Generative engine optimization (GEO): what it is and what works for software companies

What is generative engine optimization?

GEO means shaping your content and web presence so that generative AI systems mention your company, describe it accurately and cite you. “Generative engines” answer with written text instead of links: ChatGPT, Perplexity, Gemini, Claude, Microsoft Copilot, and Google’s AI Overviews and AI Mode.

Definition: GEO is optimizing for visibility inside AI-generated answers. Success isn’t position one. It’s being named for the questions that matter, described correctly and linked.

Most of these systems first retrieve relevant pages, usually from a search index. A language model then condenses what it found into an answer, adding what it learned in training. GEO works on both. This guide is written for B2B software companies selling into Germany, Austria and Switzerland, but most of it applies anywhere.

GEO, AEO, LLMO: one practice, several names

The labels stress different things, but they describe the same practice:

• GEO (generative engine optimization): the research term, focused on generative search systems.
• AEO (answer engine optimization): focused on answers rather than ranked results.
• LLMO (large language model optimization) and LLM SEO: focused on the language models.
• AI SEO: the umbrella term for the same work.

Google treats AEO and GEO as labels for this work and is blunt about where they belong: from Google Search’s perspective, optimizing for generative AI search “is optimizing for the search experience, and thus still SEO.” Other assistants run their own search crawlers, so the overlap with SEO is large there too.

Where the term comes from: the GEO paper

The term comes from a paper by Pranjal Aggarwal and five co-authors, including researchers at Princeton University and IIT Delhi. It appeared as a preprint on arXiv in November 2023 and at the KDD conference in 2024.

The team built a test engine: for each query, GPT-3.5 wrote a cited answer from the top five Google results. A language model then rewrote one source with one of nine methods, and the team measured how much of the answer that source got, across 1,000 test queries from their GEO-bench dataset.

Key findings

• Evidence wins. Adding citations, quotations and statistics did best: 30 to 40% more visibility on the word-based metric and 15 to 30% on a subjective one. That’s where the much-quoted “up to 40%” comes from.
• Keyword stuffing and tone don’t. More keywords barely helped (about 10% worse on Perplexity), and an authoritative style brought no significant gain.
• Lower-ranked sources gain most. With added citations, the source ranked fifth on Google gained about 115% in visibility, while the top-ranked source lost about 30% on average.
• Live test. On Perplexity, gains reached 22% (word-based) and 37% (subjective).

Limits of the study

• Lab conditions with GPT-3.5 and five sources. The Perplexity test used 200 queries with uploaded files, not live search.
• It measured share of the answer, not clicks or revenue. Shares add up to one, so one source’s gain is another’s loss.
• About 80% of queries were informational, on topics like history, health and science. The dataset isn’t built around software buying.
• Only pages already in the top five were optimized. How a page gets selected at all is outside the study.

In practice: specific, sourced content is more likely to make it into answers. But getting selected at all still depends on indexing, crawler access and presence on other sites.

GEO vs. SEO (and AEO, LLMO)

GEO doesn’t replace SEO, it builds on it: without indexable pages there’s little for an AI system to cite. The differences lie in the outcome you aim for and how you measure it.

The overlap is large: Google says its AI features are rooted in its core Search ranking systems, and OpenAI, Anthropic and Perplexity run their own search crawlers. Three things are new. What others write about you feeds into answers. Facts such as pricing must match everywhere. And success has no fixed position to measure. For the SEO foundation, see SEO for software companies.

Google AI Overviews and AI Mode

Only pages that are indexed and eligible to show with a snippet in regular Search can appear as sources in Google’s AI features. Google says there are no additional technical requirements.

In Germany, AI Overviews have been live since 26 March 2025, at launch for signed-in users aged 18 and over. AI Mode, a conversational search with follow-up questions, began rolling out in Germany, Austria and Switzerland on 8 October 2025.

AI Overviews cost clicks. According to SISTRIX, they appear for just over 20% of keywords in Germany, and click-through for position one then falls from about 27% to 11% (more than 100 million keywords; method not published in detail). In a Pew Research Center study of 900 US adults, users clicked a traditional result in 8% of visits with an AI summary, 15% without.

How Google builds the answers

Google retrieves pages through its regular ranking systems and grounds the answer in them. With query fan-out, it splits a question into subtopics and runs several searches at once. A question about the right tool can trigger sub-searches on pricing, integrations or alternatives, so pages that answer those clearly get more chances. Building a page per variant to manipulate answers, though, counts as scaled content abuse.

Controls and reporting in Search Console

Since 31 August 2026, every site has a Search generative AI control in Search Console. Sites are included by default. Excluding yours removes your links from AI Overviews and AI Mode; Google says the rest of Search is unaffected. Since the same date, a generative AI performance report counts impressions of those links by page, country, date and device. Its help page lists no query breakdown, so which buyer questions you appear for still takes your own sampling.

For robots.txt: Googlebot governs Search, including its AI features. Google-Extended covers Gemini, not Search (see AI crawlers).

What works for software companies, and what’s overhyped

For B2B software, the levers that pay off answer buyers’ comparison questions and keep your facts identical everywhere. None is guaranteed. And comparison work is moving into AI tools: in a survey by review platform G2 of 1,076 B2B decision makers worldwide (March 2026), 51% said they start software research with an AI chatbot more often than with Google, and 69% said a chatbot led them to a different vendor than planned. G2 has a stake in AI search; treat these as upper bounds.

