What Is Generative Engine Optimization (GEO)? How to Become Visible in AI Answers

Generative Engine Optimization (GEO) describes work on content with the aim of appearing inside the answers of generative AI systems. That means systems such as ChatGPT, Perplexity, Google Gemini or the AI Overviews in Google Search. They answer a question with a finished text instead of a result list and a few sources get quoted inside it while many others never show up at all.

So much for the common definition. The more interesting question is what it rests on.

I looked at the German search results for “generative engine optimization” on 16 August 2026. Among the first fifteen results sit eight guides from agencies and tool vendors. Between them sit the two sources the whole topic feeds on: in position 6 the research paper that invented the term, and in position 9 Google’s own documentation, which contradicts it rather directly.

Both are freely available and neither one says very much without the other. That is why they stand side by side in this article.

The route there: first the definition and the origin, then the evidence together with its limits, then the implementation. Anyone who only needs the recommendations can jump straight there through the table of contents.

The source is a paper with the preprint number arXiv:2311.09735 titled “GEO: Generative Engine Optimization”. The authors are Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande. Version 1 was submitted on 16 November 2023 and version 3 revised on 28 June 2024. The note on the arXiv page reads “Accepted to KDD 2024”, meaning the 30th ACM SIGKDD conference held in Barcelona from 25 to 29 August 2024.

The paper introduces an umbrella term first: systems such as BingChat, Google’s SGE at the time and Perplexity.ai are called “generative engines” because they search for information and generate an answer from it. The step after that is the actual contribution: a framework website owners can use to influence their visibility inside such answers.

The reasoning in the abstract is remarkably non-technical. It is about the third party in the game: “it poses a huge challenge for the third stakeholder”, and the original names websites and content creators explicitly. Anyone producing content loses control over whether and how they appear in a generated answer. The authors propose a way out of that, which they call “the first general creator-centric framework”.

Where AIO, AEO and LLMO fit into that picture. GEO targets generative answers and AEO targets direct answer engines. LLMO refers to how language models understand content. AIO is the umbrella above all of them. The boundaries are soft and interchangeable across many texts. I have compared AIO, GEO, AEO and LLMO in a separate overview. Only one point matters here: out of the four abbreviations I could find a named primary source with disclosed methodology for exactly one.

That is the origin of GEO: no product and no Google announcement, but an academic proposal with a benchmark next to it.

The paper appeared in November 2023. That GEO is now sold as a discipline of its own therefore has less to do with the paper than with the surface search results land on these days.

A generated answer settles the question on the results page already and turns the click on the source into an optional extra. This pattern is older than AI search and goes by the name zero-click search. Generated answers give it a new order of magnitude: whoever supplies the answer and still goes unnamed loses the visibility entirely rather than just a rank.

The second trigger has a date. On 15 May 2026 Search Engine Land reported on a new documentation page from Google about optimising for generative AI features. That made the surface official. A documented selection process creates a market for advice about it.

Part of the same development is the report Google now offers for it and which the documentation explicitly points to for measurement: the Generative AI performance report in Search Console. A platform does not build a report for a surface it considers unimportant.

What those two triggers do not supply is a documented statement about effect: the importance of a topic says nothing about the effect of any single measure. This is exactly the point where it gets tempting to jump straight from market observation into a list of measures. The following sections take the detour through the evidence instead.

The common comparison runs: SEO brings rankings and GEO brings mentions. That is not wrong, it merely blurs the actual difference and the comparison only gets useful once you lay the two evidence bases side by side.

Table: own comparison based on the paper “GEO: Generative Engine Optimization” (Aggarwal et al., v3 of 28 June 2024) and the Google Search Central documentation, status 10 July 2026.

The second row is the one that counts. GEO in the sense of the paper starts at a page that has already been retrieved. That is the test setup itself rather than a side condition, as the section after next shows in detail: for a page outside the candidate set even the best wording changes nothing.

This is why the formula “GEO replaces SEO” does not hold. It assumes that generative systems find their sources differently from search engines. At Google that is explicitly not the case according to the AI Optimization Guide. For the other vendors comparable documentation is missing and the claim there stays neither proven nor disproven.

The 40 % are quickly quoted. The setup behind them needs a few more paragraphs. So let me take this in order.

The benchmark. Lacking existing data, the authors build their own test set: “GEO-bench, a benchmark consisting of 10K queries from multiple sources”. Those 10,000 queries come from nine different sources. According to the paper they cover 25 domains from arts through health to games. Different difficulty levels and search intents are layered on top.

The engine. This is where it gets important for any interpretation, because per query the setup fetches the five best sources: “only the top 5 sources are fetched from the Google search engine for every query”. A language model then produces the answer: “The answer is then generated by the gpt3.5-turbo model”. To dampen random effects the authors sample five answers per query at a temperature of 0.7.

