What’s Actually Working in AI Search Right Now

Istiak: You went from leading SEO at one of Europe’s biggest price comparison sites to a 10-person AI startup. Most people saw that as a risky bet. Now Peec AI has raised $21M and you’ve grown ARR from $500K to $5M in six months. Looking back, was there a single moment that made you say “okay, I’m doing this”?

Malte:

My wife and I went on vacation to South Korea. Just for fun, I set myself the challenge not to use Google. It worked incredibly well to do everything with ChatGPT. Searching, fact checking, translations, and planning which places to visit.

When we arrived at the airport and I reflected on the vacation, I realized that how people find information and consume content is changing faster and more radically than I previously thought.

I landed in Germany on Friday, went to the Peec AI office on Saturday, signed the contract on Sunday, and quit my corporate job on Monday.

I had already been in contact with the Peec AI founders regarding an advisory role. By the way, the Peec AI office back then consisted of a few desks in the corner of another startup’s office.

Istiak: You’re looking at data across thousands of brands through Peec AI. When a brand suddenly starts showing up more in ChatGPT or Gemini answers, what did they typically do in the weeks or months before that happened? What’s the common pattern you see?

Malte:

Most brands are visible in AI search because they did good SEO and had positive brand mentions across the internet, either organically or as the result of well-executed PR campaigns. 12 months ago that used to be enough. Now it is becoming more competitive.

Istiak: Let’s get specific about content. If someone just published a 2,000-word article and it’s ranking well on Google but getting zero mentions from AI tools, what’s the first thing you’d tell them to change about that page?

Malte:

If it ranks well in Google, the first few items on my checklist are already checked. Next, I usually verify AI crawlers are not blocked. You can test this by putting a URL in a GenAI chat and asking, “What is the third paragraph about?”

Next I check the structure. Is there at least one paragraph that an LLM-based answer engine can easily cite? I care about length (1 to 3 sentences) and high entity density. The language should be clear, declarative, and authoritative.

If that does not exist, I recommend either rewriting the whole article or adding a summary or question-answer block. These usually end up getting cited the most.

Istiak: You’ve talked about GEO having both an on-page side and an off-page side. For a mid-size brand with limited resources, where should they spend 70% of their effort? Fixing what’s on their own website, or working on how others talk about them across the web?

Malte:

I would first dedicate 70% of my effort (potentially even 100%) to my own website and then gradually shift to focus on third-party sources (off-page).

Istiak: Does content need to be optimized differently for ChatGPT vs. Gemini vs. Perplexity? Or is there one playbook that works across all of them? For example, if I check my content’s visibility on ChatGPT and it’s great but Gemini ignores me, what’s likely going wrong?

Malte:

Probably yes. But just like with traditional web search engines, I recommend having one strategy. You did not create separate pages or backlinks for Bing and DuckDuckGo. So do not do it for Gemini either.

Largely, the same approach works across all LLM-based answer engines. If there is a huge discrepancy between ChatGPT and Gemini, I would first check the answers and then the sources.

Checking the answers: Is it just me who is less visible or my competitors as well? I see this often with ChatGPT and Perplexity. A brand is visible in 80% of ChatGPT answers but only 40% of Perplexity answers. If you dive deeper, you notice that the same is true for the competition. Because Perplexity only gave answers in the style of “I cannot recommend a best SEO agency. If you tell me in more detail what you are looking for, I might be able to recommend one.”

That is why in addition to Visibility (In what percentage of chats am I mentioned?) I also like to look at Share of Voice (Across all brand mentions in chats, what share goes to my own brand?).

If checking the answers does not reveal a general discrepancy between ChatGPT and Gemini, I look at the sources. Very likely, Gemini is using different sources from ChatGPT. And in these my brand is probably less popular. This has an easy fix: get mentioned in the sources favored by Gemini.

Another thing I check in the sources is, how often is my own website used as a source vs my competitors. If Gemini cites a competitor’s website more often than mine, I know I have an opportunity to optimize.

Istiak: Your team at Peec AI recently published data showing that brands using AI content generation tools often see visibility drops in both Google and LLMs. That’s a scary finding. Where’s the line between using AI as a writing assistant versus using it in a way that backfires?

Malte:

Let me start by saying that I am a huge fan of automatically created landing pages for SEO and AI content. I have personally overseen the creation of millions of landing pages and hundreds of thousands of AI-written content pieces.

But you always need to ask yourself two questions:

1. When someone wants to read AI content, why would they go to my website instead of ChatGPT or Gemini?
2. If Google could create the same content, why would they go through the effort of crawling, rendering, indexing, and ranking my content?

