GEO: how to get into AI answers
Lesson 4 of 16 in the course “Marketing 2027”. Free, with animations and a quiz.
Why AI recommends some and ignores others
More and more people don’t scroll ten links, they just ask ChatGPT, Gemini or Google’s AI Mode. The AI builds its answer from pages it found and judged useful, and names a couple of brands. If your site only says “best quality at the best price”, there’s nothing to retell, and you’re not in the answer. If you have prices, sizes, tests and honest reviews, there’s a reason to mention you. Working on getting into those answers is called GEO, generative engine optimization. You’ll also see AEO, answer engine optimization. Same idea: give the AI a reason to name you.
Case: the study that started GEO
The term GEO was coined by researchers from Princeton and colleagues: their paper was accepted at the KDD 2024 science conference. They took thousands of questions, rewrote the source pages in different ways and checked how often an AI search engine relied on them. The plain stuff worked best: adding specific numbers and statistics, quotes from experts and links to sources. Visibility in answers rose by up to 40%. The old SEO trick of stuffing text with keywords did worse than no edits at all. The authors’ own caveat matters: the effect depends a lot on the topic, so “up to 40%” is the ceiling in an experiment, not a promise to you.
Google: it’s still SEO
There’s already a pile of “secret” GEO services for sale. In July 2026 Google updated its official guide on optimizing for AI search and went through the main myths. Google Search doesn’t use llms.txt: it neither helps nor hurts, it’s just useless for Google. You don’t need to chop text into tiny chunks “for AI”, Google understands pages that cover several topics. There’s no special markup for AI; regular structured data is still useful for rich results. And chasing fake mentions does very little, because anti-spam systems filter out manipulation. Google’s main takeaway: optimizing for generative search is still SEO.
A page AI can retell
So what do you do? In the same guide Google recommends unique, first-hand content, not a rehash of what’s already everywhere. In practice it’s a page built around a customer’s question. The heading is the question itself: “Is the hoodie warm?” The first line is a direct answer with a number: “Yes, 380 g/m² fabric, comfy down to −5 °C.” Below that, a table with sizes, fabric and care. Plus your own photos and real reviews, because nobody else has those. A person gets the answer in five seconds, and the AI gets an exact fact that’s easy to quote. The numbers in the example are made up; yours should be your own, checked ones.
Give Google the data directly
There’s also a shortcut people often forget. Google directly recommends connecting Merchant Center if you sell products, and Google Business Profile if you have a shop or local services. Merchant Center is a free account where you upload a feed: a table of products with price, stock and photos. Business Profile is a business card with your address, hours and reviews. That way Google gets exact, fresh first-hand data, and it can show up in AI answers about shopping and in “near me” searches. Without them the AI guesses from other people’s sites, where the price may be a year old.
How to tell GEO is working
Without measurement GEO turns into faith. There are two honest ways. First: Google Search Console has a Generative AI report that shows how people find your pages through AI features in Search and Discover. Search Console is Google’s free tool for site owners. Second: a manual check once a month. Write down 10 customer questions, ask them in ChatGPT, Gemini and AI Mode, and mark whether you got named. AI answers wobble, so look at the trend over a few months, not a single try. And separately: Google states plainly that no third-party tool has access to its internal systems. If someone sells you an “internal AI ranking”, that’s marketing, not data.
Do it yourself: a one-evening GEO check
Five steps you can actually do in one evening. Write down 10 questions customers send you in DMs or to support: those are exactly what they ask AI. Check that your site answers each one in the first line of a page, not three screens down. Add what others don’t have: numbers, your own photos, real reviews. Connect a product feed in Merchant Center or a business profile on Google. And put a monthly AI check plus a look at the Search Console report in your calendar. No need to stuff keywords, buy mentions or pay for llms.txt “for Google”: we already know that doesn’t work.
Check yourself
- What should Nika tell the agency?
Correct answer: Say no: Google doesn’t use llms.txt and there’s no special AI markup. Google says it plainly: Google Search doesn’t use llms.txt, and there’s no special markup for AI features. That $500 is better spent on a page with honest answers and photos. - What should she add to the hoodie page?
Correct answer: An answer in line one, fabric weight, a size chart and her own photos. Specific numbers and first-hand facts give the AI something to quote and the person a quick answer. That’s exactly the kind of edit that worked best in the GEO study. - How can she check the result honestly?
Correct answer: The Generative AI report in Search Console plus a monthly 10-question check. The report gives data from Google itself, and the manual check shows what people see across different AIs. Together, month by month, that’s an honest picture. - How should she read “up to 40%” from the study?
Correct answer: It’s the ceiling in an experiment: results vary by topic, so test on your own site. The authors themselves say the tactics work differently across topics. “Up to 40%” is the best they got in their experiment; only your own measurements will show your result.