AI visibility is won per topic, not per domain

For this article I pulled the German search volumes via DataForSEO and stumbled on a number that has little to do with the actual topic and still explains a lot about where the industry stands. In Germany, “LLMO” has fallen from 720 to 210 monthly searches over twelve months. “KI-Sichtbarkeit” (AI visibility) rose from 70 to 260 in the same period. So the industry has not even agreed on a term yet. It is already busy sorting out who shows up in AI answers and who does not.

The common narrative sounds familiar. Build authority, then visibility in ChatGPT follows on its own. Sounds plausible. At its core it is the SEO logic of the past fifteen years, relabelled once.

Only it does not fit the data. Semrush, together with Kevin Indig, tracked 1,094 topic categories in ChatGPT for half a year and checked one simple question: do the usual authority metrics predict which brand dominates a topic? They barely do. The Authority Score was ahead for the topic owner in 52.5% of the pairwise comparisons. You might as well have flipped a coin.

What the study shows instead is far more useful for editorial planning than any tool recommendation. AI visibility is not a property of your domain. It emerges per topic. And the majority of topics are still free.

If you are still sorting out the terms around AIO, GEO, AEO and LLMO, read that first. This is about the question behind them: what does AI visibility actually predict, and what does it not?

According to Semrush the study covers 1,094 US categories and tracked them monthly in ChatGPT (OpenAI) from January to June 2026. The data base spans, per the authors, more than 220,000 domains and more than 50,000 brands.

A boundary that carries the rest of this article: what was measured is not general AI visibility across all platforms, but ownership inside ChatGPT answers. For Gemini, Perplexity or AI Overviews the study says nothing.

What matters is the methodology behind it. Each category was queried through five representative prompts that together map the typical course of a topic search: the definition question, the comparison, the purchase intent, the problem solution and the recommendation question. Not one question. The full range.

A brand only counts as the owner of a category when three conditions come together:

• it has the highest share of mentions in the category,
• it appears in at least four of the five prompts,
• and its lead over the runner-up is at least five percentage points.

That is a hard definition and exactly why the numbers below turn out so sobering. Whoever appears only on the definition question and disappears on the comparison does not own the topic. Whoever is narrowly ahead does not either.

This logic is not new. It is still SEO at its core, just without ranking positions. What changes is the metric. Not position one for a keyword, but share of mentions across a whole bundle of questions.

The distribution is the actual finding of the study. Of 1,094 categories, according to Semrush,

• 15.2% have a clear owner,
• 31.2% an emerging owner,
• 53.7% none at all.

More than half of all topics studied are undecided after six months of observation, because no brand there manages to be named across four of five questions with a five-point lead.

The usual reflex at this point: nice, but the free topics are surely the uninteresting ones. That is exactly what the study checks, and the result comes out the other way around.

Split the categories at the median into a high-volume and a low-volume half, and 98% of the entire AI search volume falls on the upper half, and it is precisely there that only 11.3% of topics have an owner. In the low-volume half it is 19%.

So the in-demand topics are less occupied than the niches. That is counterintuitive if you think about competition the way it works in classic web search, where high volume almost always comes with high crowding. In a niche, a few consistent mentions are enough for dominance. In a broad topic, mentions spread across so many providers that no one clears the five-point threshold. Search Engine Land highlights this point in its coverage too.

In practice that means: competition for AI visibility in the relevant topics is not yet decided. Not because no one is working there, but because so far no one is working consistently enough across all five question types.

This is the part of the study that contradicts the common narrative. The authors set the owner against the runner-up for each category and counted in how many of these pairs the owner had the better value on the classic SEO metrics.

On organic traffic the topic owner is even marginally less often ahead than the runner-up. You cannot infer any inverse relationship from that, the distance to 50% is far too small. But one thing is certain: from your traffic you cannot infer whether you own a topic in ChatGPT.

The only value with substance is branded search at 55.7%, and that fits the mechanics of the models. When people actively search for your brand name, you also appear in the texts that models learn from and cite at runtime. So recognition helps. An authority score computed from your backlink profile does not help measurably.

A clean distinction is worth making here, because two things often end up in one pot. Topical authority through content clusters is a concept for classic web search: thematic depth, internal linking, hub-and-spoke structure. That still works, and Google’s documentation remains the authoritative reference for it. Topic ownership in ChatGPT is something else. Not a ranking position, but share of mentions across a bundle of questions. The two do not correlate automatically, and that is exactly what the study measures.

The expensive fallacy now would be to replace the Authority Score as your north star with another vanity metric. A classic. An expensive classic. The study does not say authority is worthless. It says authority does not predict topic ownership. That is a weaker claim. And a far more useful one.

This number gets little attention in the German-speaking market so far, yet it is the most consequential of the whole study in practice.

In an answer, ChatGPT does two different things. It cites sources, meaning links the answer draws on. And it names brands, meaning providers recommended or listed as the answer. The study captured both separately, and the relationship is, per the authors, slightly negative, with a correlation of −0.229.

Slightly negative does not mean one rules out the other. It means: whoever is cited often tends not to be named as a brand more often, rather slightly less. Comparison and review portals get cited because they deliver structured information about a whole market, while the ones getting named are the providers they write about.

