What is GEO optimization?
GEO optimization (Generative Engine Optimization) is the systematic work of making generative AI systems mention a company in their answers, recommend it and cite its content as a source. That means ChatGPT, Google Gemini, Perplexity, Microsoft Copilot as well as Google AI Overviews and AI Mode — every interface that returns a synthesised answer instead of a list of links.
The same discipline goes by other names: AEO (Answer Engine Optimization), LLMO, AIO, GAIO. They mean essentially the same thing. We use GEO because it has become the standard term. What GEO is not: a replacement for SEO. An AI system that searches the web draws on the same search indexes as Google and Bing — a page that cannot be found there will not appear in the answer either.
„No ranking. No grounding. No AI visibility.“
Why is SEO alone no longer enough?
SEO alone is no longer enough because a good Google ranking no longer means a potential customer gets to see the page. Someone who asks ChatGPT or Google AI Overviews for suppliers receives an answer with three to five names — and only those companies get the enquiry. In our study of 449 German companies from 258 industries (3,592 buyer questions, 7,184 documented AI answers), 55 % of buyer questions went without a mention of the company examined, and in 47 % the AI named a competitor instead. Good rankings did not protect against it.
-
55 %
of buyer questions go without a mention of the company
CodaAI study 2026, 7,184 AI answers
-
47 %
are answered with a competitor instead
CodaAI study 2026
-
26 %
of brands have not a single mention in Google AI Overviews
Ahrefs, 75,000 brands
-
< 1:100
chance that ChatGPT or Google AI returns the same recommendation list twice
SparkToro, 2,961 runs
The difference is not in the craft but in the weighting and the result format. The table sets the two disciplines side by side:
| Aspect | SEO (classic search) | GEO (AI answers) |
|---|---|---|
| Goal | Rank on page one and win the click | Be named in the answer and cited as a source |
| Result format | A list of links; the user chooses | One synthesised answer with a few sources |
| Competition | Ten positions on page one | A few mentions per answer — whoever is missing is not seen |
| Unit of evaluation | The page (URL) | The passage — extracted excerpts of a few hundred words |
| What counts | Keywords, links, technical basics | The same basics plus fact density, structure, brand mentions on third-party sites |
| Measuring success | Position, impressions, clicks — reproducible | Mention rate, citation rate, share of AI search — as a probability over many runs |
| Stability | Rankings stay similar for weeks | Every answer is generated anew; the same question is answered differently |
The overlap is larger than most people think: technical accessibility, content quality and authority work on both sides. GEO shifts the priorities and adds a few requirements — it does not start from zero.
What does the study show for B2B companies?
The study shows, for B2B companies, that classic search-engine strength does not translate into AI answers. In June and July 2026 we examined 449 mid-sized German companies from 258 industries: eight qualified buyer questions each in three archetypes (market overview, use case, competitor comparison), put to ChatGPT and Google AI Overviews — 3,592 questions, 7,184 documented answers. Four findings carry this guide:
The SEO paradox. 361 of the companies examined have more than 100 top-10 rankings on Google. 35 % of them are nevertheless completely invisible in ChatGPT — zero mentions across eight questions. More than a third of the firms that did everything right on Google do not exist for the AI. The cause is in the next section: ChatGPT answers mostly from training data, and whoever is missing there is missing structurally.
Invisible in their own core business. 26 % of companies are not mentioned even for the market-overview question about their own category (“Which providers of … are considered leaders?”). That is the question a buyer starts the discovery phase with. Whoever is missing here never makes it onto the shortlist.
Competitors instead of a gap. An AI answer does not stay empty when a company is missing — it names someone else. For 47 % of companies the AI predominantly recommends competitors by name. Where companies are mentioned, they sit at position 1.8 on average — the answer is short, and the first two places decide.
Two systems, two realities. The same questions yielded an average visibility of 24.9 % in ChatGPT and 39.4 % in Google AI Overviews. 39.6 % of companies are never mentioned by ChatGPT, 15.1 % not by Google either. Google AI Overviews draw more on the live index — grounding works faster there; in ChatGPT the brand first has to make it into model knowledge or into the candidate pool via third-party sites.
What follows for B2B: the third-party studies further down measure large brands in the English-speaking market. Our figures measure German mid-sized companies with niche products — and there the gap is wider, because model knowledge about a firm with 200 employees is thin and grounding remains the only route into the answer.
How does content get into an AI answer?
Content gets into an AI answer in one of two ways: through model knowledge, built during training from sources such as Common Crawl and Wikipedia, or through web search, which the system triggers when it is not confident it can answer from memory. Model knowledge ends at the knowledge cutoff; a company that has rebranded, repositioned, spun off or merged since then does not exist there — a frequent case in mid-sized B2B, and the one the study makes visible as the “SEO paradox”. For anything current, web search is the only route, and it runs in four steps:
- Query fan-out: The system splits the question into several sub-queries and sends them to a search index.
