--- title: "AI Search FAQ 2026" description: "Malte Landwehr answers 82 questions on AI search, GEO and AEO: how answer engines work, what gets cited, how to measure it, and who owns it." url: "https://www.maltelandwehr.de/ai-search/" lang: en --- # AI Search FAQ The questions companies ask most about AI search, answered by Malte Landwehr: how answer engines pick their sources, what Generative Engine Optimization (GEO) and Answer Engine Optimization (AEO) add to Search Engine Optimization (SEO), what gets cited, how to measure it, and who should own it. Published 2 October 2026, last updated 2 October 2026. ## Summary AI search is about decisions, not clicks. Generative Engine Optimization (GEO) is a new category. Its smallest unit is a brand recommended inside an AI answer, not a URL ranking for a keyword. Search Engine Optimization (SEO) stays the foundation. Answer engines ground their answers in a web search, and a page that does not rank in Google or Bing does not make it into that search. Ranking is not enough, though. Engines cite short, self-contained passages in clear, declarative language, and they look for consensus across many sources. That makes third-party sources half of the work: review sites, Reddit, YouTube, listicles and partners. Clicks are the wrong metric. AI search turns only one to five percent of searches into a click. Measure visibility, share of voice and sources with prompt tracking, and add log files, web analytics and self-reported attribution. The SEO team usually owns GEO. It cannot execute it alone. ## Why Malte Landwehr answers these questions [Malte Landwehr](https://www.maltelandwehr.de/) is the CPO and CMO of Peec AI, the leading solution for AI search visibility and, according to Business Insider, one of Europe's fastest-growing AI startups. Before Peec AI he worked in Search Engine Optimization (SEO) for over 20 years. From 2020 to 2025 he ran SEO at idealo, the largest price comparison website in the world, first as Head of SEO and then as VP of SEO. Before that he was VP Product at Searchmetrics. He studied Computer Science at WWU Münster, with information retrieval and web crawling among his subjects. He speaks about AI search, GEO and AEO at conferences such as OMR, DMEXCO, BrightonSEO and Masters of Search. The full history is on the [speaker page](https://www.maltelandwehr.de/speaker/). The answers below are drawn from his talks, podcasts and written interviews between November 2024 and September 2026, merged into one answer per question. Every answer below is his. Contents 1. [Why Malte Landwehr answers these questions](#about) 2. [GEO, AEO and the name of the category](#category) 3. [How answer engines work](#mechanics) 4. [Platforms and market share](#platforms) 5. [Content on your own site](#on-page) 6. [Off-page: sources, PR and the places you do not own](#off-page) 7. [Measurement and attribution](#measurement) 8. [AI-generated content](#ai-content) 9. [Teams, ownership and budget](#organization) 10. [Brand perception, facts and risk](#perception) 11. [E-commerce and agentic commerce](#commerce) 12. [Zero click, publishers and the open web](#open-web) ## GEO, AEO and the name of the category ### GEO, AEO, LLMO, AI SEO: what is the right name for it? GEO, short for Generative Engine Optimization, is the term Malte Landwehr uses. He started using it consistently once Andreessen Horowitz used it prominently in May 2025. They set the standard. He also uses AEO, Answer Engine Optimization. A few big US thought leaders use it, and Google was always trying to be an answer engine anyway. LLMO was too long and nobody says it any more. DAM will not catch on either. AI search is a name for search, not a name for SEO. There is local search and there is local SEO, and the same split applies here. He usually says "search and answer engines". Saying "I used ChatGPT" is technically wrong. You went to chatgpt.com and used a chat interface on top of one version of a model with some grounding and tooling around it. ### Is GEO really a new category, or just SEO with a new label? It is a new category, and the reason is the unit of optimization. In Search Engine Optimization (SEO) the smallest unit is a URL ranking for a keyword. In Generative Engine Optimization (GEO) the smallest unit is a brand being recommended inside an AI answer. A URL has a loading time, a title tag, content, a publication date, backlinks and historical click data. A brand has none of that. A brand has text it owns, text it influences, and text other people write about it. SEO was also a traffic channel. People landed on your site, saw your identity, read your copy, used your navigation, picked up a retargeting cookie. Whether that traffic converted was as much on the product and design teams as on SEO. None of that happens in a chat. Pretending it is not a new category has a real cost. It left the field open to people with no SEO background selling things that obviously cannot work long term. Those things do work short term, so companies start trusting them. ### Will GEO become more important than SEO? That depends entirely on where the line is drawn. If the industry decides in a year that optimizing for LLMs is also SEO, then no. If it decides this is Answer Engine Optimization and SEO is one part of it, then yes. Malte Landwehr tries to avoid the naming debate. What he does believe: optimizing for AI-based search and answer systems becomes extremely important for most companies, and what is called SEO today stays at least half of that work. He expects at least half of all searches to be answered with AI support within a few years. A lot of weight shifts from index-based systems to answer-based ones, but not all of it. ### Is it still too early to invest in GEO? Everyone is entitled to their opinion. But at the very least, find out how visible your brand is in LLMs today. If you are highly visible with good sentiment, carry on as you were. If you are invisible, that is already relevant. In B2B SaaS there are companies getting more leads from ChatGPT than from Google. Leads, not clicks. Far fewer clicks, but the prospect is pre-qualified inside the chat. So the conversion rate can be twenty times higher than SEO traffic. ### Is there a reliable method for GEO, or is everyone selling air? Malte Landwehr would never use the word "rank" for AI search. But there are countless scientific papers and real case studies showing you can get your brand mentioned more often and your site cited more often. People have been optimizing for AI Overviews for years. He does not know why anyone convinces themselves that the same approaches cannot work for AI Mode, Perplexity and ChatGPT. ### Does the move to GEO make the work easier? No. Things change faster and you have to adapt faster. There are tactics that are embarrassingly easy right now, and they will stop working. On balance everything got harder, purely because the rate of change went up. ## How answer engines work ### Where does an AI answer actually come from? Three places: the knowledge and biases baked into the foundation model, whatever the system remembers about the user, and grounding. Grounding is the web search the model runs before