---
type: "BlogPosting"
title: "Case study Datura: how a Berlin nature risk consultancy becomes findable for AI assistants"
description: "Client case Datura: how we rebuilt the website of a Berlin nature risk consultancy so that Google and AI assistants can find and cite its expertise – 0 → 10 answer pages, 0 → 59 AI-readable pages. Go-live still pending."
resource: "https://www.contextstudios.ai/blog/case-study-datura-ai-visibility"
language: "en"
tags: ["Case study", "GEO", "AI visibility", "Financial sector", "AI consulting Berlin", "llms.txt"]
generated:
  by: "process:contextstudios-md/1"
  at: "2026-11-13T18:52:12.367Z"
status: "stable"
---

# Case study Datura: how a Berlin nature risk consultancy becomes findable for AI assistants

Published: 2026-11-03
Tags: Case study, GEO, AI visibility, Financial sector, AI consulting Berlin, llms.txt

![Case study Datura: how a Berlin nature risk consultancy becomes findable for AI assistants](https://wary-platypus-754.convex.cloud/api/storage/486665a9-8ef6-486b-80fe-e033f7511b11)

Datura, a Berlin consultancy for nature and climate risk in the financial sector, has worked with Context Studios to rebuild its website so that Google and AI assistants such as ChatGPT or Perplexity can find, understand and cite its expertise. Instead of 0, there are now 10 pages that each answer a specific client question directly, 59 pages that AI assistants can read as plain text, and a description for Google on 100% instead of 10% of pages.

The new website is finished and running on a test environment (staging); Datura's go-live is still pending. This case study therefore shows what was built and how we will measure its effect – no results anticipated.

## Short answer: what did the project change?

The starting question was simple: when a bank asks an AI assistant who provides nature risk data, the answer should include "Datura". A good-looking website is not enough for that. An AI assistant needs pages that answer a question in the first sentence, figures with a source, and a structure that machines can read without guesswork.

That is exactly what we built, in four steps:

1. **Listen:** reviewed the existing website and collected, together with Datura, the questions their clients ask.
2. **Build answers:** a dedicated answer page for each core question, plus a glossary and a methodology page.
3. **Check facts:** matched every figure against its original source before it was allowed onto the site.
4. **Hand over:** prepared everything so that AI assistants can read it and Datura can maintain it themselves.

| Metric (website structure) | Before (datura.io) | After (staging) |
|---|---|---|
| Pages that answer a client question directly | 0 | 10 |
| Pages AI assistants can read directly | 0 | 59 |
| Pages with a description for Google | 10% | 100% |

These figures describe the website itself, not its visibility. Whether Datura is actually named more often in AI answers is something we will measure after go-live.

## Starting point: expertise that hardly anyone found online

Datura GmbH is based in Berlin and advises banks, asset managers and pension funds on nature and climate risk: how strongly does a loan book depend on water, soils or forests? How do you assess deforestation risk in a portfolio? What do European supervisors expect? The expertise was there, backed by in-house research and experienced people. The website just didn't tell Google or ChatGPT about it.

![The previous datura.io homepage: landscape painting and the slogan "Getting a grip on your nature and climate risks"](https://www.contextstudios.ai/images/references/datura-before.avif "Before: the previous datura.io homepage – appealing to people, but without answers an AI could take over (screenshot, September 2026)")

Our audit of the old site showed a picture typical of specialist consultancies:

- **No page answered the questions banks ask.** Content was organised by services, not by questions.
- **9 out of 10 pages had no description for Google** (meta description). Search engines had to guess what each page was about.
- **No file for AI assistants:** there was no `llms.txt`, and no structured data under Schema.org describing the organisation, its services and questions in machine-readable form.

None of this is a criticism of the design. The site was made for people who already knew Datura. For people starting with a question in an AI assistant, it was invisible.

## The idea: answer pages for the questions banks ask AI assistants

Anyone researching nature risk at a bank today often asks an AI assistant first. These assistants don't look for keywords; they look for passages that answer a question clearly. Researchers from Princeton and other institutions described this field in 2023 as "Generative Engine Optimization" (GEO): shaping content so that generative search systems pick it up and name it as a source ([Aggarwal et al., GEO: Generative Engine Optimization](https://arxiv.org/abs/2311.09735)).

That led to our principle for Datura: **one page per question.** Each answer page carries the question as its heading, answers it in the first sentences and backs the answer with figures including source and year. An example of the kind of question at stake: "Who provides nature risk data for banks and investors?"