- Initial retrieval: For each sub-query, the top-ranked pages enter the candidate pool — this is the classic SEO part.
- Re-ranking: Within the candidates, individual passages are scored. Only what clears the relevance threshold moves on. Being retrieved does not mean being cited.
- Grounding snippets and answer: The most relevant excerpts are extracted and composed into the answer; the sources appear as citations.
Dan Petrovic (dejan.ai) measured how narrow this bottleneck is across 7,060 Google queries and 2,275 source pages: each query has roughly 2,000 words of grounding budget, distributed by relevance rank. The top-ranked source receives 531 words on average (28 %), the fifth still 266 words (13 %). From a typical page, 377 words are selected, in chunks of around 15 words. This yields the rule that carries everything else:
Which measures demonstrably raise the probability of being cited?
The probability of being cited rises on three levels that do not substitute for each other but multiply. Chrissy Kunisch (ONE Beyond Search) put it into a formula at the SISTRIX Meetup in September 2026: AI visibility = (technology + content + off-page) × processes. Technology gets a page into the race, content decides whether it is cited, the environment on third-party sites decides whether the brand counts as a candidate at all — and without measurement every measure remains a one-off.
Technology: the entry ticket
Technology decides whether an AI system can read the page at all. The most important difference from Google: no crawler from OpenAI, Anthropic, Perplexity, Meta or ByteDance executes JavaScript (Vercel and MERJ, December 2024). ClaudeBot loads JavaScript files in 23.8 % of its requests, GPTBot in 11.5 % — none of it is executed. Content that only appears in the browser is invisible to these systems. Gemini and Applebot, by contrast, do render.
- Allow AI crawlers in robots.txt (GPTBot, ClaudeBot, PerplexityBot, Google-Extended) and list only indexable URLs in the sitemap.
- Deliver important content as HTML rather than loading it via JavaScript; lean, semantic markup with a clear heading hierarchy H1 → H2 → H3.
- Keep response time low and put no click or login walls in front of important content — an aborted request is usually not retried.
- Internal linking as real HTML links, so that URL discovery and crawl budget land on the pages that matter.
Content: decides whether you are cited
Content gets cited when it is dense with facts, clearly structured and understandable without context. The most thorough measurement to date comes from Princeton: for KDD 2024, Aggarwal et al. tested nine text measures on 10,000 queries (GEO-bench). The three most effective add something to the text that a model cannot generate itself:
| Measure | What it does | Change |
|---|---|---|
| Add quotations (Quotation Addition) | Verbatim statements by named experts in the text | + 41 % |
| Add statistics (Statistics Addition) | Quantitative data instead of qualitative description | + 34 % |
| Cite sources (Cite Sources) | Evidence with its origin in the sentence | + 28 % · for pages ranked fifth: + 115 % |
| Improve fluency | Short, clear, conversational sentences | + 28 % |
| Keyword stuffing | Sprinkle in more relevant keywords | no gain; on Perplexity − 10 % |
The best combination in the study was fluency plus statistics. The effect on weaker-ranked pages stands out: citing sources raised the visibility of pages ranked fifth by 115.1 % — GEO works hardest where classic SEO has not yet won. The figures come from English-language tests under lab conditions; they show the direction, not a guarantee.
Add to that the structural rules that follow from how extraction works: answer first (key message in the first 30 % of the page), semantic chunking (one idea per paragraph, each paragraph understandable on its own), headings as questions and entity echoing (the answer starts with the term asked about). Comparisons belong in tables, criteria in lists: in the Wix Studio AI Search Lab analysis of 75,000 AI answers with more than one million citations (March 2026), 21.9 % of all citations went to listicles, 16.7 % to articles and 13.7 % to product pages; for commercial queries, listicles led with 40.9 %.
Freshness counts measurably: AI assistants cite content that is on average 25.7 % fresher than organic Google search (Ahrefs, 16.975 million cited URLs, July 2025). Google AI Overviews are the exception — they cite content that is actually 16 days older than organic search. A visible date, updated figures and a maintained dateModified in the schema therefore belong to routine, not to a relaunch.
Off-page: decides whether the brand counts as a candidate
Off-page signals weigh more heavily with AI systems than with Google, because the AI does not care whether a piece of information comes from your own website or from third-party sites — what counts is a consistent brand picture. Ahrefs measured for 75,000 brands which factors correlate with mentions in Google AI Overviews: brand mentions on the web correlate at 0.664, backlinks at only 0.218 (Spearman). 26 % of the brands examined had not a single mention. Correlation is not causation, and the study looks at large brands — the order of the factors is nevertheless unambiguous.
- Presence in comparison lists: In an analysis of around 1,260 B2B buying prompts (Overthink Group, July 2026), 70.8 % of all citations pointed to pages with “best”, “top” or “leading” in the title. A company missing from the lists the AI already cites is missing from the answer.
- Your own homepage: Among the 1,000 most-cited pages in ChatGPT, 23.8 % are home and landing pages (Ahrefs, October 2025) — the second-largest category after Wikipedia, and the only one the company fully owns. A consistent brand description there and on every profile is the foundation.