The second engine. The same test additionally runs against a real product: “we evaluate the same Generative Engine Optimization methods on Perplexity.ai, which is a commercially deployed generative engine”. The limitations section has the decisive number: “we rigorously test our proposed methods on two generative engines, including a publicly available one”. Two engines, then. Google’s AI Overviews appear in the paper as an example of generative systems rather than as a test environment.

The metrics. Two figures score the result: Position-Adjusted Word Count counts the words of an answer that stem from a particular source. Weighting follows position, so whatever appears at the top counts more. Subjective Impression bundles several subjective factors into one value and is assigned by a language model. Those factors include relevance and influence of the cited passage as well as its uniqueness. Both measure visibility inside a generated answer rather than rankings or clicks.

The paper draws that line cleanly itself: “owing to the black-box nature of search engine algorithms, we didn’t evaluate how GEO methods affect search rankings”.

Then there is the age of the experiment, which ran on gpt3.5-turbo and therefore on the state of 2023. Today’s models are different ones and so are the retrieval layers above them. That does not devalue the results. It limits them to a careful measurement of a system from back then.

Nine variants were tested: “We propose 9 different Generative Engine Optimization methods to optimize website content for generative engines.” Each one alters the same source text in a different way. The results sit in table 1 of the paper, in absolute values. I have reproduced them here because the absolute values are the most interesting part of the whole paper and get lost as soon as everything is boiled down to a single percentage.

Table: absolute values from table 1 of the paper “GEO: Generative Engine Optimization” (Aggarwal et al., GEO-bench test split, averaged over five runs, version v3 of 28 June 2024). Higher is better. Observed in the authors’ test setup, not general proof.

The paper summarises the leading group like this: “our top-performing methods, Cite Sources, Quotation Addition, and Statistics Addition, achieved a relative improvement of 30-40% on the Position-Adjusted Word Count metric and 15-30% on the Subjective Impression metric”. The best value sits in the table caption: “The best methods improve upon baseline by 41% and 28%”. That is the Quotation Addition row.

I find the bottom end more interesting. Keyword stuffing is arguably the best known tactic from the early days of search engine optimisation. In this visibility metric it drops below the untouched original text. The paper phrases it carefully: “we find such methods offer little to no improvement on generative engine’s responses”. On Perplexity.ai it gets clearer, where it “performs 10% worse than the baseline”.

The second engine confirms the pattern in smaller numbers. On Perplexity.ai, Quotation Addition leads “with a 22% improvement over the baseline”. Cite Sources and Statistics Addition reach “improvements of up to 9% and 37% on the two metrics” according to the paper.

One detail regularly gets lost in the shortened versions: the effect depends on the topic. The abstract says “the efficacy of these strategies varies across domains”. The results section makes that concrete: domains such as “Law & Government” and question types such as “Opinion” benefit particularly from numbers in the text according to the paper. The authors conclude from this that domain-specific optimisation is needed.

The AI Optimization Guide in the Search Central documentation now has the status of 10 July 2026. When it appeared I covered in detail why Google files GEO and AEO under SEO. The core sentence is short:

“From Google Search’s perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO.”

The reasoning sits one paragraph above. According to Google the AI features run on top of the existing systems: “our generative AI features on Google Search are rooted in our core Search ranking and quality systems”. Two techniques are named explicitly. Retrieval-augmented generation is also called grounding there and pulls current pages from the search index to support the answer. How that anchoring works I took apart in a separate piece on grounding. Query fan-out generates several related queries for one request at the same time to gather more material. The example in the documentation: “how to fix a lawn that’s full of weeds” turns into “best herbicides for lawns” and “remove weeds without chemicals” among others. What that means for the way AI Overviews work is already covered on this blog too.

The documentation gets more explicit in a section headed “Mythbusting generative AI search: what you don’t need to do”. It lists what you can ignore for Google Search:

• llms.txt and similar files: “You don’t need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search”. Google Search simply does not use them. What the format is about and who else reads it sits in my llms.txt guide.
• Content chunking: “There’s no requirement to break your content into tiny pieces for AI to better understand it.”
• Rewriting for AI: “You don’t need to write in a specific way just for generative AI search.”
• Bought mentions: “seeking inauthentic ‘mentions’ across the web isn’t as helpful as it might seem”.
• Structured data as an AI lever: “Structured data isn’t required for generative AI search, and there’s no special schema.org markup you need to add.” The same passage keeps it useful for rich results, which I work through in my piece on structured data and AI Overviews.

On the terms themselves the documentation gets unusually direct. It writes that many of the circulating tricks “aren’t effective or supported by how Google Search actually works”. Then it adds a pointer: “If you’re considering third-party ‘AEO’ or ‘GEO’ advice or services, review our guidance on evaluating third-party SEO advice.”