If you do not have good answers to these questions, you should reconsider creating such content in the first place.

There are many cases where good answers to both questions exist:

• You have a list of hotels with attributes (pool, distance to beach, etc.) and want to describe each in a single sentence.
• You have a website with weather data. The data refreshes every 60 minutes. On each city-level page you want to display the current weather.
• You have a product details page with user-written product reviews. You want to summarize these reviews.
• For a stock market index (like the S&P 500) you want to have a page that always reports what happened in the last 3 hours.

But if your idea is to publish 10,000 LLM-written product reviews or blog posts, that is not something anybody wants to read. So Google, OpenAI & Co will try not to crawl, rank, cite, or recommend it.

Istiak: You’ve said that writing about new topics (new books, new shows, recent events) is one of the clearest ways to get cited by AI because the LLM has no choice but to use RAG. How quickly does that window close? Is “being first” a real competitive advantage in GEO the way it used to be in SEO?

Malte:

Every SEO advantage is still a GEO advantage. Because SEO helps you to get your own documents into the pool of candidates for source selection.

So being first is an advantage. But less so in GEO than in SEO. A cleaner written, better structured, and more authoritative page ranking on position two has a better chance of being cited than an older and sloppy page ranking on position one.

Istiak: Your recent LinkedIn post was about choosing the right prompts for LLM tracking. For someone just starting to monitor their AI visibility, what’s the simplest way to figure out what prompts matter for their brand?

Malte:

First of all, don’t stress over it too much. LLMs are great at understanding the topic, context, and intent of a prompt. So, opposite to SEO, you do not need to know word-for-word which prompts your target audience is typing into ChatGPT & Co. And even if you did know, most prompts are unique anyways.

Think about the topics, intents, maybe personas, and maybe sales funnel stages you care about. These are your dimensions. Then create a few prompts for each dimension and combination of dimensions.

To understand the vocabulary your audience is using, look at your customer support tickets, your recorded sales meetings, or the internal search bar and chatbox on your own website. Take that as input to create prompts.

To create the actual prompts, using ChatGPT or Claude can be very helpful.

Istiak: You wrote about the end of click-based attribution, what you called “The Dark Chat Manifesto.” If a CMO can’t track clicks from AI search the way they tracked Google clicks, how should they measure whether their GEO efforts are actually working?

Malte:

The same way they have been tracking the impact of brand advertising. Treat prompt tracking like a panel! What I mean by that is you do not need perfect data. You do not need to track every single prompt. You just need a set of questions (your prompts) and a panel (the LLM-based answer engines). And then you ask your panel these questions. Over time, you look at the delta to see trends.

As long as the data is directionally correct, you can see everything you need. How are you performing vs the competition, how is your performance developing, and what are opportunities you should tackle.

Additionally, CMOs can track self-reported attribution. Simply ask each new user or lead where they heard about your brand.

While many performance marketers and SEOs struggle with these approaches, both are established techniques that have been working well for decades. Of course the data is less perfect than a report from PPC campaigns. But it works!

Istiak: If someone reading this interview could only do one thing this week to improve their content’s chances of being cited by AI, just one thing, what would you tell them?

Malte:

To get cited: look at your 5 content pieces that have the most traditional SEO traffic. Ask yourself if they could be improved by adding a summary on top or a few questions and answers on the bottom. Then add them where it makes sense.

To make sure LLMs understand your brand: describe yourself consistently on the internet. Do not call yourself a lawyer for SEOs on your website, a legal visibility consultant on X, and an SEO agency for juristical topics in your press releases. Pick one version. Use that everywhere. This often works very well for very small brands and personal brands.

Istiak: You hold both the CPO and CMO title at Peec AI. How has the CMO job changed now that your “audience” includes language models, not just humans? Are you literally writing content with LLMs as readers in mind?

Malte:

When auditing or creating content, I ask myself certain questions because of AI. Stuff like “Should I add a summary for this article / paragraph / video / table / chart?” or “Can I add three related questions and their answers?” But I always include the needs of human readers when answering these questions.

The way I write has changed. But I would never write just for LLMs.

Istiak: Last one. Who’s one person you think we should interview next on this topic? Someone doing really interesting work in AI visibility or content optimization that more people should know about?

A few people I always like to hear from are Julian Redlich (Permira), Ethan Smith (Graphite), Lily Ray (Amsive), Niklas Buschner (Radyant), and Norman Nielsen (idealo).