or while it answers. It uses its own index plus Google, Bing or Brave depending on the product, pulls content out of the documents it finds, and generates from that. Only one of those three is easy to influence, and that is grounding. That is where a mix of SEO and the right content strategy pays off, sometimes within minutes. ### Can you influence the foundation model itself? Practically, no. Once the model is trained, that version says what it says until somebody spends millions retraining it. If the training data said "Nike is good" a thousand times and "Adidas is bad" a thousand times, that version will prefer Nike forever. The only reliable lever is having positive text about you in dozens of books and thousands of websites. That is the territory of nation state actors, large lobby groups and PR associations, the people who can fund huge campaigns to make oil, sugar or tobacco look good. For everyone else it is irrelevant, and the testing cadence is absurdly long anyway. ### What are fanout queries and why should I care? When an LLM cannot answer from model knowledge, it grounds. It does not search for your prompt. It distills the prompt into several fanout queries and searches for those. Ask "what is the weather next week in the capitals of Germany and France" and it searches "weather Berlin" and "weather Paris". It knows the capitals. It does not know the weather. The interesting part is the terms the engine adds that were never in the prompt. In e-commerce, ChatGPT puts "reviews" into roughly ten percent of fanout queries. For a while it added "Reddit" to everything. Recently it started adding "official". Year numbers are extremely common. That is why putting the current year in a title genuinely works. Do not optimize for every individual fanout query. Look for the recurring concepts, pick the two or three most impactful, and write for those, either on your own site or on a publisher. ### Which search index is each answer engine using? Yahoo Scout and Copilot use Bing. ChatGPT uses Google, sometimes Bing, and obfuscates it heavily. Grok uses Google, Bing and Brave. Claude uses Brave. Grok additionally grounds on X, which makes it the easiest engine to influence if you have reach there. Grok is also the most aggressive at adding terms to fanout queries. If you read Google's fanout patents and implement them literally, you end up with something very close to what Grok does. It may be a decent proxy for what Google runs behind the scenes. ### Are AI Overviews technically the same as normal Google search? No. It is a stripped-down algorithm with fewer spam filters, feeding a stripped-down version of Gemini. That is why you see domains cited in AI Overviews that no longer rank in normal Google search at all. Every second Google can save matters. So it pushes a lot of candidate documents at the model quickly and filters less. Bartosz Goralewicz showed the same effect back in the SGE days: if JavaScript rendering takes too long, documents get considered far less often. Anything slow and complicated cannot happen when you want an answer in one or two seconds. ### Why does ChatGPT keep quoting my old pricing? LLMs look for consensus, not for your pricing page. If you only state your pricing on your own site, and five Reddit threads and two review profiles still carry the old number, the old number is the answer. So when something about you changes, publish it everywhere. G2, Yelp, your social profiles, the footer of your press releases, your help center, your product docs. Several assets telling the same story from different angles let the engine find agreement fast. ### Do backlinks still matter for AI search? Yes, but indirectly. Grounding triggers a search in the background, and backlinks are a ranking factor there. Once the document has been retrieved, the model is not looking at a link graph. Links get you into the candidate pool. They do not decide the answer. Domains with more referring domains and higher authority do get cited more often. Malte Landwehr would bet his reputation on that correlation. He cannot offer scientific proof of causation. In practice, some backlinks are usually needed at domain level to rank at all. Once a URL ranks for the fanout queries, content type, style and formatting matter far more. ## Platforms and market share ### Which answer engines should I actually track? AI Overviews, ChatGPT and Gemini first. Consider AI Mode. For B2B add Microsoft Copilot. Plenty of large enterprises forbid their staff from using anything else, and that is the engine people forget. If you sell an API and want developers to adopt it, Gemini and Claude matter more. Beyond that, follow your own data. Look at where competitors get traffic from, and look at the sources that multiple engines cite consistently rather than what one engine cited today. ### What happened to Perplexity? It stopped being relevant. People say "ChatGPT and Perplexity" out of habit, but Perplexity is not even a top five answer engine any more. ChatGPT, Gemini, Grok, AI Overviews and Claude are all bigger. Rufus will likely pass it too. Claude overtook it in many countries despite being a largely B2B product used heavily through APIs. Malte Landwehr's prediction has been consistent: Perplexity usage drops hard, Gemini usage climbs hard, ChatGPT keeps growing in absolute numbers. ### Are we underrating Google AI Overviews? AI Overviews are the most underrated LLM surface, probably because they do not look like a chat interface. In terms of reach they are easily twice the size of ChatGPT. More than two and a half billion people see them, whether they chose to or not. ### When does AI Mode become Google's default? Not soon, and the reason is monetization. Google will pull elements of AI Mode into normal search step by step instead. A privately held Google could risk a quarter of revenue to move fast. A publicly traded one cannot. The share price drops, and then either you dilute employees or they leave because their equity is worth less. Logan Kilpatrick tweeting that AI Mode is about to become the default does not change that. He is at DeepMind and has no say over Google Search, and he was promptly made to issue a statement he obviously did not write. ### Have we hit peak ChatGPT? Not in absolute user numbers. But Gemini keeps taking share, Perplexity keeps losing it, and in many countries Grok has already overtaken Perplexity, especially in Asia where the political correctness debate barely registers. India alone has hundreds of millions of Grok users. ### How big is AI search as a share of discovery? Published estimates run from five to twenty percent. The spread exists because nobody has the equivalent of search volume or Search Console for LLMs. Not every prompt is a search, and not every Google search is a real search either. Malte Landwehr thinks five percent is simply wrong. Ignoring AI Overviews, AI search is already at ten to twenty percent. By 2027 a large share of searching will be done by agents rather than people. A single prompt can trigger a hundred searches. Thinking in commercial journeys rather than queries, he would expect ninety percent of them to be partly AI. ### Will there be