That is a statement about Google Search and says nothing about how ChatGPT, Perplexity or Claude pick their sources. The distinction matters because both camps like to skip it.

A generative answer comes about in two steps: first the system looks for sources and afterwards a language model formulates an answer from them, deciding which passages it takes over and which source it names.

Google’s documentation talks about the first step. Retrieval-augmented generation and query fan-out describe how pages enter the candidate set. The paper talks about the second. All nine methods alter the text of a page that has already been retrieved.

The evidence for that split sits inside the paper’s own test setup. Its sources come from Google Search and specifically five per query. Whatever is missing from those five takes no part in the measurement at all. Put differently: the experiment presupposes classic ranking instead of replacing it.

That dissolves the two positions. “Up to 40 % more visibility” does not mean “40 % more visibility than without SEO”. It means that among the pages retrieved anyway, the citation shifts. And Google’s “it’s still SEO” does not mean wording is irrelevant. It means Google recognises no second rulebook for it.

An order of work follows from that in practice: first get into the candidate set, then work on the place inside the answer. Anyone tackling the second stage without holding the first is optimising a text that no system retrieves. I took that apart elsewhere: how source selection differs between AI Overviews and AI Mode.

Look at the ranking in the table above once more. At the top sit verbatim quotations, numbers and citations, below them flow and readability and at the very bottom the tactic that stuffs the text with keywords.

That is not an AI discipline. That is the description of a well sourced specialist article.

The order in which to tackle this. The following sequence is my rule of thumb rather than a requirement from either source. It follows the split from the previous section: get retrieved first, get quoted second.

1. Check indexability and ranking. Pages that do not sit in the top results for the question appear in no candidate set at all. That is classic work and it comes first.
2. Actually answer the question in the text. An answer that sits in the text can be taken over. One that is merely implied cannot.
3. Build in the evidence. Numbers, quotations and linked sources right where the statement falls. Those are the three methods with the best values in the test.
4. Bring the readability along. Fluency and comprehensibility landed above the baseline in the test without any change to the substance. That is the cheapest step of them all.

What I derive from that for my own work is a rule of thumb too. I treat those four points as quality rules for any specialist article rather than as an AI measure. The reason is simple. They sit on my editorial checklist anyway and cost nothing extra. They also still work when the next engine picks differently.

The reverse applies as well: everything Google’s documentation explicitly calls unnecessary costs time without a documented return. For Google Search at least. If you maintain an llms.txt because another system reads it, that is a separate decision with its own reasoning. Whether your text is more than a prettier version of the usual is what my commodity self-test settles.

The measurement question is more awkward with GEO than with classic SEO. A position in a result list you can look up and whether your text appears in a generated answer depends on the individual query and on the moment and on the model.

Google’s documentation names exactly one source: “To measure how your content is performing in generative AI features on Google Search and Discover, use the Generative AI performance report in Search Console.” How that report is built and what it shows sits in my analysis of the generative AI performance reports in Search Console.

Three things help when reading those numbers:

• Impressions and clicks drift apart. A mention inside an answer produces an impression. The click is then decided by a reader who already holds the answer.
• Visibility spreads per topic rather than per domain. The same website can get quoted regularly on one topic and never on the next. Why that is I described under AI visibility per topic rather than per domain.
• Single spot checks do not count as measurement. The same query returns different sources depending on the run and the paper’s authors sampled five answers per query for exactly that reason.

For ChatGPT, Perplexity, Claude and Gemini an official reporting surface for website owners is missing and what remains in the end is your own spot checks and the numbers of third-party vendors. Both are estimates and both should be called that.

By now there is a range of tools reporting an AI visibility. Between what such a tool can do and what it claims runs a fairly clear dividing line.

Observing is something a tool can do. It puts questions to the engines and logs the answers and counts the mentions of a domain in them. That is a real measurement with the limits from the previous section: a spot check at one moment and with a question set somebody picked.

Reading out is something it cannot do. Google’s documentation is unmistakable here: “No third-party tool has access to our internal ranking or AI systems.” A reported “GEO score” is therefore a model built by the vendor rather than information out of the system.

A usable purchasing check follows from that in two questions. First: which questions does the tool ask and how often and against which engines? Second: how does the reported score come about? On the first question there is a verifiable answer and on the second you will hear either a disclosed formula or an explanation of why there is none.

Two sources, one term, two perspectives. The paper shows that the wording of a text shifts how a generative answer quotes it. Google’s documentation shows that the selection of those sources still happens where it always did. Taken together the two findings paint a fairly unexcited picture of an abbreviation that is now being sold as its own discipline.

One thing stands out: none of the nine tested methods is a trick. Quoting, sourcing, quantifying and writing readably. That is the work a specialist article demands anyway, except that in a generated answer it becomes visible faster than in a classic result list.