← All articlesReputation and brand6 min read
GEO: how a brand gets into ChatGPT, Perplexity and Google AI answers
What generative engine optimization is, how it differs from SEO, which tactics are backed by research and which are still opinion. And why brand mentions from creators and communities are now part of the job.
A buyer's path to a brand used to start with ten blue links. Now it increasingly starts with a finished answer that names three or four companies. If yours is not among them, that person will probably never reach your site. The work of getting named in those answers has a name of its own: GEO, generative engine optimization. Here is what in it is backed by research, and what is being sold on hype alone.
Where the term came from
The word GEO comes from a paper by researchers at Princeton and partner universities (Aggarwal et al., first version November 2023, published at KDD 2024). On 10,000 queries they tested which changes to a page raised its chance of being included in a generative answer. Three tactics produced a clear effect: citing sources, adding direct quotations and adding statistics to the text. Visibility gains reached up to 40%. The important word is "up to". That is the ceiling in the experiment, not the average, although plenty of agencies quote it as a typical result.
Since then GEO has become an industry. Ahrefs, Semrush and a dozen new tools sell tracking of brand mentions in AI answers. Keep that in mind when reading their studies: almost every number below is vendor data and correlational.
How big the shift is
Estimates of how often Google shows an AI Overview vary a lot. Comscore counted 39.4% of US desktop queries (June 2026). Semrush, on a sample of commercial queries, found around 90% (October 2026). The honest phrasing is: from about 40% of all queries to 90% of commercial ones, depending on the method.
Ahrefs measured what that does to clicks on 300,000 keywords (February 2026). When an AI Overview is present, click-through on position one is 58% lower, position two 51% lower, position ten 19% lower. A year earlier the drop was estimated at 34.5%. This is a correlation, as Ahrefs itself notes, but the direction is consistent.
Direct traffic from chatbots is still tiny. SE Ranking's 2026 study puts referrals from AI systems at 0.32% of all web traffic. Two years earlier it was 0.02%. A sixteenfold rise on a very small base. The takeaway: GEO today is mostly about not being erased from the answer, and much less about a new stream of visits.
How an AI decides whom to name
This is where the data gets useful. Ahrefs compared 75,000 brands against their appearance in Google's AI Overviews (May 2025). The strongest correlation with being mentioned was the number of brand mentions across the web in general (coefficient 0.66). Backlinks to the site correlated far more weakly (0.22). In plain terms, how much people talk about you on other sites matters more to these systems than how many links point to yours.
Semrush showed where those mentions come from, on 100 million citations (November 2025). ChatGPT's top sources are Wikipedia and Reddit. Google's AI Mode favors LinkedIn and YouTube. Reddit and LinkedIn sit in the top five cited domains on all three systems, Perplexity included. Platforms built on user content account for 17% of the domains ChatGPT cites, while traditional publishers account for 4% (Growth Memo, secondary data).
Two more observations come from Kevin Indig's analysis of 1.2 million answers: 44% of ChatGPT's citations are taken from the first 30% of a page, and within a single topic around 30 domains collect two thirds of all citations. Put the main point first, and aim for a place among those thirty.
What is proven and what is not
Proven by experiment: statistics with a named source, expert quotes and references to research inside the text. That is the Princeton work, and in tests on Perplexity the effect reached 37%.
Supported by correlation: brand mentions on third-party sites, freshness (Ahrefs, 17 million citations: chatbots cite pages that are on average 26% newer than organic results, although Google's AI Overviews lean slightly older), and answering the question in the first third of the page.
Backed by nothing except vendor claims: the llms.txt file. Ahrefs checked 137,000 sites (June 2026). 28% have the file, yet 97% of those files received no request from any AI system in a month. Google's John Mueller said it outright: no AI system currently uses it. You can add one, it does no harm, but it is a checkbox, not work.
Google's position on all this is public. Its guide to AI features (May 2026) says no special markup or separate "files for AI" are needed, and the same quality signals as in regular search apply. Critics reply that, by vendor measurements, organic results and Google's AI Mode overlap by domain only about half the time, so "just do SEO" explains only part of the picture.
Why this is a social media problem
If the strongest signal for AI answers is brand mentions on other people's sites, and the most cited domains are communities and video platforms, GEO stops being a job for the SEO specialist alone. A creator's video that names your product, a thread in a community, reviews on third-party sites: that is the raw material of AI answers.
We found no direct study showing that creator content specifically ends up in citations, so no promises here. But it fits the Ahrefs correlation between brand mentions across the web and appearance in answers: more mentions on other people's sites, higher chance of being named. That is why we now treat creator campaigns and UGC content as work on AI visibility as well as on reach.
What to do in practice
On your own site: move the main answer into the first paragraph, add numbers with sources and dates to your texts, and update older pages instead of only writing new ones. On other sites: watch the context in which you are mentioned and grow mentions where AI systems take their citations from. Track frequency of appearance, not position. In SparkToro's measurements (January 2026), ChatGPT and Google returned the same list of brands in under 1% of repeated runs. Position is random, and frequency is the number worth tracking.
What people say about you online and how to influence it is the subject of our online reputation service. How mention monitoring fits among the other classic tools is covered in the article on classic social media tools.