GEO: Staying Visible When AI Gives the Answers
An observation from our own consulting practice: inquiries start more and more often with "ChatGPT recommended you …". The journey to a provider no longer begins on a search results page but in a chat window. There are no ten results there, just one answer with three or four cited sources. Either you are one of them, or you do not exist for that prospect.
This guide explains what changes compared to SEO, shows the three measures with the best effort-to-impact ratio, describes how AI search actually selects its sources, and says honestly what you can expect in which timeframe.
What changes and what stays
| Classic SEO | GEO | |
|---|---|---|
| Goal | Ranking on the results page | Being cited in the AI answer |
| Unit | Page ranks for keyword | Statement answers question |
| Click | User comes to the website | Answer often without a click; the brand must appear in the answer |
| Success metrics | Position, clicks, CTR | Citations, brand mentions, AI referrals |
| Foundation | Crawlability, links, content | unchanged, plus structure and verifiability |
The table shows: GEO does not replace SEO, it raises its quality bar. AI systems reward what good editors always wanted: clear statements instead of marketing fog, backed numbers instead of superlatives, answered questions instead of keyword density. If you have seriously invested in content quality over recent years, you start with a head start.
Measure 1: Find the questions your pages must answer
AI systems extract answers. A page that answers "What does a website relaunch cost?" with ranges, influencing factors and examples is citable. A page that promises "individual solutions for your success" is not.
The harder task is usually not answering but finding the right questions. You already own three sources for them:
- Google Search Console. Filter the queries report for question words like "how", "what", "why" and "cost". There you find the phrasings users already use to find you, and the questions where you sit just behind page one.
- Sales inbox and first calls. The questions prospects ask before commissioning you are the same ones they ask an AI beforehand. Collect the ten most frequent from the past six months.
- Support tickets and questions from existing customers. They show where explanation is needed that your website does not yet cover.
From this collection a simple mapping emerges: one core question per important page, answered completely on that page, with numbers, ranges and at least one example. Pages that answer no recognisable question get merged or reworked.
Measure 2: Structure and schema machines can read
A clean heading hierarchy, a summary at the top of the page and Schema.org markup are craft. But they decide whether a model attributes your content correctly. For a B2B website, this order of schema types has proven itself in our projects:
- 1. Organization: establishes who you are, with name, address, logo and profiles. The base all other markup refers to.
- 2. Person: makes authors and contacts identifiable and links content to verifiable expertise.
- 3. Article or BlogPosting: ties articles to an author and a date, so authorship and freshness are machine-readable.
- 4. FAQPage: marks up answered questions so they can be extracted directly as question-answer pairs.
- 5. Service: describes your services with provider, region and scope.
The order follows the effort-to-impact ratio: Organization and Person take a few hours and affect every page. One rule applies to all types: markup only describes what is visibly on the page. Markup without content does more harm than good.
Measure 3: Show expertise instead of claiming it (E-E-A-T)
AI systems weight identifiable authorship and consistent company signals. Behind the acronym E-E-A-T (Experience, Expertise, Authoritativeness, Trustworthiness) sit three very concrete work sites.
Author pages. Every article needs a real person as its sender: a profile page with photo, role, focus areas and links to further posts by the same person. An anonymous "editorial team" is equally unconvincing to humans and machines.
Consistent company data. Name, address and phone number must be identical on your website, Google Business Profile, LinkedIn and in industry directories. Contradictions there are noise in exactly the data machines use to pin down your identity.
Experience in the text. Experience shows in passages only someone who has done the work can write: concrete project observations ("in our projects this typically takes four to six weeks"), real numbers with context, screenshots from your own systems, even named failures. Generic advice paragraphs that could appear word for word on any competitor's site do not deliver this signal.
How AI search selects its sources
So the three measures do not read like articles of faith, it helps to look at the mechanism. The large systems work in two stages. First they search for candidates: Google AI Overviews draw on the normal Google index, ChatGPT and Perplexity use their own crawlers and search partners. At this stage, the classic ranking factors still apply. If you cannot be found on a topic, you never enter the selection.
In the second stage, the model decides which passages to use and cite in the answer. Here the winner is not the page but the passage: a paragraph that answers the question directly, completely and verifiably. Measures 1 to 3 pay into exactly this stage. How this selection works technically is explained in our knowledge article on RAG.
An honest caveat belongs here: the systems differ in detail, their selection logic is not public, and it changes continuously. What we describe is the common base mechanism as it can be read today from vendor statements and our own tests, not a guarantee in any individual case.
What you can realistically expect
GEO is not a switch. As a rough timeline from our client projects: technical changes like schema markup get recrawled within days to weeks. Until reworked content appears in AI answers, expect weeks to months, depending on topic, competition and how often the respective system refreshes its sources.
What you should not expect: a quick traffic jump, guaranteed citations on your preferred topics, or control over the wording of answers. The same question may cite your source today and a different one tomorrow; answers vary by session and phrasing.
What you may expect: first citations on sharp specialist questions where you have the best answer in the market, and a small but growing segment of AI referrals. In conversations with our clients, exactly these visitors stand out as valuable: they arrive pre-qualified, because an AI has already used your content as an answer.
How to recognise progress
GEO success is measurable, just differently. Watch AI referrals in your web analytics, meaning visitors from chatgpt.com, perplexity.ai and similar sources; how to set this up cleanly is shown in our post on measuring AI referrals. Check regularly how AI systems talk about your company and your services. Track brand mentions in AI answers on your most important topics. We built a free GEO Score Check that tests the technical side in seconds.
Do we need an llms.txt?
Not strictly, at the moment. llms.txt is a proposal that lets websites offer AI systems a curated content overview. No major system has so far committed to evaluating the file. It costs little and does no harm, but it replaces none of the three measures in this article. We use it on our own website, but treat it as a bet, not an obligation.
Will we lose traffic to AI Overviews?
For purely informational queries, click-through rates drop; that matches industry studies and our observation in client projects. What matters is what happens with purchase-related queries: there the click survives more often, and whoever is named in the AI answer receives the better-prepared visitor. The more important question is therefore not whether traffic drops, but whether your brand appears in the answers.
Is GEO worthwhile for small websites too?
Yes, often especially so. AI answers cite the best passage for a concrete question, not automatically the biggest brand. A small website with genuine specialist expertise can be cited on sharp questions where large generalists offer nothing substantial. The prerequisite remains that the content is findable, structured and carries an identifiable author.
How do we measure citations concretely?
On three levels. First, AI referrals in your web analytics as the hard number. Second, regular spot checks with the same prompts in ChatGPT, Perplexity and Google to see who gets named on your core topics. Third, brand mentions in the answers even without a link. The approach with a concrete setup is described in our post Measuring AI referrals.
The first step for this week
There is less room in AI answers than on results pages. Where three sources are cited instead of ten links, the early mover occupies a spot that becomes expensive for latecomers. The sensible start costs one afternoon: pick your most important service page, formulate the one question that page should answer, and check honestly whether the complete answer is there. Our GEO Score Check tests the technical side in seconds. Once both are in place, move on to the next page.
Go deeper in our knowledge base
Want to know what these topics mean for your company? The Future Check shows you the biggest levers within 2–4 weeks.
