GEO and AI visibility: what the research actually says
GEO (generative engine optimization) is work on your visibility in AI answers. In the research the term comes from, two changes scored highest in the main test: adding quotations and adding statistics. That was measured in a test setup, with a source's share of the answer as the yardstick. Google calls GEO for Search "still SEO", and since February 2026 Bing offers a report that shows how often your pages are cited as a source.
Until 14 November 2026 this page attributed five percentages to one study, among them plus 40 percent from statistics and plus 30 to 40 percent from structured data. When we opened the study, two of the five rows turned out to be about something it did not test. This version follows the sources: the study, the documentation of Google, Bing, OpenAI and Perplexity, and what we found on our own site.
What is GEO?
GEO stands for generative engine optimization: work that raises the chance that an AI answer uses and cites your page as a source. The term comes from the study GEO: Generative Engine Optimization by researchers from, among others, Princeton University and the Indian Institute of Technology Delhi.
The researchers describe the problem like this: an AI search engine fetches sources and summarises them, and the maker of a website has little to no control over when and how their content ends up in that answer. GEO is their name for changes to the text of a page that enlarge that share.
Is GEO different from SEO?
Not for Google Search. Google writes in its guide to generative AI in Search: "From Google Search's perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO."
Bing does use the term. When it introduced a new report in Bing Webmaster Tools, Microsoft wrote on 10 February 2026 that the release is "an early step toward Generative Engine Optimization (GEO) tooling". So one search engine uses the word and the other treats it as part of SEO. How the three terms SEO, GEO and agent-ready relate is covered in SEO vs GEO vs agent-ready.
What does the research say about GEO?
Adding quotations and adding statistics scored highest on both metrics in the study's main test. The researchers write that GEO can raise visibility "by up to 40%".
This is how it was measured. The researchers built their own AI search engine: for every query the top five Google results were fetched, after which the language model GPT-3.5 Turbo wrote an answer with citations. They compiled a set of 10,000 queries and tested on its test split, 1,000 queries. Per query they picked one source at random and changed it in each of nine ways. Then they looked at how much of the answer was attached to that source.
These are the scores from the main table of the study. Higher is better.
| Change to the page | First metric | Second metric |
|---|---|---|
| No change | 19.3 | 19.3 |
| Adding quotations | 27.2 | 24.7 |
| Adding statistics | 25.2 | 23.7 |
| Writing more fluently | 24.7 | 21.9 |
| Citing sources | 24.6 | 21.9 |
| Adding technical terms | 22.7 | 21.4 |
| Writing more simply | 22.0 | 20.5 |
| A more authoritative tone | 21.3 | 22.9 |
| Adding unusual words | 20.5 | 20.4 |
| Repeating keywords | 17.7 | 20.2 |
The first metric counts how many words of the answer are attached to a source, with a citation further down the answer weighing less. The second is a language model's judgement of, among other things, the relevance and influence of the citation. With the table the researchers write that the best methods come out 41 and 28 percent above the baseline. The bottom two rows are listed in the table under "Non-Performing".
In the running text the researchers call citing sources, adding quotations and adding statistics their three top-performing methods, with "a relative improvement of 30-40%" on the first metric. Writing more fluently and more simply gave, according to them, "a significant visibility boost of 15-30%". Put those sentences next to the main table and writing more fluently sits beside citing sources on the first metric: 24.7 and 24.6. The appendix has a second table with deviations, and there citing sources comes out higher: 25.3 against 24.4, with a margin of 0.6 and 0.8. So the difference between those two methods is small and depends on the table you read.
Two other outcomes are interesting for a small business:
- With three methods, the page ranked fifth gained the most. This comes from a separate test in which all sources were changed at once. The table for that test shows five methods. There, citing sources gave the page ranked fifth in Google "a substantial 115.1% increase in visibility", while the page ranked first lost 30.3 percent on average. Adding quotations and statistics showed the same pattern; a more authoritative tone and more fluent writing did not.
- The effect differs per topic. The researchers write that "the efficacy of these strategies varies across domains".
The researchers repeated the test on Perplexity with 200 queries from the test split, supplying the source text as a file. The table for that test has seven of the nine methods. There, according to the researchers, adding quotations gave a 22 percent improvement on the first metric, and repeating keywords performed "10% worse than the baseline". The table shows that on the first metric: 21.9 against 24.1. On the second metric, repeating keywords came out higher than doing nothing: 28.1 against 24.7. Citing sources came out lower on the second metric: 19.0 against 24.7.
What does the research not prove?
The study does not prove extra visitors, customers or revenue. It measures a source's share of an answer, in a setup where that source was already among the top five search results.
Four more limits apply:
- The page has to be found first. Only pages from the top five search results took part. A page that is not there did not appear in the test.
- The main test ran on GPT-3.5 Turbo. The researchers write that methods "may need to adapt over time" as AI search engines change.
- The effect on your position in Google was not measured. Literally: "we didn't evaluate how GEO methods affect search rankings".
- The changes were written by a language model. That model added statistics and quotations to existing text. For your own site a stricter rule applies: a statistic or quotation has to be correct and have a source.
That a language model wrote the changes is in the study. The requirement we attach to it is our own addition. We mention it because the temptation to invent statistics is strong once you read that statistics help.
How AI-ready is your business?
Take the free AI Readiness Quiz: 7 questions, 2 minutes.
