GEO Playbook
Get your brand into AI answers, step by step
How to set up an AI-assisted GEO project for your brand and keep improving its AI visibility.
Sascha KirsteinVersion 1.0Contents
Part A
The basics
How an AI answer comes together
ChatGPT, Gemini and the rest don't answer from training alone. For most questions, they search the web, often several times and often in English, and then decide which pages to fetch and read. Understanding this is the key: it shows you where to start improving your AI visibility.
Question
„Welche PR-Tools sind DSGVO-konform und aus der EU?“
Language model
Trained knowledge plus search results
Web searches
- EU PR media database tool GDPR compliant Germany press distributionEnglish
- PR software Europe GDPR media monitoring press distribution EU companyEnglish
- + 1 more search
Fetch and read pages
- presspage.pro/de/pr-distribution-software
- pressfile.de/en/advantages/gdpr
- + 4 more pages
Answer (excerpt, in German as asked)
Für PR in Deutschland/EU würde ich zwischen Mediendatenbank/Media Relations, Presseversand/Newsroom und Medienmonitoring unterscheiden. Geeignete EU-Anbieter sind zum Beispiel:
- PressFile · Deutschlandpressfile.de
- Presspage · Niederlande/EUpresspage.pro
- Mynewsdesk · Schweden/EUmynewsdesk.com
- …
Sources
- presspage.pro
- wiztrust.com
- cision.de
- pressfile.de
- queuecraft.eu
- mynewsdesk.com
Which search engine sits behind it depends on the provider:
| Search index | Used by |
|---|---|
| Gemini, Google AI Mode | |
| Bing | ChatGPT, Microsoft Copilot |
| Brave | Claude, Mistral |
| Its own | Perplexity |
So SEO is not the opposite of GEO. It is part of it, just across more than one index.
From our own measurement: For questions asked in German, ChatGPT ran 92% of its searches in English. Our English comparison page was cited ten times as often as the German one.
Mentioned vs. cited
Brands get mentioned when others write about them. Pages get cited when they answer the question.
Mentioned
The brand appears in the answer text. That is mostly down to others: trade media, comparison sites, reviews.
Cited
A page from the brand is a source for the answer. Here you do the talking: you set the narrative and shape the information that goes into the answer.
How to work with the playbook
A single answer is chance. Asking the same questions regularly is what makes the results statistically meaningful. That is what the project is for.
There is nothing to install. Everything lives in one project folder that you work in directly with your AI.
- Download the project folder and name it after the brand, for example
geo-acme. - Open it in your AI, for example Claude or Codex with folder access.
- Type "Start". The AI walks you through the steps in Part B, does the research itself and only asks you where your decision matters.
From then on, "Start" is how you open every session, ideally once a week. The AI checks how far the project has come, what is missing, what is due and whether there is a new version of the playbook. If there is, it installs it when you say so.
In ChatGPT, Copilot or Gemini without folder access, upload the files to a project and paste the contents of AGENTS.md as the instructions.
Decide first: by hand or with a tool?
This sets the scope of the whole project, so it comes first.
| By hand | With a tool | |
|---|---|---|
| Questions | 10 to 15 | from 20, grows with the project |
| Models | 2 to 3 | 4 or more |
| Measurement | spot checks, e.g. monthly | daily, with history |
| Good for | a first impression | trends, reporting, actions with proof of impact |
Both work with the same folder, and you can switch later at any time.
Try it without the folder
Copy this into an AI with web search:
Part B
The project
Check the starting point
The basics have to be right. Often something fundamental is missing here, such as a website that AI models can't read, one that blocks AI search, or a name that several companies share.
Is there an English version of the website? Most models search mainly in English, even for German questions. So an English version pays off even if your audience only speaks German.
The AI
- researches the website, the offering and the name
- checks whether there is an English version
- checks whether AI search can find and read the website
You
- say what already exists, so the work builds on it
Understand the brand and its audiences
The most important step. You measure what your audiences ask, not how the brand describes itself. Mostly that means customers, and depending on your goals also job seekers or an expert audience.
Add concrete goals with a time frame, for example: "By the end of 2027, we are mentioned in every third answer to buying questions in Austria."
The AI
- drafts a hypothesis of the ideal customer: who buys, with what problem, in which language
- helps you phrase goals so they can be measured
You
- check it with sales: who actually buys?
- say whether job seekers or an expert audience count too
Define brand recognition
This step often gets forgotten. AI models often write a brand differently, for example with the product name instead of the company name. Count only the official spelling and your numbers are wrong.
The AI
- collects spellings, product names and domains
- filters out ambiguous variants
You
- add names that are only known internally
From a project: 0% visibility after the first runs. The models did mention the brand, just under its product name.