![The new Datura answer page "Who provides nature risk data for banks and investors?" with three sourced figures](https://www.contextstudios.ai/images/references/datura-hub.avif "After: an answer page on staging.datura.io – question as heading, answer in the first paragraph, three sourced figures on the right")

This page states, for example, that nearly 75% of euro-area corporate bank lending goes to firms highly dependent on at least one ecosystem service. The figure comes from a European Central Bank paper ([ECB Occasional Paper No. 380, 2025](https://www.ecb.europa.eu/pub/pdf/scpops/ecb.op380.en.pdf)) and appears on the page with its source and year. An AI assistant can take it over without first having to reconstruct the context.

## Implementation in four steps

### 1. Listen

We didn't start with a keyword tool but with a conversation. We reviewed the existing website technically and in terms of content, and collected with Datura the questions clients actually ask: about deforestation risk, water risk, EBA and ECB expectations, and reporting standards such as TNFD and CSRD. These questions became the structure of the new site.

One guardrail applied from the start: **we invent nothing.** Where content was missing, we left a clearly marked gap and asked Datura to supply it. Subject-matter statements come from Datura, not from us.

### 2. Build answers

Each core question got its own answer page with a fixed structure: question as heading, short answer, sourced key figures, technical terms linked to the glossary, a section of frequently asked questions and a list of sources. Two pages were added that all answers refer to:

- **A glossary** with definitions of the most important technical terms, each with a primary source and date. Every term has its own address so that an AI assistant can cite exactly that definition.
- **A methodology page** describing Datura's approach step by step.

Comparisons with competitors work without vendor names. That was Datura's wish and it fits the goal: the site should explain Datura, not others.

### 3. Check facts

A website meant to be cited by AI assistants must not contain wrong figures, because those are exactly the figures that would be passed on. That is why we matched every statistic against its original source before publication.

This turned up **two swapped sources and one wrong year**: references where the link led to a different study than stated, and a paper dated to the wrong year of publication. Everything was found and corrected before it went online. For a consultancy whose business rests on reliable data, this step is not a detail; it is part of its credibility.

### 4. Hand over

The last step was the technical layer that people don't see but AI assistants and search engines very much do:

- **`llms.txt`:** a file in the root directory that tells a language model in brief what Datura is and which pages it should read. The format is an open proposal ([llmstxt.org](https://llmstxt.org/)), not a standard, but it is quick to build.
- **`.md` twins:** every answer page, every glossary term and the methodology also exist as plain text in Markdown format, without navigation, scripts or styling. Append `.md` to the address and you get the content on its own.
- **Structured data under Schema.org:** the organisation, questions and answers ([FAQPage](https://schema.org/FAQPage)), technical terms ([DefinedTerm](https://schema.org/DefinedTerm)) and the page hierarchy are marked up in machine-readable form.
- **A machine-readable company profile** with positioning, address and team, so that search systems recognise Datura unambiguously as an organisation.
- **A clear rule for AI crawlers:** the test environment stays closed to all crawlers; as soon as the same site runs under the real address, search and AI crawlers are explicitly allowed. No code change is needed at go-live.

"Hand over" also means that Datura should be able to maintain the site themselves. The content management system has dedicated building blocks for short answers, key figures with a mandatory source field, and questions with answers. Quotes can only be published once they are explicitly approved. An editorial guide directly in the system explains the three rules: invent nothing, every figure with a source, quotes only after approval.

## Result: a website for people and for AI

The same content now works twice: as a readable page for people and as a clear answer that an AI assistant can take over and cite. In figures, measured on the finished site in the test environment:

- **0 → 10** pages that answer a client question directly
- **0 → 59** pages that AI assistants can read directly as text
- **10% → 100%** of pages with a description for Google

The client contact is Malte Hessenius, COO of Datura.

On the status, stated honestly: the new website is built and signed off on `staging.datura.io`. The old site is still running under `datura.io`. Datura decides when the switch happens. Until then, the new site is closed to search engines and therefore cannot have any effect yet. We do not promise rankings or mentions in AI answers. We measure them.

> “Working with Context Studios on our website was a very smooth process. They upgraded our SEO and AEO, and the handover was clean that we can now maintain everything ourselves. Beyond the project, they were a great sparring partner for our next AI steps.”
>
> — Malte Hessenius, COO, Datura

## How we measure

To be able to prove anything after go-live, you need a starting value. That is why a baseline measurement ran at the end of September 2026, before go-live:

- **8 questions** of the kind banks and investors ask about nature risk,
- put to **4 AI tools**,
- each question in **3 runs**, because AI answers vary from one request to the next,
- **96 queries** in total. For each one we record whether Datura is named or linked as a source.

We are deliberately not publishing the results of this measurement here. It is the reference point for the repeat measurements after go-live, which we will carry out over the coming weeks and months. Only the comparison says something.