Take the quiz →How often do people click through when Google shows an AI summary?
About half as often, in one American measurement of searching in Google. Pew Research Center followed the browsing of 900 American adults in March 2025. On Google search pages with an AI summary they clicked on a regular search result in 8 percent of visits; without a summary that was 15 percent. A link in the summary itself was clicked in 1 percent of visits.
Separately, using the Datos panel, SparkToro calculated that in 2024, 59.7 percent of Google searches in the European Union and 58.5 percent in the United States ended without a click. That figure is about Google searches in general. SparkToro notes on the article itself that the figures may be out of date and points to a newer article.
We found no separate figures for the Netherlands. What you can take from this: being cited in an answer and getting visits are two different things, and the first does not automatically produce the second.
What can you do today?
Start with access, then content. In this order:
- Allow the search robots of AI services. OpenAI writes about its robots: "OAI-SearchBot is used to surface websites in search results in ChatGPT's search features." According to OpenAI, a site that blocks that robot is not shown in ChatGPT search answers, though it can still appear as a navigational link. Perplexity recommends allowing PerplexityBot in your robots.txt. For Google's AI features a page has to be indexed and eligible to be shown with a snippet. Google's guide names one more requirement: the site must be included in Search generative AI features in Search Console.
- Write what only you can write. According to Google this probably weighs most: content that people find unique and useful "will likely influence your website's presence in generative AI search in the long run more than any of the other suggestions in this guide". First-hand experience, your own measurements and your own point of view fall under that.
- Use real quotations and statistics, and cite your sources. Adding quotations and statistics scored highest in the main test. In the test where all sources were changed at once, citing sources gave the page ranked fifth the largest gain of the five methods in that table. A statistic without a source is worth nothing to your reader, and a risk to your credibility.
- Write clearly. In the study's main table, writing more fluently came out about level with citing sources on the first metric.
- Keep your SEO in order. A page that is not indexed cannot serve as a source. The study started from the top five search results.
What you can leave aside, according to Google: "For Google Search, you can ignore tactics like "chunking" content, creating unnecessary AI text files (like llms.txt), or pursuing inauthentic mentions." About structured data Google writes in its documentation on AI features: "There's also no special schema.org structured data that you need to add." Structured data remains useful for your regular search results; it is not a condition for Google's AI answers. What an llms.txt does and does not do is covered in our llms.txt guide.
How do you measure whether it works?
With the search engines' own reports, and by hand for the rest.
| Where | What you see |
|---|---|
| Google Search Console | In its guide Google points to the "Generative AI performance report" for your visibility in the AI features of Search |
| Bing Webmaster Tools | The AI Performance report, in public preview since 10 February 2026: how often your pages are cited as a source in Microsoft Copilot, in AI summaries in Bing and at some partners, and which pages those are |
| ChatGPT, Perplexity and others | As far as we know there is no report for site owners. You can ask questions yourself and note each time you are mentioned |
Microsoft adds a caveat to the Bing report itself: the citation total says how often you are cited, "without indicating placement or presentation within a specific answer". For questions asked by hand, answers can differ per attempt and per user. Ask the same questions at fixed moments and note what you see. How to set that up is covered in measuring AI visibility.
What did we find on our own site?
We went through mindsora.io in November 2026. What that produced for this topic:
- Access. Our robots.txt allows OAI-SearchBot, PerplexityBot, Googlebot and Bingbot.
- An error our own check missed. On 116 of the 146 pages the head of the page (title, description and references) after loading differed from the HTML the server sent. Our check only read that HTML. That has been repaired: after the fix we counted 0 of 146. More on that in SEO vs GEO vs agent-ready.
- This page itself. The previous version attributed five percentages to the study. Two of the five rows were about something the study did not test: structured data and authority signals. The word "schema" does not occur in the study, and what it tests under "Authoritative" is a more assertive writing tone. We did not find the percentage we gave for technical terms in it either. We replaced the table with the scores from the study itself.
- What we do not know yet. We read a baseline in Search Console on 14 November 2026. Whether our changes make a difference will only show at the next measurement.
The lesson for us: read the source before you copy a figure, and with a study, read the tables and the appendix too.
Frequently asked questions
Does GEO work for a small business too? In the test where all sources were changed at once, the page ranked fifth gained the most with three of the five methods tested. That was in a test setup with the top five search results. It promises nothing for your site.
Do I need structured data to get into AI answers? Not for Google's AI features: Google writes that you do not need to add special structured data for them. For your regular search results it remains useful.
Do I need an llms.txt? Not for Google Search; Google calls the file a tactic you can ignore. More in our llms.txt guide.
Can I measure GEO? Partly. Google and Bing have a report for it. For ChatGPT and Perplexity you depend on asking questions yourself and keeping track.
Our free Agent-Ready Grader tests one point from this article: whether your robots.txt allows AI robots. Apart from the score it requests your homepage with a user-agent string that carries the name of three robots: OAI-SearchBot, Claude-SearchBot and CCBot. The score from 0 to 100 follows our own weighting: 20 points for an llms.txt and 30 for structured data, two things Google does not require for its AI features. It does not test whether an AI answer cites you.
Want to work out how many hours a recurring task costs you? The free AI Readiness Scan does that: 7 questions about your business processes.
Ready to automate?
Book a free consultation and discover what AI automation can do for your business. Or take the 2-minute AI Readiness Quiz first.
Book a call