advertising inside LLMs? Something will come. The labs have hired a lot of advertising executives from Facebook. But Malte Landwehr cannot see classic banner or AdSense-style advertising working. A chat pre-qualifies the user so heavily that Google can charge ten times as much per click in AI Mode, or move to revenue share. The format that could work is the e-commerce slider. Click a product in ChatGPT and you see up to three shops. OpenAI taking a commission on those does not make the experience worse. Advertising that changes which product gets recommended would drive users to another engine. ## Content on your own site ### If I can only change one thing on a page, what should it be? Add a summary. Two or three sentences at the top of a long article. The same for a very long paragraph, a table, or a chart. Across dozens of websites, multiple languages and every engine Malte Landwehr has looked at, these short summaries are the single thing cited most often. The trick with tables and charts is to write the insight, not the description. Not "this table shows population growth in German state capitals" but "based on this data from the Federal Statistical Office, Berlin is the fastest growing state capital and Hanover the slowest". The first is not worth quoting. The second is. ### What does a citable paragraph actually look like? Around fifty words. Self-contained, so it still means something when lifted out of context. High entity density, with things named by their real names. Declarative and authoritative language, no "maybe" and no "probably". Cite your own sources inside it. Test it by pulling it out. "For the above mentioned reasons, the second solution is the best for this kind of problem" carries zero information alone. "Based on reviews in three print magazines and ten thousand Reddit comments, hiking boots from Jack Wolfskin are the best option for a beginner to intermediate hiker who wants one comfortable, reliable pair at an affordable but not cheap price" survives the extraction. Make especially sure you have one of these on your homepage. ### Should I chunk my content? Chunking has become a dirty word, and Malte Landwehr keeps his distance from over-engineering. You do not need to turn your text into bullet points the way some people do. What you need is that somewhere, usually high on the page, there is a self-contained passage in clear declarative language. LLMs do not take your whole document when they generate an answer. They take individual pieces, maybe twenty words, maybe two hundred. Keep it in one document rather than splitting it across pages. The first hurdle is still ranking in Bing or Google so the document enters the retrieval set. Then structure it well inside. ### Does the question-and-answer format help? A lot. Either make your H2 and H3 headings questions, or put a FAQ block with Q&A schema at the bottom. Malte Landwehr does not believe there is a magic dial where the markup moves you three positions. What happens is that the format forces you to write the exact condensed answer that fits as a citation. Correlation data also shows content performing well in ChatGPT tends to have a correct H1 to H3 structure, lists, at least three schema types, exactly one H1, and FAQ schema. That is correlation only, so do not treat it as proof. ### My article ranks on Google but gets zero citations. What do I fix first? If it ranks, the first boxes are already ticked. Then check in this order. Are AI crawlers blocked? Test it by putting the URL in a chat and asking what the third paragraph is about. Is there at least one paragraph an engine can easily cite, one to three sentences, high entity density, clear and declarative? If not, either rewrite the article or add a summary on top and a question-and-answer block at the bottom. And make sure important content is not hidden behind JavaScript. ### Does freshness matter in AI search? Yes, more than in traditional search. Every engine Malte Landwehr has looked at has a recency bias compared with classic web search. Republishing and updating works, and so does a machine-readable last-updated date that actually changes. That said, he is strongly against updating for the sake of updating. Do a proper content audit: delete the obsolete, leave alone what still converts and is accurate, merge what overlaps, and genuinely update the rest. ### Should I show the publish date or the updated date? There are two positions. One says be transparent and show both. The other says Google will keep showing the old publication date in the snippet, so delete it. What Malte Landwehr would actually do is display both, but make the original date hard for a crawler to read, in JavaScript or with a line break removed by CSS, so the updated date is the one every crawler sees. The maximum short-term impact and the long-term trust position are unfortunately different here. ### Is cannibalization still something to avoid? Malte Landwehr was always firmly in the anti-cannibalization camp, mostly because he worked on sites with a million or more URLs where it becomes a real problem. Now it genuinely makes sense to have several pages on one topic so engines can find consensus. But keep the intent distinct at the title level. Ten best health insurance providers, providers that won awards, best provider per state, best for people earning above a certain amount. Overlapping, but distinguishable. He still does not want three pages with the same title and nearly the same content. ### Do self-promotional listicles on my own site work? It is the most embarrassing tip Malte Landwehr gives, and it works. Ask for the best CRM and the sources are HubSpot and Zoho. SuperOffice has an article on the best CRMs where SuperOffice wins. Nightwatch has one where the best SEO tool is Nightwatch and the second best is the Nightwatch Chrome plugin. Monday, ClickUp, Dropbox, Blaze, all of them. Across industries, professional services is the worst offender: twenty percent of cited listicles are self-referential, and that includes SEO agencies. An agency in Berlin invented a matcha powder, put it in three listicles, and LLMs now sometimes recommend it. It has no company website at all. The warning matters. Publish hundreds of these, especially AI-written, and you lose your Google rankings. Lose your rankings and you are not found during grounding either. ### Is llms.txt worth building? Malte Landwehr tells clients no. It is a text file like any other, so it can be indexed and it can show up in grounding. He has seen it cited, but not because of special treatment. It is cited because most websites are so bad that a plain text document with no markup and high information density looks attractive, the way print views and PDFs used to. No answer engine will ever build special handling for a file where webmasters declare what they want to rank for. Meta keywords already showed that. Where llms.txt is genuinely useful is pointing agents and vibe-coding tools at your API or MCP server and explaining how your site is structured. ### Should I serve Markdown to LLM crawlers? Not as a parallel .md copy of every page. That splits crawl budget and gives any human who lands there a dead end with nothing to click. What can make sense is serving Markdown from the same URL based on user agent. That is a form of cloaking, so Malte Landwehr would not do it for Google. For a ChatGPT crawler, why not. It helps a lot if your site is heavy on client-side rendering. Where it gets dangerous is what Time is reportedly doing, injecting ads that only the model sees. At some point a product manager at OpenAI or Anthropic says that is not acceptable. ### What makes content invisible to LLMs? - **Semantic clutter.