Comparison and alternatives pages

Buyers ask for “alternatives to X”, “X vs. Y” and “the best software for Z”. Honest answers give the system something to quote: criteria, who each product fits, dated prices. Pure self-promotion offers little to repeat neutrally; in the GEO paper, too, tone didn’t help and evidence did.

Consistent facts on pricing and features

AI answers assemble your product from your site, review profiles and old articles. If they disagree, an outdated version can end up in the answer. Keep pricing model, features, target customer and company name consistent everywhere, and on your site as text, not just images or PDFs.

Review sites and trade media

Google says its AI features can show what’s being said about products and services across the web, including blogs, videos and forums, while seeking inauthentic mentions isn’t as helpful as it might seem. For software sold in DACH: maintained profiles on OMR Reviews, Capterra or G2, genuine reviews, and mentions in trade articles and comparison lists.

Crawler access and structured data

If robots.txt or bot protection blocks the AI vendors’ search crawlers, your pages will rarely appear as sources (which crawlers matter). Structured data such as Organization and SoftwareApplication makes facts machine-readable, but Google’s AI features don’t require it, and it should match the visible text.

What’s overhyped

• llms.txt: proposed by Jeremy Howard in 2024 as a Markdown summary of a site for language models. Google says Google Search doesn’t use such files; they neither help nor hurt. As of 17 October 2026 we found no confirmation from OpenAI, Anthropic or Perplexity that they read it for answers, and their crawler docs don’t name it as a signal for search or answers. Some vendors publish an llms.txt for their own developer docs, which is a different matter.
• Chunking content or rewriting it for AI: unnecessary, according to Google.
• Keyword stuffing: no effect in the GEO paper.
• Tracking positions in AI answers: the order is close to random (see measurement).

For the ChatGPT-specific playbook, see our guide on getting mentioned in ChatGPT.

Local-language prompts, English sources

Even when your buyers ask in German, an AI assistant may run part of its research in English. So English sources matter even if you only sell in DACH.

Berlin-based tracking vendor Peec AI analyzed more than 10 million prompts and 20 million background searches from its platform in February 2026, limited to cases where language and location matched (such as German from Germany). 43% of ChatGPT’s background searches for non-English prompts ran in English, and in about 78% of cases ChatGPT added English searches to its research.

A caveat: Peec sells visibility tracking, the method isn’t independently reviewed, and we found no official OpenAI description of how ChatGPT picks its search language. Check it yourself: ask ChatGPT a German question about your category with web search on, and note the language of the cited sources.

• Make your core facts (product, category, pricing, integrations) findable in English too, on your site or in English-language profiles.
• Include English review sites and comparison lists in your category.
• Keep pricing and feature names identical across languages, but don’t machine-translate everything: your buyers still read German.

Takeaway: If you only talk about yourself in German, you may be missing from the part of the research AI assistants do in English.

AI crawlers: training vs. search

The major AI vendors separate training crawlers from crawlers for search and answers. Mix them up, and you may block yourself from answers when you only meant to opt out of training.

Fetchers acting on a user’s direct request (ChatGPT-User, Claude-User, Perplexity-User) follow their own rules: OpenAI says robots.txt may not apply to ChatGPT-User, and Perplexity says Perplexity-User generally ignores it. Anthropic says its bots honor robots.txt.

Our recommendation: allow the search crawlers. Training crawlers are a business decision; OpenAI and Google both say blocking them doesn’t affect search. One exception: Google-Extended also controls whether the Gemini app uses your content to ground its answers. Also check your firewall and CDN bot protection, which can block crawlers regardless of robots.txt. OpenAI applies robots.txt changes in about 24 hours, Perplexity within 24 hours.

Measuring GEO: mention rates, not single answers

One AI answer is not a metric: the same prompt gives a different answer on each run. Numbers only mean something across many prompts and repeated runs.

An experiment by SparkToro and tracking vendor Gumshoe (January 2026) shows the spread: 600 volunteers ran 12 recommendation prompts through ChatGPT, Claude and Google’s AI search 2,961 times. On ChatGPT and Google’s AI search, the chance of getting the same brand list twice was below 1 in 100, and the same order closer to 1 in 1,000. How often a brand appeared at all was more stable. The study isn’t peer-reviewed, and Gumshoe sells tracking.

How to measure properly

1. Prompt set: 30 to 100 buyer questions on category, comparisons, alternatives and problems, in your buyers’ language and in English.
2. Systems: at least ChatGPT, Perplexity, Gemini and AI Overviews, plus Copilot and Claude if your buyers use them.
3. Repeats: run each prompt several times, on different days.
4. Metric: share of answers that mention you (mention rate) versus competitors, plus factual accuracy and cited sources.
5. Cadence: monthly, with the same prompt set.

Add the platforms’ own data: the generative AI report in Search Console, AI Performance in Bing Webmaster Tools (in preview since February 2026, with citations in Copilot and Bing), visits from chatgpt.com or perplexity.ai in your analytics, and a “How did you hear about us?” field on your demo form.

Takeaway: Measure how often you’re mentioned across many prompts and runs. Positions in individual answers tell you almost nothing.

GEO is ongoing work, hard to fit in without an SEO hire. XEO is preparing exactly this work as a done-for-you GEO agency for software companies; the waitlist is open.

Frequently asked questions