How to do GEO step by step: 7 steps to get cited in AI answers

What is GEO optimization?

GEO optimization (generative engine optimization) is the practice of preparing content so that generative AI systems such as ChatGPT, Perplexity, Google AI Overviews and Gemini cite and link it as a source. The goal is not first place in a list of links but a mention in the answer the AI generates for the user. Why such mentions carry even more weight for AI systems than backlinks is covered in our guide on digital PR for AI visibility.

The difference from classic search matters for every step below. On Google, the user sees ten blue links and decides which one to click. In an AI answer, the user gets a finished paragraph, and only the two to five sources the model used appear as a footnote or link. If you are not among them, you do not exist for that user at that moment.

A concrete example: when someone asks ChatGPT "Which SEO agency in eastern Switzerland has experience with AI visibility?", the model does not search the whole web live. It draws on its training knowledge and connected search sources for pages that answer the topic clearly, with evidence and in a machine-readable way. An agency website that only says "We are your partner for digital marketing" gives the model nothing to extract. A page that states in its first sentence what it offers, which canton it is based in and which results it can prove does.

The terms and the theory behind them are explained in our guide What is GEO?. This article focuses on the practical steps.

GEO vs. SEO: where is the difference?

The core difference: SEO competes for a position in the list of links, GEO for a citation in the AI answer. SEO works to make your page appear high in Google or Bing results and get clicked; GEO works to make a language model use your content as evidence in its generated answer, often without any click.

Both share the same foundation. For a language model to cite your page, it first has to be crawlable, technically clean and relevant, which is exactly what classic SEO delivers. GEO builds on that and adds how the content is prepared: extractable direct answers instead of long introductions, structured facts instead of walls of text, sourced numbers instead of claims. If you do GEO without SEO, you optimize a page the model never finds. How to add GEO to existing SEO work is covered in the guide Combining SEO and GEO.

Measurement shifts as well. In SEO you measure rankings, clicks and click-through rate in Google Search Console. In GEO you measure whether and how often your content appears in the models' answers, a metric often called share of model or AI visibility. The table compares the two directly:

In practice, you do not throw away your SEO; you extend it. A well-ranking service page stays important, but you add the building blocks a model needs to cite it. Both channels draw on the same content, just through different levers. A deeper comparison, including synergies, is in the article LLMO vs. SEO: differences & synergies.

How does an AI decide whom to cite?

An AI cites the sources that answer a question most clearly, with the best evidence and in the most extractable form, not necessarily those with the most backlinks. Language models look for self-contained statements they can use in their answer without rewording, and they prefer content whose statements they find confirmed across several sources.

Three factors carry particular weight. First, extractability: a sentence like "Since January 1, 2024, the VAT rate in Switzerland has been 8.1 percent" can be quoted directly; a paragraph that only gets to the point after three sentences cannot. Second, verifiability: models give more weight to statements backed by a source, a date or a traceable number. Third, a consistent entity: if your brand, location and offer are named the same way across the whole page, the model can attribute the content to you without doubt.

A practical example from everyday SME work: two fiduciary firms in St. Gallen both answer the question "When do I have to register my company for VAT?". One writes promotional copy about its expertise; the other puts the question in an H2 and answers it in the first sentence with the specific revenue threshold and a source. The model will very likely cite the second, because its answer can be used as is. How AI search systems weigh and select sources internally is covered in the article How AI search systems select sources.

GEO optimization in 7 steps

These seven steps turn an existing page into a source AI systems can cite. Work through them in order. You do not have to rewrite everything: usually it is enough to restructure existing content, add evidence and make it machine-readable.