** The recipe that opens with where the grandmother was born. Unfocused content is very bad here. Write about the thing. - **Embedding collision.** Rewrite a Wikipedia article word by word and it is unique as a string but identical as information. The engine will always pick Wikipedia. Ten years ago that trick ranked. Not any more. - **No trust signals.** You can debate E-E-A-T as a concept, but as a mental model it holds. Engines want reputable sources, so send every signal you can. - **Client-side JavaScript.** Google can handle it. Most LLM crawlers do not even try to render. - **Blocked crawlers.** Check robots.txt and check your CDN. Cloudflare in particular now ships defaults that block some of these crawlers. ### Do I need English content if I only sell in Germany? Yes. Ask a question in German from a German IP and you would expect German sources. But a meaningful share of fanout queries still come out in English, and English sources get pulled in. A German brand that only communicates in German has no chance of influencing those. So make sure some information about you exists in English. An interview, a content marketing story, a digital PR campaign. Whatever brand wins the US market often becomes very visible in other markets even where it has no presence. And be aware that engines have never heard of hreflang and happily mix language versions. That is a serious problem for car manufacturers whose South American sites describe a differently equipped version of the same model. ### One playbook for all engines, or one per engine? One playbook. You did not build separate pages and backlinks for Bing and DuckDuckGo, so do not do it for Gemini. If there is a big gap between two engines, check the answers first. Is it only you who is less visible, or everyone? Perplexity often refuses to recommend at all, which drags everyone down. That is why Malte Landwehr looks at share of voice alongside visibility. If the answers look comparable, check the sources. Gemini is probably using different sources where you are less present, and that has an easy fix. ### Is good SEO enough for AI search? It is the foundation and it is not enough. Malte Landwehr's favorite case is a specialist financial services firm in the US. Perfect SEO, dominating Google and Bing, and zero percent visibility in Perplexity, despite being the most used source. Their page said: "we are not the right solution for everyone, here are five competitors we trust." The engine found the most important document on the topic, took the passage with the most entities in declarative language, and recommended the competitors. They rewrote the heading and put themselves at number one. Overnight they became the most visible brand in their segment. Great Search Engine Optimization is the perfect foundation. It is also not the finish line. ## Off-page: sources, PR and the places you do not own ### What is the single fastest-working GEO tactic right now? Closing source gaps. Run a few prompts, look at the sources the engines cite for your topics, and find the ones where your competitors are named and you are not. Then get yourself named there. A polite email, a comment, money, whatever is appropriate. It works very, very well. ### How do I turn a list of sources into an action plan? Work it in order. Run your prompts several times. Look at the brands currently winning for inspiration. Then look at the sources at both URL and domain level. - **Existing URLs:** can you get added? Target the ones that already list several competitors, because asking to be added is reasonable there. An interview with a competitor's CEO is a lost cause. - **Domains:** can you create new content there? Digital PR, a press release, the commercial content team, an advertorial, an affiliate deal, or just joining the community. - **Fanout queries:** harvest the terms the engine adds on its own and build for the recurring ones. ### Should I spend on digital PR before anyone has proved it drives citations? For an established brand, spend the digital PR budget on content living on third-party sites that can get cited. For a brand new website, spend it on links to strengthen your own domain so you can rank and be cited in the first place. ### Is it the link or the mention that counts? The mention. For the answer the engine generates, a link simply does not matter. Of course the link raises your chances of ranking in web search, which raises your chances of being a source, so indirectly it helps. But if somebody is already handling SEO and you are adding what LLMs need, the link is not the point. Every guest article, interview and podcast transcript about you is a potential source. When ten documents are pulled in and five of them say you are number one, you get recommended as number one. ### Do paid advertorials still work? Yes, and that is uncomfortable. Malte Landwehr has been tracking a German insurance prompt set since October. Around two percent of all sources are paid advertorials, mostly Bild and Handelsblatt, with Wirtschaftswoche close behind. The title literally says "Anzeige" and the engines cite them anyway. No engine has done anything about it. The useful part for the advertiser is that the mentions column shows only one brand. It is a bought article. If you have too much money, this is the simplest way to buy LLM visibility. It works less well than a year ago, and at some point it will stop. ### What do I do when Reddit, YouTube and social are the top sources? Match the source type to the team that can act on it. LinkedIn means getting your CEO publishing, and note that personal profiles get cited far more than company pages unless the prompt names the company. YouTube, TikTok and Instagram mean working with your influencer and creator people, and asking them to put a few terms in titles and descriptions. Reddit depends on how shady you are willing to be: most subreddits hate brand accounts, some accept them. Malte Landwehr knows marketers working with fake accounts. He does not recommend it. He has observed it working. If it is a review site, optimize the profile and get the reviews up. If it is corporate sites, look for partner directories or interviews. If it is e-commerce sites, look at their retail media offering and pick the formats that produce crawlable text, not banners. ### Why is Reddit so valuable to these systems? It is user generated content moderated by upvotes. Heavily moderated UGC is one of the most valuable things on the internet for both model training and ranking. There is really only Reddit, Quora, which is already spammed, and Wikipedia, which has a completely different register. On Reddit you find a person who survived a specific cancer writing about their experience with wigs, and hundreds of other people with real experience commenting. You do not find that on Wikipedia or in journalism. And unlike Facebook, a handful of loud voices with a fanbase do not dominate. ### Do I need a Wikipedia article? You do not even need a live one. Malte Landwehr has been watching a company whose article was deleted and locked in November. ChatGPT has cited it at roughly the same frequency ever since. He has seen the same with many other deleted articles. So in theory, if your brand can never qualify for a real article, you could create one, make sure ChatGPT sees it quickly, and it gets deleted after ten hours but keeps being cited for at least six months. Handle with care. You will probably never get a proper article afterwards. He mentions it mainly so people know to watch for it being done to them. ### What are the three cheapest things that actually move the needle? - **Consistency.