For practice this forces an uncomfortable decision up front, because “be visible in AI answers” is not one goal. It is two:

• Be cited as a source. For that you need extractable, precise, current information on a matter. The path there runs through content that answers a question conclusively, rather than leading up to it. I took apart what that looks like in detail in becoming the cited source.
• Be named as a brand. For that you have to appear in the comparisons, lists and recommendations of others. That is closer to PR and market presence than to on-page optimisation.

A tool that hands you “AI visibility” as a single number mixes both. Want to know why you do not show up in a category? Then you need the separation, otherwise you optimise the wrong role for months.

The high stability is first of all good news for anyone who wants to invest. Topic ownership is not volatile. Whoever was ahead in May was still ahead in June with roughly 90% probability, and that is exactly what justifies long-term content work on one topic instead of scattered effort across twenty.

The flip side is more interesting. The study separated the stable from the flipped categories and compared the median lead: 2.9 percentage points for the stable ones, 1.3 for the ones where the lead changed hands.

Both are small numbers, and that is exactly the point. The difference between “it is mine” and “it can flip any month” is a handful of percentage points of mention share. A narrow first place is not ownership. It is a snapshot.

From this follows a concrete consequence for your target metric: the sensible one is not whether you lead the category, but how large your lead over the runner-up is. A lead of one percentage point is a reason to keep working, and not a reason to tick the topic off.

This also explains why ownership so rarely materialises in the 53.7% of unclaimed categories. Not because no one is present there, but because presence spreads across many providers and in the end no one builds the necessary lead.

“53.7% are free” does not grant permission to get broader. It implies the opposite.

1. Pick topics instead of collecting keywords. Ownership emerges per category, and the definition demands presence across four of five question types, so for a single topic you need several pieces of content that together cover a bundle of questions. With two or three topics that is realistic. With twelve it is not.

2. Use the five question types as a grid. Definition, comparison, purchase intent, problem solution, recommendation. Go through your existing content on your core topic and enter it into this grid. The typical finding: several articles stack up on the definition question while nothing exists on the recommendation question. Four pieces, one question type. By the ownership definition that counts as one.

3. Measure your own coverage, do not estimate it. I recently rebuilt my own search index for seo-kreativ.de. 173 crawled URLs went in, 159 indexed came out. Eight of them were 404s that were not in the sitemap at all, and the crawler only found them through internal links. Until then I had a different picture of what is on my own site. This gap between assumed and actual coverage is the normal case.

4. Extractability over volume. For the citation role, what counts is whether a model can cleanly lift the answer out of your text: a clear question-answer structure, concrete figures with a source, a date. Structured data for AI Overviews helps with machine parsing. But it does not replace an answer that is actually in the text.

5. Measure what is measurable. There is no Search Console for ChatGPT. The only free source with real numbers from my own operation is the AI performance data in Bing Webmaster Tools. It covers Copilot, not ChatGPT, and is therefore an approximation. That is still more honest than a score that blends citations and mentions into one number.

An article that holds a study up against an industry narrative has to hold that study to the same standard. Four limitations belong on the table.

The publisher sells the tool. Semrush publishes the study and at the same time sells the toolkit the data was collected with. That does not make the numbers wrong, and the methodology is disclosed. But it does mean the choice of what gets measured and what gets emphasised does not come about neutrally.

US categories, one model, six months. What was studied are US categories in ChatGPT, which is why the distributions do not transfer to the German-speaking market without further ado, since the topic landscape here is smaller. Other systems behave measurably differently too. For Gemini, Semrush puts the overlap between named brands and cited domains at up to 30% in a separate analysis. You cannot compute that directly against the 21% from this study, the two values are defined differently. As a pointer to platform differences it is enough.

The data contains no topic-level ranking values. Kevin Indig points out in his own write-up that the data set contains no visibility or ranking data at the topic level, so the authority metrics remain domain values, not topic values. It is therefore not ruled out that a topic-level authority measure would predict better. It just does not exist in this data set.

Non-prediction is not ineffectiveness. Indig puts it most sharply in the study himself: “Treat that as a hypothesis, not a finding. What the data actually shows is: traditional SEO metrics aren’t enough to explain who owns a topic.” The data shows that the classic metrics are not sufficient to explain ownership, but it does not show that these metrics are irrelevant. His detailed take is in the Growth Memo.

This limit is not fine print. Whoever turns it into “backlinks are dead” has turned a clean study into a bad headline.

The most useful insight from 1,094 categories is not that classic metrics fail. It is that ownership emerges at a level where most people do not plan at all. Not on the domain. Not on the single keyword. On the topic as a bundle of questions.

That is the cheaper piece of news of the two. A domain authority that keeps up with the big players is not something you build in a year, but covering five question types across two topics cleanly and currently is editorial work within reach.

Status: June 2026. All information without guarantee. Despite careful research, no warranty is given as to topicality or completeness. The study results shown are based on the figures published by Semrush, for whose accuracy no warranty is assumed. The search volumes cited are from own research via DataForSEO and are estimates. All brands and product names mentioned are the property of their respective owners. This article does not replace individual legal or professional advice.