaries.

What the paper did not show

The "up to 40%" figure has been repeated so widely that it now functions as a sales claim. Four limits keep it honest:

• It is a maximum, not an average, and it is relative. Thirty to forty percent describes the best-performing methods on one metric against an unoptimised baseline, with low-ranked sources benefiting disproportionately. It is not a 40% traffic uplift and was never claimed as one.
• It was not measured on Google. The work used a constructed generative-engine pipeline plus validation on a live citation-first engine. Google's AI surfaces were not the test bed, and Google has since published its own contradicting guidance about what it uses.
• It assumes you are already retrieved. The experiment rewrites a document that is already in the candidate set for a query. Nothing in the paper addresses getting into that set — which, in production, is where most sites actually fail.
• Visibility is not value. Occupying more of an answer is not the same as earning a click, a customer or a sale. Later academic work has criticised this whole benchmark family for measuring prominence rather than usefulness to the person asking.

Read with those limits in place, the paper is a solid, narrow result: on a fixed set of candidate sources, making a document more evidentially specific and better written makes a generative engine quote it more. That is genuinely useful. It is also a long way from a discipline.

Google's official position, in its own words

Google's guide to optimizing for generative AI features tackles the terminology directly. It acknowledges that AEO and GEO are terms used to describe work focused on AI search visibility, then states that from Google Search's perspective, optimising for generative AI search is optimising for the search experience, and thus still SEO. It then links readers to its guidance on evaluating third-party SEO advice and services.

The reasoning is architectural rather than rhetorical. Google's generative features are rooted in its core Search ranking and quality systems: retrieval-augmented generation pulls pages from the same Search index, and query fan-out issues related searches against that same index. There is no second pipeline to optimise for. A page has to be indexed and eligible to appear with a snippet before it can appear in an AI surface at all.

The guide then lists what Google says you can ignore for Google Search. This is the most commercially significant paragraph published on the subject in 2026:

That last row matters more than it looks. Fan-out is real and Google documents it, so covering the questions around your topic genuinely helps. Spawning a thin page for each variant is the version that gets you penalised. Depth on a page and breadth across a genuine cluster; not one page per query.

Can the research and Google both be right?

Yes, and noticing why is the useful part.

Line the two up. The paper's winning methods were: add statistics, add quotations, cite sources, write more fluently. Google's advice is: create non-commodity content with a unique point of view, organise it clearly for readers, avoid recycling what everyone else has said. Those are the same instruction expressed in two vocabularies. Nothing in the peer-reviewed GEO result requires a tactic Google warns against — and the one tactic the paper tested that is a classic manipulation, keyword stuffing, performed worse than doing nothing.

So the disagreement is not between researchers and Google. It is between both of them and a vendor category that took the paper's name, discarded its methods, and attached it to files, chunking and mention-buying — none of which the paper tested and all of which Google names as ignorable. When a proposal cites the 40% figure and then sells you an llms.txt file, it is quoting a study it has not read.

The practical reconciliation, and our position: GEO is not a separate discipline, but it is a useful name for a surface. The work is SEO plus an engineering layer — crawl access for AI agents, index coverage beyond Google, render-side visibility, feeds — that traditional SEO scope never had to cover. The sibling label, answer engine optimization, names the same work aimed at assistants outside Google, where index coverage and agent access diverge most from a Google-only workflow. The evidence ranking behind that engineering layer is laid out in our breakdown of the factors that decide AI citations.

If GEO is just SEO, why do we still use the word?

A fair challenge to put to any agency that sells it, so here is our answer plainly.

People search for it. Business owners arrive asking whether they need GEO because a competitor, a consultant or an assistant used the term. Refusing to use the word means being absent from the conversation where the question is actually asked. The label also does one honest job that "SEO" does not: it names the surface. When a client says "we rank fine but ChatGPT describes us wrongly", they have a real problem that the word SEO does not evoke, even though the fix lives in the same toolbox.

What we do not do is price the label as a separate product with separate mechanics. Anyone quoting you for GEO as a distinct service, on top of SEO, with its own tactics and guaranteed AI rankings, is selling a vocabulary. Google's own guidance warns about third-party services promising ranking success and about tools claiming access to internal metrics — no such access exists.

If you want a second opinion on a GEO proposal you have been sent, a free consultation costs nothing and will usually tell you within ten minutes whether the tactics in it are supported by anything.

What to actually do this quarter

Reading the paper and Google's guide together produces a short, unglamorous list. In priority order:

1. Confirm you are indexed and snippet-eligible. Google states a page must be indexed and eligible to show with a snippet to appear in generative features at all. Check for stray noindex and preview directives first — this is a floor, not a tactic.
2. Confirm AI retrieval agents can fetch your pages. Blocking the crawler that gathers training data is a policy choice; blocking the fetcher that answers live questions removes you from answers. Our guide to robots.txt for AI crawlers covers which agent does what.
3. Apply the paper's actual findings to your top pages. Replace vague claims with numbers, attribute facts to sources, quote credible ones, and tighten the prose. This is the only tactic list in this article with a peer-reviewed result behind it.
4. Cover the fan-out on one strong page, rather than spawning a page per variation.
5. Stop doing the mythbusted five. No chunking, no AI-specific rewriting, no mention buying, no llms.txt as a ranking play, no schema as a substitute for content.
6. Measure in Search Console, using the generative AI performance report, and treat third-party AI-visibility scores as directional at best.

If that list looks like SEO with sharper edges, that is the finding. The genuinely new work sits underneath it in the engineering layer, and that is the half we build for clients as part of our SEO, AEO and GEO service — crawl access, indexing, structure, agent-readiness and the content that earns the citation once the plumbing works.