Choose competitors
Pick the providers that actually matter in your target market. The list doesn't need to be perfect at the start: the first answers show which competitors the AI names, and you can add them at any time.
The AI
- researches providers in the target market and explains the selection
You
- say who you really lose deals to
From projects: Once, four US providers were dropped and regional ones added. Another time, the models named more than 30 tools nobody had thought of.
Develop the questions
10 to 20 questions are a good start. Phrase them the way your audience would ask an AI, including jargon if they use it themselves. What matters is a core set you measure for as long as possible, so trends show. More questions come naturally over time, and many projects grow to 100 to 200.
| Perspective | Example | What you measure |
|---|---|---|
| Customer | "Which tool can I use to measure whether our PR is working?" | Is the brand mentioned as a solution? |
| Talent | "I'm a data scientist moving to Innsbruck. Which companies there are good employers?" | Is it mentioned as an employer? |
| Thought leadership | "Why doesn't my brand show up in ChatGPT, and what can I do about it?" | Is it mentioned or cited as an expert? |
| Brand | "What does [brand] do, and who is it for?" | What is said, and is it correct? |
Analyse each perspective on its own. Watch out for questions tailored too closely to your own brand, for example because they use its self-description or its own buzzwords: almost only one answer fits, and visibility looks higher than it is.
Double-check
Measure twice, cut once. When the question set is ready, take a step back and check your assumptions one last time. Change a running set later and you break the time series. Three checks help:
- Would your audience ask it like this? Read every question once from the audience's point of view. Do they know every word?
- Is the set balanced? Are all the important topics covered, and does each topic have enough questions to be analysed on its own? Are there comparison and alternative questions about the usual competitors? Are there questions about your own brand, to see what is said about it?
- A second, independent look. A fresh chat rates the whole set, with strengths and weaknesses. The guide includes the prompt for this.
Set the scope
Once the question set is ready, it gets concrete: which models should you measure? Measure the models your audience actually uses. 20 questions on 4 models every day add up to about 2,400 answers a month.
Check the first run
Check the first run straight away and read the answers yourself. Setup mistakes show up right away, and you don't spend weeks measuring nothing. If your AI can control a browser, it asks the questions itself in ChatGPT, Gemini and the rest and collects the answers with their sources.
- Brand not recognised? A spelling is missing.
- Unknown providers in the answers? Add them as competitors.
- Questions with near-identical answers? Pause one of them.
- Something wrong about the brand? Note it down with the source.
- No providers in the answer? The question only returns general information instead of specific companies or products. Rephrase it, for example with "Which tool …".
Derive actions
Visibility and share of voice are results. The data behind them shows where to act: the sources the AI reads and the searches it runs. No fishing in the dark. The key question: which pages list your competitors but not your brand?
| Finding | Action |
|---|---|
| Competitors on portals and lists, your brand is not | Request a listing and supply the facts. Often the fastest win. |
| Trade media as a frequent source | Get on the media list, offer topics and data. |
| Nobody answers the question well | Write your own page: fair, backed by evidence, in English if needed. |
| Mentioned, but only backed by your own website | Build independent proof: press, reviews, expert articles. |
| False statements | Find the source and get it corrected. |
From our own measurement: When an answer cited one of the two comparison sites we are listed on, Gemini mentioned us 33% of the time. Otherwise it was 12%.
Every action gets an owner and a date. Measure again after four to six weeks.
Keep improving
"GEO is a real opportunity for smaller companies. Even without a big budget, you can improve your visibility in AI answers. At aclipp, it takes us less than five hours a week on average."
Sascha Kirstein, founder of aclipp
A GEO project is a cycle: find a gap, take action, measure the effect.
| When | What |
|---|---|
| Weekly | "Start" in the AI, kick off one action |
| Monthly | Progress against competitors, effect of the actions |
| Quarterly | Review: are the audiences, competitors, questions and goals still right? |
To keep numbers comparable over time, the folder has rules: delete nothing, only archive, and if you rephrase a question, create a new one. The AI knows these rules and sticks to them.
From our own measurement: Our citations rose from 7 to 1,368 between June and September. Part of that came only from measuring more models. Only the comparison with the market leader showed the real progress.
Citations in AI answers per month
- aclipp
- Market leader
- 734
Jun
- 4251
Jul
- 7241,345
Aug
- 1,3681,867
Sep
- 1,6071,224
Oct(to 18th)
The playbook grows with our projects. The AI tells you about new versions on its own and applies them to your folder when you ask.
Measure your AI visibility with aclipp
aclipp tracks where your brand and competitors appear in AI answers every day. Your AI uses the results to suggest what to do next.
For Claude and ChatGPT: set up the connector.