- Review platforms, trade media, digital PR: mentions in sources the AI already cites in your industry — even without a link. In B2B that is not Trustpilot or Google reviews but industry portals, trade media, association directories and software comparison platforms such as Capterra or G2. Named experts with a quote are picked up more often than anonymous editorial teams.
- Your own videos with transcripts and entities in knowledge bases: both are sources models are trained on and prefer during grounding.
What does not work in GEO?
Much of what is currently being sold does not work in GEO. The following five points are measured, not opined — and they save budget:
| Measure | Finding | Source |
|---|---|---|
| Keyword stuffing | No gain; on Perplexity 10 % worse than the unchanged text | Princeton GEO study, KDD 2024 |
| llms.txt | No significant relationship with citations across roughly 300,000 domains; 97 % of the files across 137,210 domains examined were never requested; according to Google, no major provider uses the file | SE Ranking · Ahrefs · SISTRIX |
| Schema markup as a citation guarantee | No correlation between schema coverage and citation rate (Search Atlas, December 2024). Schema helps Google and Bing understand entities — it is infrastructure, not a lever | Search Engine Land, March 2026 |
| Server logs as proof of visibility | Google AI Overviews and AI Mode generally do not fetch live; a request by ChatGPT-User means “checked”, not “cited” | SISTRIX, March 2026 |
| Ranking promises | Across 2,961 runs, the chance that ChatGPT or Google AI returned the same list twice was below 1 in 100 | SparkToro, January 2026 |
„Anyone who says today “I only want to be visible in Perplexity” is optimizing for a very small window. Platform-specific fine-tuning comes in a second step — once the basics are in place.“
For the same reason there are six things we deliberately do not offer — among them llms.txt as a service, Reddit seeding in B2B and ranking guarantees. The list is on the service page.
How do you measure the success of GEO optimization?
The success of GEO optimization is measured as a probability, not as a position. A ranking is reproducible: same keyword, same location, same list. An AI answer is generated anew every time. Rand Fishkin (SparkToro) demonstrated this with 2,961 runs across twelve prompts in ChatGPT, Claude and Google AI: the chance of getting the same recommendation list twice was below 1 in 100; for the same order, below 1 in 1,000. A tool that reports a single visibility score is therefore measuring noise.
What holds up is a fixed prompt set, queried repeatedly and regularly, and three metrics derived from it: the mention rate (in how many answers the company is named), the citation rate (in how many its own page is linked as a source) and the share of AI search compared with competitors. Plus sentiment: mentioned does not mean recommended, and a false statement about the company is more urgent than none.
- Choose prompts along the buying decision. Recommendation, comparison and trust questions are closest to the decision; general research questions are worth the least. The table below shows the categories with B2B examples.
- A baseline measurement before the first measure. Without a reference value, nothing can be proven eight weeks later.
- Query model knowledge separately. What does the AI say about the company without web search? Wrong service, old location — then the task lies with the brand description, not the content.
- Source analysis: Which third-party sites does the AI cite for your prompts? That is the list you need to be on.
| Category | B2B example | Influence on the buying decision |
|---|---|---|
| General research | “What matters when choosing a conveyor system?” | low — no buying intent visible |
| Problem solving | “How can cutting fluids be used longer in machining?” | medium — problem exists, supplier still open |
| Market overview / recommendation | “Which manufacturers of aluminium profile systems are considered leaders in Europe?” | high — this is where the shortlist is formed |
| Comparison | “Supplier A or supplier B for surface finishing of small parts?” | high — just before the decision |
| Trust / reputation | “Is there any experience with supplier A on large production runs?” | high — whoever fails to convince here has lost the buyer |
| Purchase / closing | “Where can I get supplier A with a lead time under four weeks?” | medium — brand is set, availability counts |
This is also how our Digital Visibility Audit measures: on real buyer questions, in ChatGPT and Google AI Overviews, as a share over runs — with the competitors named instead shown alongside.
Where do you start with GEO optimization?
GEO optimization starts with what lays the foundation for all platforms at once — not with fine-tuning for a single system. The order that follows from the evidence above:
- Measure. Define the prompt set, run the baseline measurement, query model knowledge. Only then do you know whether the problem lies with technology, content or brand.
- Become readable. robots.txt, JavaScript-free content, response time, clean structure — Tier 1 · Found.
- Write to be cited. Answer first, one idea per paragraph, figures with their source in the sentence, comparisons as tables, a visible date. Existing pages first, then new ones.
- Be present where the AI gets its evidence. Consistent brand description on every profile, your industry's review platforms, comparison lists, trade media — Tier 2 · Recommended.
- Become the source yourself. Your own data, your own studies, named experts — content for which there is only one source: your company. Tier 3 · Cited.
To go deeper into the terms, see the GEO Glossary — 56 terms, one page each, in the order in which content makes its way into an AI answer.