SIXT wanted to know what it could do to get cited more often in AI search. So Dr. ElMalah ran a controlled experiment to test a new approach to content. The results?
320% increase in citation frequency across full-treatment domains
+5.8 visibility during the experiment
+1.6 share of voice against tracked competitors
About
SIXT is one of the world’s largest mobility providers, offering car rental, car sharing, ride-hailing, and other mobility services in more than 100 countries.
Travelers now use AI assistants to explore destinations and compare transport options. When they ask questions SIXT can answer, SIXT wants to be among the sources shaping the answer.
As Dr. ElMalah explained, "AI is being built into users’ and businesses’ day-to-day at an exponential rate. Monthly active users of AI tools and features are already in the billions, and that shift disrupted traditional search behavior. Search has been pivoting toward AI, and our travelers and users are no exception. When they research or plan a trip, more and more of that journey starts in an AI answer, where SIXT needs to be visible."
SIXT was already using Peec AI to measure its presence in AI search: which prompts it appeared for, which sources AI platforms cited, and how its visibility compared with competitors.
The challenge: you can see what AI cites, but not why
Seeing which sources get cited doesn’t reveal the signals that caused an AI model to select them.
Part of the reason is how AI platforms build their answers. AI platforms retrieve relevant information, then make choices about what ultimately appears and gets cited. The models and signals behind those choices aren’t public. And as Peec’s own research has shown, ranking highly doesn’t necessarily mean you’ll get the citation.
That left SIXT with a practical problem: figuring out what it could change when the reasons one source gets picked over another aren’t visible.
Rather than trying to solve that in the abstract, Dr. Hadeer ElMalah, senior SEO and GEO manager at SIXT, turned it into an experiment. The resulting research, presented at SEASON 2026 in Hamburg, tests whether a specific approach to content can increase citation frequency in live AI search.
The hypothesis: cover the topic, not just the query
The experiment started from the idea that being the closest match to a query might not be enough.
The approach draws on a 2025 paper by Raja and Vats on Semantic Compression. Imagine asking 10 people the same question and getting 10 nearly identical answers. You’d probably learn more if some of them brought different, relevant perspectives. Semantic Compression explores a similar idea in retrieval: instead of selecting only the closest matches, select a compact set drawn from different semantic clusters that covers more of the topic.
ElMalah builds on that idea with an approach she calls Semantic Multi-Cluster Coverage. Instead of creating content around one narrow prompt, the approach aims to cover different angles of the topic and its sub-topics and questions surrounding its underlying intent.
Take a hypothetical search for “best car rental options for a Berlin trip.” A narrowly focused page might only cover rental prices at Berlin airport. But a useful answer could also include pickup and drop-off logistics, insurance, available vehicle types, and comparisons with the other sources heavily cited for this prompt, which may include competitors.
The SIXT experiment tested whether content built this way would be cited more often in AI-generated answers.
The test: track the prompts across different buckets
Instead of guessing what might get cited, SIXT started by tracking prompts across topic buckets, then looked at what AI platforms were already citing for them. Dr. ElMalah used those sources to shape new content, then tested the approach across 24 prompts and 11 partner listings.
1. SIXT tracked prompts in buckets, not as one list
Dr. ElMalah organized 24 prompts into buckets designed to cover different angles of the same topic. The goal was to capture different topic angles and sub-topics and expand semantic clusters around specific customer needs, to improve LLM retrieval rates and post-retrieval citation frequency. (If you’re just getting started with prompt tracking, you can check out Peec’s guide on choosing the right prompts for LLM tracking.)
2. SIXT leaned on AI to produce relevant content at scale
Writing multi-cluster content by hand across dozens of prompts isn't realistic. SIXT built an AI writing agent using five inputs, including four from Peec AI.
the target prompts
the URLs most often cited for those prompts
fan-out queries, or related searches around each prompt, and recurring terms
The agent mapped each prompt’s intent to relevant SIXT services and fleet information, then generated content covering multiple semantic clusters around that intent.