The measurement plan answers three questions:

1. Do ChatGPT, Perplexity and Google's AI answers name Datura for the questions that now have an answer page?
2. Do the answer pages appear on Google for exactly those questions?
3. Do new enquiries from banks and investors come in this way?

A note for context: Google itself states that there are no additional requirements and no special optimisations needed for AI Overviews and AI Mode in Search; the familiar SEO fundamentals apply ([Google Search Central, "AI features and your website"](https://developers.google.com/search/docs/appearance/ai-features)). That is why a large part of the project was classic diligence: clear pages, clean descriptions, sourced content. The machine-readable layer comes on top, but it does not replace these fundamentals.

## What other consultancies can learn from this

Datura is a case from sustainable finance, but the pattern fits many specialist consultancies, whether in the financial sector, in law or in industry. Five points carry over:

1. **Start with your clients' questions, not with keywords.** The best list of questions is held by your own people in sales and advisory. A workshop with them replaces many hours of keyword research.
2. **Answer in the first sentence.** A page that only gets to the point after three paragraphs is rarely cited by an AI assistant. Question as heading, answer directly below.
3. **Every figure with a source and year.** Especially in regulated industries, a wrong figure is more costly than none. The fact check belongs before publication, not after.
4. **The technical layer is the smaller part.** `llms.txt`, Markdown versions and structured data are quick to build once the content is clean. Without clean content they help little.
5. **Measure before go-live.** Without a starting value you cannot say later what has changed. A baseline measurement with fixed questions, several tools and repeat runs costs little and makes results reliable.

If you want to check how your own website performs in AI answers, our [GEO and AEO optimisation](/services/geo-aeo-optimierung) service describes the approach. You will find the broader picture, from strategy to implementation, under [AI consulting](/ai-consulting); if you are based in the capital, under [AI consulting Berlin](/ai-consulting-berlin). The short version of this client case is also [on our homepage](/#proof).

## FAQ

### What does AI visibility or GEO mean?

GEO stands for Generative Engine Optimization: preparing content so that AI assistants such as ChatGPT, Perplexity or Google's AI answers find it, understand it and name it as a source. At its core are pages that answer a specific question immediately, sourced figures and a machine-readable structure. Classic search engine optimisation remains the foundation.

### Is the new Datura website live yet?

Not yet. The new site is built and signed off on `staging.datura.io`; the old site is still running under `datura.io`. Datura decides the timing of the switch. The test environment is closed to search engines; access for search and AI crawlers takes effect automatically as soon as the site runs under the real address.

### What results has the project delivered so far?

What has measurably changed is the website itself: 10 instead of 0 pages that answer a client question directly, 59 instead of 0 pages that AI assistants can read as text, and 100% instead of 10% of pages with a description for Google. Whether Datura is named more often in AI answers as a result will only become clear from the comparison with the baseline measurement after go-live.

### Does every website need an llms.txt?

No, it is not a must. `llms.txt` is an open proposal, not a binding standard, and Google does not list any special requirements for its AI features. But the file takes little effort to build and gives language models a clear overview of the most important pages. Good content matters more; the file is an addition.

### Can the approach be transferred to other consultancies, and what does it cost?

Yes. The pattern of collecting questions, building answer pages, checking facts and measuring before go-live works for any consultancy with expertise that needs explaining, especially in the financial sector. We don't quote flat project prices because scope and starting point vary widely: you receive a fixed price after scoping, with the offer within 48 hours. Get in touch via our [contact page](/contact).

## Sources

- Aggarwal, P. et al.: [GEO: Generative Engine Optimization](https://arxiv.org/abs/2311.09735), arXiv 2311.09735, 2023
- European Central Bank: [Occasional Paper Series No. 380, "Nature at risk"](https://www.ecb.europa.eu/pub/pdf/scpops/ecb.op380.en.pdf), 2025
- Google Search Central: [AI features and your website](https://developers.google.com/search/docs/appearance/ai-features), accessed 2 November 2026
- [The /llms.txt file](https://llmstxt.org/), proposal for an overview file for language models, accessed 2 November 2026
- Schema.org: [FAQPage](https://schema.org/FAQPage) and [DefinedTerm](https://schema.org/DefinedTerm)
- Project figures: our own review of datura.io and staging.datura.io, September 2026


## Related

- [AI Consulting](https://www.contextstudios.ai/ai-consulting.md)
- [AI Development](https://www.contextstudios.ai/ai-development.md)
- [GEO optimization: visibility in AI answers](https://www.contextstudios.ai/geo-optimization.md)
- [AI Agency](https://www.contextstudios.ai/ai-development-company.md)