** Describe yourself the same way everywhere. Not "SEO freelancer for lawyers" on your site, "SEO service provider in the legal field" on LinkedIn, and "visibility consultant for jurists" in your press releases. One wording, used in every profile you own. - **Fanout queries.** Look at what the engines are actually searching for and create content for it. Seriously, not as spam. - **Sources.** Look at the most frequent sources and find your way in. A YouTube channel you could collaborate with, a Reddit thread you could join, a local top-three list you could buy into. Consistency is the one that can change answers literally overnight for personal brands and small brands. It goes from "I do not know who Hans-Peter Müller is, maybe this third-league footballer" to "Hans-Peter Müller is a graphic designer for growth-stage startups." ## Measurement and attribution ### Why are clicks the wrong metric for AI search? Traditional web search converts thirty to forty percent of searches into a click. AI search converts one to five percent. Judging the channel on clicks makes you underestimate it severely. ChatGPT looks smaller than DuckDuckGo if you only count referrals. In one case Malte Landwehr has seen, one percent of clicks came from ChatGPT and twenty percent of new leads self-reported ChatGPT. And for most businesses the click was never the point. For BMW it matters far more that an answer says BMW makes a great car than that BMW gets cited for an emissions standard. The only businesses that genuinely need the click are publishers. ### What is dark search? Someone chats with an LLM until they have decided on a brand. There is no link, so they type the brand into Google or the address bar. That traffic gets attributed to brand or paid search. The decision happened in the chat and your analytics never saw it. It behaves like word of mouth. If one person tells another they are happy with their iPhone, Apple has no attribution for that recommendation, and does not need any. AI search needs to be tracked like brand marketing, not like performance marketing. It is actually harder than TV: with TV you at least know when the ad ran. ### What should I measure instead of clicks? Four things, and you need all four for a full picture. - **Prompt tracking.** Synthetic, but it shows the actual answer. Gives you a competitor benchmark, share of voice, the sources, and the ability to filter by topic, persona and funnel stage. It is the only one of the four that tells you who your competitors actually are. - **Log file analysis.** Real data, verified by reverse IP lookup. Tells you which URLs get retrieved. Does not tell you whether the content was used, whether you made it into the answer, or how you compare with anyone else. And caching plus engines that lie about their user agent make it incomplete. - **Web analytics.** Real data on the few clicks you do get. With CTR under five percent it captures a fraction, but a fraction is not nothing: one shop Malte Landwehr knows makes two million euros a month purely from ChatGPT clicks. - **Self-reported attribution.** Ask every new lead where they heard about you. Imperfect, and worth adding a channel you do not run at all so you can measure your own error rate. Still the best tool there is for dark search. ### Which GEO KPIs do I put on the slide? - **Visibility.** In what percentage of answers is your brand mentioned? The primary metric for most business models. - **Share of voice.** Of all brand mentions in those answers, what share is yours? Useful when an engine is generally reluctant to recommend. - **Citation share.** What percentage of citations comes from your domain? Matters most for publishers, affiliates, comparison sites and some shops. - **Source coverage.** Across the top ten, hundred or thousand source URLs, how many mention your brand? This has a huge influence on visibility. - **Sentiment and factual accuracy.** Especially on branded prompts, where visibility is a hundred percent anyway. For a board in B2B, Malte Landwehr would lead with self-reported attribution, use visibility and share of voice as the competitive benchmark, and treat citation rate as an internal KPI for the GEO team. ### How do I pick which prompts to track? Do not stress about it. LLMs are very good at understanding topic, context and intent. So unlike SEO you do not need the exact words your audience types. Most prompts are unique anyway. Real prompt volume is almost always one. The prompts with high volume are "hello", "thanks" and "again". Think in dimensions instead: topics, intents, maybe personas, maybe funnel stages, maybe locations. Write a few prompts for each combination. Encode the context in the prompt itself, as plainly as "I am a father of three, I commute every day, which car should I buy?" Then analyze topics and tags, never individual prompts. ### How do I find the words my audience actually uses? Customer support tickets, recorded sales calls, customer reviews, your internal site search, the chat widget on your website. Pull the vocabulary out with an LLM. If people searching for a car say "rasanter Flitzer" rather than "sports car" sixteen times across inquiries, put that in the prompt. Jargon changes results. Someone asking about cables with advanced dielectric insulation and low skin-effect distortion gets different brands than someone asking for a cable to connect their speakers. Someone asking about a heart attack gets different results than someone asking about myocardial infarction. ### Can I see LLM clicks in Google Analytics? Yes, with a custom segment. Malte Landwehr keeps a GitHub repository with around forty LLM referrer URLs. Turn that into a regex and drop it into GA4. The models will happily explain where to click, or you can screenshot GA4 and ask. Just remember the clicks are not the point for most companies. The recommendation is. ### Can I find out how often my brand is mentioned in total? Not really, and no engine reveals it. There are three approaches and each has a flaw. Tracking your own prompt set gives you a share, not an absolute number. Malte Landwehr finds that far more useful for daily optimization anyway. Prompt databases are dubious: one large SEO tool recently published a top-sources list with LinkedIn at number two and a completely unknown site in the top twenty, and those prompt sets contain things like "6 plus 3". Clickstream data tells you where clicks go after ChatGPT, which is not the same as a brand mention. ## AI-generated content ### What guardrails do I use if I publish AI content at scale? Build a quick detector and run your own content through it. Four measures, benchmarked against a set of human-written texts in the same format and topic: - **Perplexity.