SEO vs. GEO SEO alone is no longer enough: How your website remains visible in times of Generative Engine Optimisation (GEO)

Juri Maier

System integration and APIsArtificial intelligence is considered a milestone in digitalisation - but as computing power increases, so does the demand for resources. Large language models (LLMs for short) in particular, such as ChatGPT, Gemini, DeepSeek and others, are suspected of being power guzzlers. How high is the actual energy consumption? Where do the greatest burdens arise - and where can emissions be avoided? We have checked the facts.

Training of LLMs will reach physical limit

The development of large LLMs is literally a tour de force. Months of training on high-performance computers consume huge amounts of energy. A meta-study conducted by the Öko-Institut on behalf of Greenpeace is sounding the alarm: if the training effort continues to double every five months, there is a risk of an exponential increase in energy consumption.

However, this forecast is based on a trend projection up to 2030, while the International Energy Agency (IEA) has looked a little further into the future and analysed various scenarios. According to this, the curve could flatten out from 2028 and emissions from data centres could even fall from 2030 and in the medium scenario if efficiency gains, the expansion of renewable energies and new cooling technologies are used. Another reason for assuming that the hunger for energy will fall again is that AI training is subject to a limit, as training data is not an inexhaustible resource. According to calculations by Epoch.ai, the available data could be exhausted between 2027 and 2032 - the end of growth would therefore be foreseeable.

‘Please and thank you’ cost energy

Not only the training, but also the operation of AI applications costs energy. ChatGPT allegedly consumes 2.9 watt hours per enquiry - almost ten times as much as a Google search. However, this figure is no longer entirely new. Current models such as GPT-4o work much more efficiently: Epoch.ai, an AI-friendly think tank, measured only around 0.3 Wh per 1,000 tokens. The progress lies in more efficient hardware and optimised models. However, this can only apply to requests that LLMs make to themselves. As soon as the Internet is also searched, it is obvious that the energy expenditure must tend to be higher than that of a single search engine query - a classic search engine query, mind you, because Google now offers information generated by its AI in addition to search results without being asked.

By the way: If you ask precise questions, leave out ‘please and thank you’ and pay attention to your spelling, you will save quite a bit of energy. Above all, a well-positioned prompt with all the important information on relevant context and the expected format of the result avoids follow-up questions - which reduces consumption.

Local vs. API: Does local hosting pay off?

The idea sounds charming: host your own AI models locally, run them on green electricity - and thus work in a more climate-friendly way. In practice, this situation is - as assumed - a little more complex.

Example: The LLaMA-2-13B model on a German server powered by green electricity. A consumption of approx. 0.58 Wh per 1,000 tokens was measured - around twice as much as with GPT-4o. Theoretically climate-neutral thanks to green electricity? Not quite. Even green electricity from the German grid remains part of the general electricity mix with its CO₂ share.

Even more serious: according to Boris Ruf's carbon footprint model, locally operated AI - even with green electricity - causes up to six times more emissions than efficiently hosted models, for example at Mistral, operated in French or Swedish data centres with hydroelectric and nuclear power.

It therefore makes no sense to rely on local AI for environmental reasons alone. Nevertheless, we do so if it is necessary due to aspects such as data protection or provider independence, or if the providers' APIs do not allow embedding in our customers' processes. In principle, the use of efficiently hosted AI services via their APIs makes much more ecological sense.

Text, speech, video: The data type makes the subtle difference

Not every use is equally data-intensive. Text requests are comparatively frugal. Voice input can also be quite data-intensive, but only if the speech recognition built into the operating system is used. Voice output and image processing require significantly more computing power, and moving images top everything, even without AI. Watching a 10-minute explanatory video on YouTube generated without AI can therefore consume significantly more energy than a 10-minute text-based dialogue with an AI. Everything is relative.

Infrastructure: More than just electricity

Data centres don't just need electricity - they also need space, water for cooling and hardware. The manufacture of GPUs and servers in particular has an environmental impact due to mining, rare earths and energy-intensive production processes. What's more, the useful life is often short. Faster chips make old hardware inefficient - and turn it into electronic waste. That is a problem.

However, continuing to use outdated hardware is not always the best alternative. New chips are sometimes not only more powerful, but also more economical. If you use outdated technology for too long, you may even worsen the environmental balance. If you want to conserve resources, you should use new hardware selectively and consciously - it is more important to look for efficient data centres that require as little additional energy as possible for cooling, use space efficiently, ideally reuse waste heat and, above all, are operated with electricity from renewable sources.

Conclusion

AI has a direct impact on our natural resources and the climate. We would therefore like to see labels for data centres and transparency obligations for AI providers. For us, we have defined this as a conscious and responsible approach:

  •     We focus on word processing when using AI models. Not only for energy reasons, but also for aesthetic reasons.
  •     Where possible, we rely on open source and self-hosted models that we can operate in data centres that work with green energy, effective cooling and a long hardware lifecycle.
  •     Where self-hosted models are not sufficient, we favour providers that have the lowest possible carbon footprint in terms of both training and operation.

Unfortunately, in the absence of better options, we sometimes have to make compromises. When in doubt, we therefore consider nuclear power suppliers to be the lesser evil compared to electricity sourced from coal-fired power plants (even if this is sometimes ‘improved’ by certificates). We would also like our standard hosting partner to be more innovative when it comes to land consumption and waste heat. However, their servers, which are powered by 100% hydropower, have an average lifespan of eight years and require no electricity for active cooling on a record-breaking 358 days a year.