The agent then verified topical coverage and benchmarked cosine similarity: does this cover the topic comprehensively, and is the content close in meaning to what AI platforms most often cite for these prompts? SIXT targeted 95% topical alignment with the top-cited URLs and a cosine similarity of at least 0.85 with the prompts' top-cited URLs, a measure of how closely the content matches those URLs in meaning.
3. SIXT published content on external partner sites
SIXT applied the approach to 11 listings across external partner sites, without changing anything on its own website.
SIXT chose established direct travel-niche domains in the same geographic market.The treatment sites together received more than 740,000 organic clicks per month.
4. SIXT measured changes against a control
Four domains were included in the experiment: two received the full treatment, one received a partial treatment, and one was left unchanged as a control.
SIXT compared a two-week baseline with the two weeks after the changes went live. The partial-treatment domain was optimized under operational constraints, and its content page changes went live late, so it was measured for 12 days rather than 14.
The main metric was citation frequency: URL retrievals multiplied by citation rate. Being retrieved means a source made it into the information an AI platform considered; being cited means it appeared as a source in the answer. The metric captures both.
The unchanged domain provided a reference point for the movement on the treated listings.
The result: the treated listings outperformed the control
Increased citations across the treated listings: Citation frequency on the two full-treatment domains increased 320% (~308% once adjusted for normal background movement), while the partial-treatment domain increased 200% (~188% adjusted). The control domain rose 12.2%, in line with normal AI search fluctuation.
Gains in visibility and share of voice: Over the same period, SIXT posted the biggest increase in both Visibility (+5.8) and Share of Voice (+1.6) among tracked brands in Peec AI.
A framework for future tests: Dr. ElMalah built a structured way to find content opportunities through prompt-bucket tracking and produce content at scale with her writing agent. Replicating the approach in other locations is one of the paper’s stated next steps.
Dr. ElMalah describes the findings as preliminary, based on one brand and four domains. One full-treatment domain also continued climbing after the experiment, from 28.5 during the test to 83.9 and then 102 over the next four weeks. The paper tracks only this one full-treatment domain beyond the test period, with no concurrent control.
Figure 1. Citation frequency progression for one full-treatment domain across baseline, test, and two post-test periods.
Key takeaways for marketers
Three things are worth taking from the SIXT experiment:
Cover the wider need, not just the closest query: SIXT’s experiment was built around the idea that relevance alone may not be enough. Instead of optimizing around one narrow query, Dr. ElMalah used its writing agent to cover multiple semantic clusters around the underlying intent. The results make that approach worth testing in other categories and markets.
Think beyond your website: SIXT made no changes to its own properties. The experiment ran entirely across listings on external partner sites. For brands whose customers already encounter them through publishers, marketplaces, directories, or other third-party sites, those surfaces belong in the GEO strategy too.
Use a control when you can: AI search is changing too quickly to assume that every increase came from your intervention. SIXT’s unchanged domain gave Dr. ElMalah a reference point: while citation frequency rose 200–320% on the treated domains, the control rose 12.2%, in line with normal AI search fluctuation. That makes the result much more informative than simply comparing citations before and after.
For Dr. ElMalah, the experiment also reinforced a broader lesson about building a sustainable GEO strategy:
“GEO is optimizing for a moving target. A trend you see on LinkedIn might work today, but that doesn’t mean it will keep working. GEO is about mapping your own AI-tracked data in your industry and markets, and creating a sustainable GEO strategy grounded in that data and in how large language models and retrieval-augmented generation work, a strategy that adapts as new patterns arise.
Peec gives us a comprehensive, granular, and historical view of our AI search performance, which enables us to spot these patterns and separate real ones from AI volatility. We then draw insights, make them actionable, and run new tests.”
Want to see how your brand appears across ChatGPT, AI Overviews, AI Mode, Perplexity, and Gemini? Start a free trial or request a demo of Peec AI.
Source: ElMalah, H. (2026). Semantic Multi-Cluster Coverage and LLM Citation Frequency: A Controlled Generative Engine Optimization Experiment. SEASON 2026, September 15–17, Hamburg, Germany. https://doi.org/10.48441/4427.3734