** How predictable the next word is given the previous ones. If your AI text scores much higher, adjust the prompt. - **Compression rate.** Borrowed from email spam filtering. Roughly, how many words you could remove without losing information. Generic "write 600 words about X" prompts with no data score badly. - **Jaccard similarity.** Catches text where it is the same sentence everywhere with a few words swapped. - **Cosine similarity.** Catches text where every word is different but the information is identical. Claude will write you the Python script. You do not need to understand the math to use it. Four measures expose most AI footprints. ### Where should the human sit in an AI content pipeline? At the brief, not at the edit. Write the brief, have a second prompt check whether the brief makes sense, brief each paragraph separately, write one paragraph per prompt, then run a fact checker, then check for empty paragraphs and repeated concepts, and go round again. That costs maybe five to seven and a half euros in tokens per piece, and it is usually really good content. Think of it as briefing a new author who is very smart, knows nothing about your business, and will bluff unless you are explicit. Moving four sentences around after the fact is rubber stamping, not editing. ### What is the biggest mistake marketers are making right now? Scaling AI-written content past any reasonable limit. AirOps, ClickUp, Gong, Sprout Social, Monday and Webflow all lost organic Google rankings over scaled AI content. Losing rankings costs you LLM visibility as a consequence. If you are a top player in your niche, use it sparingly and precisely. The risk is not worth the short-term reward. If you are the tenth largest CMS for dentists with six months of runway, go all in. It might save the company. ## Teams, ownership and budget ### Who owns GEO in the organization? Usually the SEO team, and that makes sense. They know what to do and they own most of the foundational work. But they cannot execute it alone. They will not run your influencer program, moderate your branded subreddit, publish YouTube videos, write press releases or maintain the help center, and all of those drive AI visibility. For software companies the documentation is often a major source of citations, and that usually sits with customer experience or engineering. At one Fortune 500 company Malte Landwehr knows, sixteen separate teams have KPIs supporting AI search, with one full-time person just telling those departments how often their content was cited. What smart organizations do is create a separate AEO or GEO team and move the best SEO people into it. The actual visibility owner, though, is the CMO. ### How much budget should move from SEO to AEO? An established brand with good SEO: roughly ninety percent SEO, ten percent AEO. A brand new company: ninety-nine percent AEO. For a brand new B2B SaaS, Malte Landwehr would plan every priority around the assumption that there are no more clicks coming from search, only recommendations. ### Does company maturity change the approach? Yes. In large mature companies the two challenges are guiding a marketing organization from a link-and-click world to a recommendation world, and establishing goals and responsibilities at all. Their advantage is existing SEO performance. They already rank, engines already retrieve and cite them, so the work is making content more citable and adjusting the content pipeline. Their website is their biggest asset. At the same time they need to manage expectations around stagnating or declining SEO traffic. ### What do the teams who win at GEO actually do differently? - They understand SEO is the foundation of GEO, and that SEO alone is not enough. - They have internalized that AI search leads to decisions, not clicks, and have pointed their strategy at that. - They treat GEO as a separate channel with its own budget, responsibility and measurement, and make it a goal for multiple teams. - They embrace the uncertainty. No prompt volume, no knowledge of exact prompts, broken attribution. None of that stops you from doubling your visibility. The ones who improved fastest did one of three things: created a lot of content, improved the structure and tone of existing content, or got their brand into relevant third-party sources. ### How do I build a strategy that does not expire every six months? Strategy should not change constantly. Tactics will need to change every six months if you want to stay on the cutting edge. Separate the two. For strategy, answer three questions. Do you have a team responsible for AI search visibility? Is that team actually enabled to influence it? Can you measure impact in a way stakeholders believe? Once all three are yes, look at the sources that multiple engines cite consistently, abstract them into types of site (UGC, editorial, competitors), and build from there. Then make sure it lines up with the overall marketing strategy. ### How has AI search changed the marketing job? When auditing or creating content, Malte Landwehr now asks things like "should this article, paragraph, video, table or chart have a summary?" and "can I add three related questions and answers?". The way he writes has changed. But he always weighs the human reader when answering those questions. He would never write purely for the models. The bigger change is orchestration. You are no longer writing tickets for developers and asking an author to work in a few keywords. You are coordinating a dozen stakeholders who each own a piece of your visibility. ### How long does the shift to AI search take a company? It depends on how mature your product, technology and marketing teams are and how well they work together. A brand that already invested heavily in SEO and embedded it in the product organization can shift resources very quickly. No new hires, no new roles, just a change of focus. The timeline is the harder question, because there is no end state. You can adapt to what currently works within a year, maybe six months in a small company. But in twelve months new technologies will exist. There is no timeline to finish AI readiness. There is only a timeline to catch up with the status quo and then reach a state where you can keep iterating. ## Brand perception, facts and risk ### Why track brand perception and not just visibility? Sentiment, brand perception and factual accuracy are dimensions that simply did not exist in rankings and clicks. On branded prompts, where your visibility is a hundred percent anyway, perception is the only interesting thing to look at. Sentiment in isolation means little. A cigarette brand will never show a hundred percent green. It gets interesting