1. Put a direct answer at the start of every section. Goal: a sentence the AI can lift as is. Action: start every section with one or two sentences that answer the heading's question fully and on their own, with no preamble and no reference back to the previous paragraph. If you write about "the cost of a website", the first sentence is not "Many factors play a role" but "A professional SME website in Switzerland typically costs between X and Y CHF, depending on …". That is the sentence the AI extracts. Check: read only the first sentence under each heading. Does it answer the question?
2. Add FAQ schema (FAQPage JSON-LD) with 40–60-word answers. Goal: ready-made question-and-answer pairs. Action: collect four to ten real questions your audience actually asks (from Search Console or Google's "People also ask" box) and answer each one on its own in 40 to 60 words. Also add these questions as FAQPage schema in the source code. The page then hands the model clearly marked question-and-answer pairs, the format AI answers use most readily. Check: every visible FAQ answer matches its schema text and stands on its own.
3. Name entities clearly. Goal: the model knows exactly who and what the page is about. Action: spell out brands, tools, places and standards instead of talking about "solutions" or "platforms". Write "Google Analytics 4" instead of "our analytics tool", "Canton of Thurgau" instead of "the region", "CHF" instead of "franc amounts". The clearer the entity, the more reliably the model can attribute your content to a specific provider, place or topic, and the more likely it is to name you. Check: your brand, location and offer are named the same way throughout the page.
4. Back up facts: number plus source plus date. Goal: statements the model can verify. Action: every number in your content needs a source and a time reference, for example "according to the Federal Statistical Office, as of 2025". Invented or vague statistics hurt twice: they cost trust with people, and models that cross-check statements are more likely to discard them. A visible "Updated: June 2026" on the page also signals freshness; models prefer current, dated content. Check: no number on the page is without a source and date.
5. Structure content for machines. Goal: content a model can parse cleanly. Action: turn running text into structures: comparison tables for "X vs. Y", numbered lists for processes, short paragraphs of two to four sentences and descriptive subheadings that are real questions or statements. A table comparing three pricing options gets extracted far more often than the same content in three nested sentences. Check: comparisons sit in tables, processes in numbered lists, and no paragraph runs longer than four sentences.
6. Link internally with descriptive anchor text. Goal: the model understands how your pages relate. Action: connect related pages with anchor texts that name the target, such as "guide to AI visibility" instead of "click here". This helps models understand the topical connection between your pages and strengthens the page you point to as the main source on the topic. Build a cluster of several articles around your core topic that link to each other. Check: no "click here" links remain, and every cluster article links to the main page.
7. Measure and refine. Goal: know for which questions you are cited. Action: regularly check across several models whether, and for which questions, you get cited by entering your target questions directly in ChatGPT, Perplexity and Gemini and logging the sources they name. If you do not appear, refine the section in question, usually the direct answer or the evidence. GEO is not a one-off project but ongoing tuning. Check: you have a log per model and per question that you repeat at regular intervals.

Which content gets cited most often?

The content cited most often answers a clearly stated question in an extractable first sentence and backs it with a verifiable number or source. Definitions, how-to guides, comparisons and fact-dense FAQ sections are the formats language models prefer as evidence.

In concrete terms, four patterns recur. Definitions ("What is X?") give the model a finished explanation it can use almost word for word. How-to content in numbered steps matches the way AI presents instructions. Comparisons in table form ("X vs. Y") provide structured side-by-sides a model can pour straight into an answer. And FAQ blocks with short, self-contained answers are the ideal question-and-answer format for AI systems.

So do not write the tenth generic industry article; answer your customers' specific questions with a local angle. A plumbing business in Winterthur that answers "What does an emergency plumber cost on a weekend in the Canton of Zurich?" with a concrete price range and a source has a real chance of being cited, because hardly anyone else answers that precise, local question with facts. Generic content without substance, or marketing phrases without facts, practically never gets cited.

How do you measure the success of your GEO work?

You measure GEO success by whether your content appears as a source in the answers of ChatGPT, Perplexity, Google AI Overviews and Gemini, and for how many of your target questions. This metric is often called share of model or AI visibility, the GEO counterpart to ranking position in classic SEO.

The simplest method is a manual spot check: enter your most important questions one after the other in the different models and note whether your domain appears as a source. Check across several models: a page can be cited in Perplexity but not in ChatGPT, because the systems use different sources and weightings. If you want to do this systematically and repeatably, use tools that run these queries automatically across several models.

Google Search Console helps as well: rising impressions for question-style searches are often an early sign that AI Overviews are including your page too. A full guide to methods and tools is in our article Measuring AI visibility. What changes in AI search are coming in 2026 and how measurement is evolving is covered in the Google I/O 2026 GEO roadmap.

GEO optimization for Swiss SMEs: is it worth it?

For small and mid-sized companies, GEO is worth it especially when potential customers research before choosing a provider, because the AI answer increasingly makes the first shortlist before anyone opens a website. For specific, local questions, the leverage is often greater than in classic SEO, because the clearest, best-sourced answer tends to win, not the largest domain. How to plan this with limited budget and staff is covered in our GEO strategy for SMEs.

SEOX serves more than 50 satisfied clients and has achieved more than 1,000 top-1 rankings over 25+ years, with 95 percent client satisfaction (source: SEOX, as of June 2026). This experience from classic search is the foundation GEO builds on: without a clean technical base and relevant content, there is no citation. If you want to know where your company stands in AI answers today, start with our free analysis or look at our LLMO service. And if you want to keep an eye on the next stage of discoverability, read about Agentic Resource Discovery (ARD), the new standard AI agents use to find and verify services.

If you appear as a source in the AI answer, you make the shortlist before the first click. That is what gives GEO its leverage for SMEs, provided every statement is extractable and backed by evidence.