GEO: how to get cited by ChatGPT, Perplexity and co.

What this guide covers

This guide is the practical side of GEO: what the research actually measured, how AI search tools fetch their sources, and what to do with your setup and content. If you’re a founder who first wants to understand why AI recommendations matter for a startup, start with our business guide to generative engine optimization.

What the GEO study really measured

The term GEO comes from a study by researchers at Princeton and other institutions (Aggarwal et al., first published November 2023, accepted at KDD 2024). It’s widely quoted but rarely precisely. The essentials:

Source: Aggarwal et al., GEO: Generative Engine Optimization, arXiv 2311.09735 and full text v3.

What it means for you: the study shows that concrete, sourced content gets used more often and more prominently in AI answers. It doesn’t show that a particular trick works the same way in ChatGPT, Gemini or Google’s AI Overviews today; it tested a GPT-3.5 simulation and a Perplexity sample. Results also varied by topic domain.

How AI search tools fetch sources

AI assistants answer many questions from training data. For current or specialist questions they search the web, read a few pages and summarise them. To be part of that, the search has to find your page and be allowed to read it.

• Training and search are separate: you can block GPTBot and still appear in ChatGPT search results via OAI-SearchBot.
• Check your firewall and CDN: Perplexity notes that bot protection can block its crawler. Your server logs show whether crawlers get through.
• For Google: if you’re indexed in regular search and allow snippets, you’re eligible for AI Overviews.

My rule of thumb: don’t block AI bots by reflex. If you want to be cited in AI answers, you have to be readable.

The four levers in practice

Writing extractable content

• Answer first: answer the page’s core question in two or three sentences at the top.
• Numbers with source and date: exactly what worked best in the study.
• Clear structure: question-style headings, short paragraphs, tables for comparisons.
• Your own data: what only you have can only be cited from you.
• No keyword stuffing: ineffective to harmful in the study.

My rule of thumb: entity clarity before content volume. If an AI isn’t sure who you are, it would rather name someone it is sure about.

llms.txt: what it is and what it isn’t (yet)

llms.txt is a proposal by Jeremy Howard from September 2024: a Markdown file at /llms.txt that gives AI agents a short overview of a site and links to easily readable content. llmstxt.org describes it as a proposal, not a formal standard.

• What’s documented: per llmstxt.org, OpenAI, Anthropic and Google publish llms.txt files for their developer docs, and many documentation platforms generate them automatically.
• What isn’t: that ChatGPT, Gemini or Perplexity read or prefer llms.txt when answering questions.
• Google says explicitly: AI Overviews don’t need AI text files.

Honest verdict: llms.txt is cheap, harmless and may help agents that read pages deliberately. Based on what’s known today, it isn’t a ranking lever. Invest in the four levers above first.

Measuring: are you being named?

There’s no Search Console for ChatGPT yet. So measuring means asking fixed questions regularly and recording who gets named.

Note: answers vary from run to run. One query isn’t a measurement; only the pattern across several runs counts.

To compare AI recommendations, there’s our AI duel (in German): Wen empfiehlt die KI?

Tools for ongoing measurement

Prices per vendor pages, excluding tax. The number of tracked questions drives the price; start small.

“Try” links are affiliate links (advertising): I earn a commission at no extra cost to you.

Checklist

• robots.txt allows OAI-SearchBot, PerplexityBot and Googlebot; training crawlers decided deliberately
• Bot protection and CDN let search crawlers through (check logs)
• Key pages indexed and snippets allowed
• Every core page starts with a clear answer
• Numbers with source and date, your own data where possible
• Name, description and contact details consistent everywhere, structured data with sameAs
• Mentions built on independent sites
• Monthly spot check with fixed questions, results recorded
• llms.txt optional, as an add-on, not a strategy

Before and after: making a paragraph extractable

What does “extractable” look like in practice? An example of the structure, with placeholders instead of invented numbers:

Important: the numbers must be real. An invented statistic isn’t just misleading; AI systems may pass it on, with your name attached.

A 30-day plan to start

After 30 days, repeat the week-1 spot check. Don’t expect quick jumps: AI search needs time to crawl new content and pick up mentions. The plan lays the foundation on which visibility builds over months.

My rule of thumb: when personal AI agents pre-sort what customers see at all, your marketing has to be readable for the agent. Clarity beats creativity, at least on the pages you want an AI to cite.

When AI says the wrong thing about you

Sometimes AI assistants describe a company wrongly: an old name, the wrong offer, a competitor under your name. There’s usually no direct way to correct it. But you can work on the sources the answer comes from.

1. Find the source: ask the AI with web search what it’s relying on; check the pages it names.
2. Fix your own pages: bring your about page, imprint and structured data up to the same current state.
3. Align profiles: directories, LinkedIn, review sites, Wikidata, as far as you’re allowed to edit them.
4. Other people’s pages: politely ask the owner for a correction, with evidence.
5. Measure again: ask the same question again after a few weeks.

Many AI assistants also offer feedback on individual answers. It’s no guaranteed route to a correction, but it doesn’t hurt.

12 prompts for GEO

My audience is [audience], my offer [offer]. List 20 questions these people would ask an AI assistant before looking for a provider like me. Without my brand name, in their words.

Answer this question as you would for a customer: [question]. Name specific providers and the sources you rely on. (With web search; record the date.)

For the question [question], AI names [competitor] instead of me. Here’s their website copy and mine: [texts]. Which differences in clarity, evidence, mentions and structure might explain that?

For this page [text], write a two-to-three-sentence opening that answers the core question [question] directly, with a sourced number if available. No marketing language.

Rewrite this section so it’s easy to cite: clear claim, number with source and date, one sentence that can stand alone: [section]. Don’t invent numbers; mark gaps as [source missing].

From these customer questions [list], write an FAQ with short, unambiguous answers from my content [texts], plus the FAQPage JSON-LD.

Here are my company descriptions on the website, LinkedIn, directories and imprint: [texts]. Where do they contradict each other (name, founding year, offer, location)? Suggest one consistent description.

Write Organization JSON-LD for [company] with name, URL, logo, description, founding date, address and sameAs links to: [profiles]. Only details I give you.

Which kinds of independent sites do AI assistants typically cite for [topic] in [country]? Suggest 15 specific types (trade media, comparison sites, associations) and how I can honestly appear there.

Here’s my robots.txt: [content]. Which AI crawlers may access which areas? Am I allowing the search crawlers of OpenAI, Perplexity and Google? Suggest a version that allows search and deliberately handles training.

Write an llms.txt for my site in the llmstxt.org format: H1 with the name, a short summary, sections with the key pages: [page list with descriptions].

Here are my spot checks from two months: [table with question, date, vendors named, sources]. What changed, where am I newly named or no longer named, and which sources keep coming up?

The most common mistakes

1. Blocking AI crawlers wholesale. Then AI search can’t read you.
2. Betting on a trick. llms.txt or special markup don’t replace clear, sourced content.
3. Treating one query as a measurement. AI answers vary; only repeated spot checks count.
4. Overstretching studies. The GEO study shows trends in a simulation, not a guarantee for ChatGPT or Google.
5. Inconsistent company details. Different names and descriptions confuse AI systems too.
6. Keyword stuffing. Ineffective or harmful in the study.