relative to competitors. When one competitor is portrayed much worse, go and look at which prompts and which sources are doing it. Adoption here is behind. GEO was driven out of SEO teams, and perception is not their bread and butter. It belongs with brand management and PR, and those teams started later. ### Why do I look fast when asked about my brand, but vanish when asked about fast brands? The two questions pull from completely different sources. Malte Landwehr asked LLMs hundreds of times to name ten attributes of various car brands. For Tesla, speed came back nearly every time. Then he asked hundreds of times for ten fast car brands. Tesla appeared in under ten percent of answers. Ask for attributes of a brand and the brand's own site plus profile-type documents get cited. Ask for brands with an attribute and you get third-party content, mostly listicles. The second question is the one that matters across most of the funnel. So check both directions, brand-first and attribute-first, and point your PR and content strategy at whichever gap you find. ### Which sources will answer engines consider credible going forward? Until around August, ChatGPT leaned heavily on UGC, Reddit and YouTube, with Wikipedia an absolute top source in the first half of the year. It has shifted much more toward brand-owned domains. That also means competitor domains and domains that merely sound like the brand. The last one hits brands whose name is also a normal word particularly hard. Riot, Discord, Slack. Logically the engines will find ways to identify authoritative domains, the way Google did. Maybe Reddit gets heavily moderated, maybe a new ID-verified network appears. But whatever they decide, people who used to work in SEO will find ways around it. Malte Landwehr has been that person for twenty years. Nothing the engines have today would stop him. It stays a cat and mouse game for a while. ## E-commerce and agentic commerce ### How do I optimize product pages for AI search? The foundations are the same as SEO: specific, non-overlapping category and product pages with as much relevant content as possible without distracting from the commercial intent. On a listing page, show twenty or thirty products instead of ten before pagination, each with a couple of attributes. That is the highest-quality content you can put on a category page, and it almost always lifts traffic. On product pages, more data in the spec sheet nearly always helps. Small tricks matter: Apple products are not gray or blue, they are Midnight, so write "blue" next to it. Aggregated user reviews are excellent. Customers use different words than your copywriters and care about things you never considered. Then take the terms ChatGPT reveals about the category, the words it adds to shopping fanout queries and the columns it uses in comparison tables, and work them into your pages and your feed. As a brand, put them into the descriptions you give Amazon, Tmall, Mercado Libre and eBay too. ### What does the perfect product detail page look like? Everything above, plus three summaries. Summarize the ten user reviews in two or three sentences. If the product is unusually cheap right now, say so explicitly: "this product is currently ten percent below its three-month average". If you have a long spec table, write one or two sentences on what makes this product distinctive. Those condensed passages are very often what gets quoted. They contain all the important information in one chunk. And remember that in e-commerce the citation is a side effect. Getting into the shopping boxes matters more. ### What is agentic commerce, really? A shopping experience where the buyer no longer does the research and purchase steps. They state a desire by voice or text, and an agent that can reach the web, shops, marketplaces and search engines picks the product type, picks a brand, picks a specific product, finds a shop and orders it. "I want to go camping next week, buy me everything I need." That is the fully automated version, and Malte Landwehr does not think it becomes the norm for consumer goods. Most people want to stay in the loop. B2B is where full automation lands first: ordering twenty thousand screws of a specific type from ten possible vendors where only price, availability and reliability matter. No human needs to be in that negotiation. Humans make everything more complicated. For consumers, the research steps go agentic first. "I am going hiking, what do I buy?" And once the decision is made and the buyer genuinely does not care whether it ships from Amazon, Walmart or Zalando, the agent can handle the buying. ### Which agentic commerce protocols do I need to care about? The acronyms overlap badly. MCP wraps an API or app so models and agents can use it, and it is widely adopted. It will recede somewhat now that agents handle plain APIs and MCP is one more hallucination layer. A2A hands a subtask from one agent to another. AP2 handles payment with pre-authorized limits. UCP from Google describes how commerce should run end to end and is the more open option. ACP from OpenAI and Stripe handles the full checkout inside a chat. WebMCP lets the browser know where to click on a shop. Realistically some of these will turn out to be irrelevant and some will be abstracted away, the way you build a website today without thinking about DNS, TCP, TLS or 802.11. Stripe is a plausible architect of that abstraction. If you are in e-commerce, look at UCP and ACP. Nobody knows which wins. ### What is agent experience, and is it the same interface as UX? It is exactly the discipline of UX, except you develop empathy for agents instead of humans. And it is very likely a different interface. An agent needs Markdown, API endpoints or an MCP. It does not need a styled headline it will flatten anyway, a cover image, mood photography, color coding or a slider. It can take a table or a list. What an agent probably values is precision. Not "these are the Mammut ski boots, size 42 to 43, delivery usually under seven days" but "these are the Mammut ski boots, manufactured 2025, sold in Switzerland under serial number 77935, shoe size 45.2 which runs slightly large, delivery in 18 hours". And whether you are blocking agents or giving them a real interface. ### What do I measure for agents? Split it in two. For AI search: visibility, the win rate of your products when several are compared, whether your shop is listed first when several shops appear, whether your site is used as a source, and how your brand and products are talked about. For the agentic transaction: the same things you would measure for humans, filtered for agents. Conversion rate of agent visits. Cart abandonment. Can they use a filter on a category page and then select a product? What tool calls are they making against your MCP? Are they repeating or rephrasing their question because the first answer was not good? Many agents do not load your analytics JavaScript and do not accept your cookie banner. So that means log file analysis again. ### Can a smaller brand compete with the established players in AI search? Established brands have an advantage, but it is smaller in AI search than in SEO, at