Generative engine optimization: how AI engines pick sources
Generative engine optimization (GEO) is the work of making your pages likely to be used and cited in answers that an AI engine writes. An AI engine finds pages, picks some, and writes an answer with links to them. GEO targets the picking step. The term comes from a 2023 research paper. Google's own guidance says that optimizing for generative AI search is still SEO. This article explains how engines pick sources, what the paper tested, what Google says, and how to measure the result.
What GEO means and what people call it
The same idea goes by several names: GEO, LLM SEO (LLM means large language model), "SEO for AI search" and, in some articles, answer engine optimization. No standard body defines these terms. This article uses GEO for one job: raising the chance that an AI engine cites your page. The sibling article on answer engine optimization covers the page-format side.
Google gives its own view in its guide to generative AI features: optimizing for generative AI search "is optimizing for the search experience", so it counts as SEO. Treat GEO as a label for a part of SEO, not as a separate discipline with its own rules.
How AI engines pick sources
Each engine has its own pipeline. Google does describe how AI Overviews and AI Mode work. The documentation says the features may use "issuing multiple related searches across subtopics and data sources" to build a response. Google calls this query fan-out. While the answer is written, its systems identify more supporting web pages. The aim is to show a wider and more diverse set of helpful links than classic search does.
To be a supporting link, a page must be "indexed and eligible to be shown in Google Search with a snippet". Google adds that there are no additional technical requirements, and that "indexing and serving isn't guaranteed". So the first GEO step is not new. Make sure Google can crawl, index and show the page.
The GEO paper used a simpler engine for its tests. It fetched the top five Google results for a query, then asked a language model (gpt-3.5-turbo) to write an answer from those pages. Real engines differ. Keep that in mind for every number below.
Where GEO differs from classic SEO
In classic results, the unit is a ranked link. In an AI answer, the unit is a source cited inside written text. The paper reflects this in its first metric: it adds up the words of the sentences that cite your source, and it weights each sentence by where it sits in the answer. A source that is cited early and often counts for more.
Two Google statements frame the rest. AI Overviews appear only when Google's systems determine that they are "additive to classic Search". And sites that appear in AI features are included in the overall search traffic that Search Console reports. You do not get a separate channel to manage. You get the same pages, judged in a new format.
What the GEO paper tested
Aggarwal et al. introduced GEO as a black-box framework for improving how visible a source is in a generative engine's answer. The authors built GEO-bench, a set of 10,000 queries across 25 domains. Each query came with the cleaned text of its top five Google results.
They tested nine ways to edit a source page. The methods include:
- adding citations from credible sources (Cite Sources);
- adding quotations (Quotation Addition);
- adding statistics (Statistics Addition);
- making the text easier to understand or more fluent;
- adding keywords from the query (Keyword Stuffing).
The paper scored each edit with two visibility metrics. Position-Adjusted Word Count counts the words of the sentences that cite your source and weights them by position. Subjective Impression uses a language model (GPT-3.5) to rate facets such as relevance and the likelihood of a click.
What the statistics say
The abstract says GEO can "boost visibility by up to 40% in generative engine responses." Read that number with its setup:
- It is a relative change in the paper's own metrics. It is not a change in clicks, traffic or rankings.
- The paper reports that Cite Sources, Quotation Addition and Statistics Addition gave the largest gains, 30 to 40 percent relative on Position-Adjusted Word Count and 15 to 30 percent on Subjective Impression.
- Results depended on the source's rank. In a test where the authors applied each method to all five sources at once (section 5.2), Cite Sources raised the fifth-ranked source's visibility by 115.1 percent. It lowered the top-ranked source's visibility by 30.3 percent on average. The paper's table reports this as "relative improvement in visibility" and does not name the metric.
- Keyword Stuffing, the classic SEO edit, gave "little to no improvement". On Perplexity.ai it did 10 percent worse than the baseline.
- The authors conclude that "lower-ranked websites, which typically struggle for visibility, benefit significantly more from GEO." This holds for their setup. It says nothing about how a real engine treats a small site.
- On 200 test queries with Perplexity.ai, the authors uploaded the source texts as files and had Perplexity answer only from them. The best methods gained 22 percent on Position-Adjusted Word Count (Quotation Addition) and 37 percent on Subjective Impression.
The authors list limits. They tested only two generative engines. They did not measure effects on search rankings. They also expect engines and query behavior to change. Use the paper as evidence of what can influence an engine in a lab setup, not as a promise about ChatGPT or Google.
What Google says to do
Google's guide puts one practice first: create "non-commodity" content. It defines this as content that provides "unique expert or experienced takes that go beyond common knowledge and the ordinary." The guide also asks for a clear technical structure and, for businesses, accurate local and product details.
The same guide lists things you do not need to do:
- Create
llms.txtfiles or other special markup. Google says Search "doesn't use them." See llms.txt for the format. - Split content into tiny pieces for AI.
- Write in a specific way just for generative AI search.
- Seek inauthentic mentions across the web.
- Add special structured data for generative AI search.
These points do not conflict with the paper. The paper tests edits that add citations and facts to a page. Google's advice is to publish original content that adds something. Adding sourced facts to a page is not a trick. It is a reason to cite the page.
Google also explains how to measure. It says to use the Generative AI performance report in Search Console, and it states that "no third-party tool has access to our internal ranking or AI systems."
What no one can promise
Google states that "indexing and serving isn't guaranteed". No vendor controls which pages an engine cites, and the paper tested only a lab setup. Treat any offer that guarantees a citation, or that sells inauthentic mentions across the web, as a risk. Google says seeking such mentions "isn't as helpful as it might seem."
Let the search crawlers in
Other engines run their own crawlers. OpenAI says OAI-SearchBot is for search and recommends that you allow it in robots.txt. It also says sites that opt out "will not be shown in ChatGPT search answers." GPTBot is a separate crawler for content that may train models. Perplexity describes PerplexityBot as designed to surface and link websites in its search results.
Each setting is independent, so you can allow a search bot and block a training bot. The user-request fetchers behave differently. Perplexity says Perplexity-User "generally ignores robots.txt rules" because a user asked for the fetch, and OpenAI says ChatGPT-User does not decide whether content appears in search. Both vendors say a change can take up to about 24 hours to take effect. The sibling article How to rank in ChatGPT covers the ChatGPT side in more depth.
How to measure GEO
You cannot read an engine's ranking. You can ask it your real questions and record what it cites. A useful record has the prompt, the engine, the date, the cited URLs, whether your brand appears and which competitors appear. Repeat the same prompts on a schedule, because answers vary between runs. The paper handles this too: it sampled five responses per query at a temperature of 0.7 before it compared sources. One answer is an anecdote. A trend over many runs is evidence.
Strategist's AI answers view at /performance/ai-answers runs your saved prompts across OpenAI, Anthropic, Google, xAI, z.ai and Perplexity with your own API keys, and trends how often each cites you (Do AI assistants cite you?). For how tracking tools work in general, see AI visibility tracking. For Google's own features, see Google AI Overviews and SEO.
A short GEO checklist
- Confirm the page is indexed and can show a snippet.
- Allow the search crawlers you care about in
robots.txt. - Add facts a reader can check: statistics and quotations with a named source.
- Say something original that the pages ranking above you do not.
- Track the same prompts over time and compare.
FAQ
Is answer engine optimization the same as GEO?
Not exactly. Both aim at AI answers, but answer engine optimization also covers snippets and direct answers in classic results. The answer engine optimization article draws the line.
Next steps
- Read Answer engine optimization to format pages for direct answers.
- Read the GEO paper to check its setup against your own case.
- Compare your current plan with Google's guide to generative AI features.