least for now. Very few brands have a real GEO practice. So small and nimble brands have an outsized opportunity to influence the grounding process. A few thousand euros on third-party listicles and being cringe enough to publish "x vs y" and "best x" content gets you a long way. Visibility differences are massive and extremely specific, though. Kellogg's being visible for sugary breakfast in the US says nothing about healthy breakfast, nothing about Canada, and nothing about Gemini. Someone who loves audio gets a completely different set of headphone brands than a fifteen-year-old who wants a cheap gaming headset. ### What are the specific risks for a premium brand? Three. First, invisibility. If three brands come up for best winter jackets and you are not one of them, you never even enter discovery. Second, losing your USPs. If agents compare strictly on a spec table, premium brands are at a disadvantage. The jackets are the same size, similar material, often from the same factories in Turkey, Portugal, China or India. You need attributes an agent will actually register. A patented cooling technology that is really just a particular use of merino gets its own check mark, and then it is a differentiator. Third, the sale going to a marketplace at a worse margin, or to a counterfeit. The products may come from the same place, but the premium one gets quality assurance on things like dyes and the cheap one does not. ### Who ends up with the power: brand, marketplace or AI platform? The answer engines, and they will be even more extreme gatekeepers than search engines were. The current US administration is unlikely to regulate. The EU has stopped enforcing the DMA against US gatekeepers. Antitrust is hard to apply when market share shifts every quarter. On top of that, AI makes it cheap to build a large shopping graph of products, offers and reviews. The companies aggregating demand need marketplaces less, and less reliance means fewer incentives to send users anywhere. The biggest losers are undifferentiated merchants and marketplaces. Malte Landwehr expects factory-to-LLM cases where Chinese manufacturers sell directly over these protocols. ### Morgan Stanley says half of online shoppers will use agents by 2030. Is that realistic? The adoption curve part is easy. Every new technology of the last decades reached a hundred million or a billion users faster than the one before, AI set records, and new products keep breaking them. The shift to AI will be much quicker than the shift from offline retail to e-commerce. On the exact number, Malte Landwehr has no idea. What he is confident about is that some purchases are made for the experience of making them. Buy a Rolex without ever setting foot in a Rolex store and you are missing out, and you will very likely enjoy it less. The same is said to be true of certain handbags. ### What is holding agentic commerce back today? Most people do not know it is possible, and trust. Do you hand an agent your credit card? It introduces new risk, so it needs to be a hundred times better, and right now it is only slightly more convenient. The version that wins is a notification saying milk ran out, order milk, one tap, the right milk, the right quantity, delivered when you are home. That does not exist yet. There is also no advertising surface for this. Zalando and MediaMarkt are not telling anyone to buy through ChatGPT. They want you on their site. So nobody knows which shops support what, and there is no way to advertise an agentic shopping surface to agents. ## Zero click, publishers and the open web ### How bad is zero click now? Google sits around sixty-five percent zero click, up from sixty, probably driven by AI Overviews. Keywords with an overview lose roughly thirty-five to forty-five percent of traffic, depending on which study you believe. Malte Landwehr looked at seventeen studies with forty-four data points: exactly one found AI Overviews increase clicks, and only for branded keywords. Google's AI Mode is around ninety-five percent zero click. ChatGPT was around ninety-nine percent for search-like prompts, and has recently come down to roughly ninety-five as it started linking brand mentions. ### Is top-of-funnel content still worth producing? It depends entirely on why you were producing it. If the model was people landing on informational content and clicking an ad, stop. Publisher traffic is down about twenty-five percent in a year and far more for many. Nobody needs to click through for that any more. If you are Salesforce, top-of-funnel content about CRMs makes complete sense. You get cited as the expert and you can recommend your own solution inside it. Plenty of companies are producing more of this, especially lists, precisely because it shapes the answers. But if the only purpose was ad revenue, do not bother. There is too much competition and there are barely any clicks. Malte Landwehr is convinced Google eventually stops crawling content it could generate itself in a second with Gemini. There is no incentive for it to spend the money. ### What happens to publishers and affiliates? For publishers it gets brutal. Clicks from AI answers are eighty to ninety percent below classic results. "Reach for display ads" and "SEO traffic for affiliate commission" do not work in a world where the answer is complete. One publisher told Malte Landwehr it lost seventy percent of traffic on evergreen content. Business Insider shut down content commerce in the US. Affiliates have the same click problem and one real advantage: LLMs love exactly the formats affiliates publish. Listicles, comparisons, "the ten best X". So affiliates stop being a click-and-commission machine and become a citation surface. Brands will ask how many LLM answers your article appears in, not how many clicks you sent. That breaks CPC and CPO pricing. He expects pay-per-citation or flat fees. ### Will websites still matter in five years? Many sites will get less traffic, but the website stays valuable as the one place a brand links all its profiles from. If you only have profiles on platforms, anyone can create a profile in your name. The thing that makes yours the real one is the link from your own domain. For a restaurant, the Google Maps profile is already more important than the website. The site still matters for trust, for certain audiences, for retention, and for cutting out the intermediary. And for shops specifically, Malte Landwehr would work now on being usable by agents, including an MCP server. That is one of the few genuinely good uses for llms.txt: pointing at it. ### Did we pass peak SEO traffic? 2024 or 2025 was the peak year for free clicks from Google. Malte Landwehr is confident about that. Look at publishers, affiliates or Wikipedia and it is obviously true. The total pie of organic clicks is not growing, and a lot of what remains goes to YouTube, Reddit and a few other large players. That does not mean every site is losing. Some are doing very well. It means you should not plan revenue, budget or headcount on the assumption that organic traffic stays where it is. ### What is the one thing to remember about AI search? AI search is about decisions, not clicks.