There are plenty of interesting numbers in this analysis. This is the only one where the relationship is so stark you look twice.
We sorted every answer by whether the brand's own domain was among the cited sources or not, then checked how often the brand got mentioned in each group.
That's not a difference, that's a switch. And it shows up the same way in two independent surveys with the same model: once 92.9 versus 0.0 percent, once 87.7 versus 1.6 percent.
What came before is even more remarkable. In the first survey, the AI didn't pull from the brand's own domain in a single one of 1,134 citations. Not once. The brand only entered the conversation through retailer pages, review sites and advice portals.
Three weeks later, the same domain was the single most-cited source overall, with 1,237 of 5,387 citations, ahead of Reddit at 13.7 percent.
We can't say where that jump came from. A model switch happened between the two surveys, and measurably nothing had changed on the website itself. The most plausible explanation is that the newer model accesses manufacturer sources differently. That's not something this data can prove, it would take a repeat measurement within the same model.
Not necessarily, and that's an important caveat. It could just as easily run the other way: in answers where the brand is already the topic anyway, it's natural for the AI to also pull from its website. Cause and effect can't be cleanly separated here.
What can be said: when the own domain is missing, the brand almost never gets mentioned. In practice, that means a brand whose website AI systems never treat as a source depends entirely on what third parties write about it. Whether that's cause or symptom, it's a bad starting position either way.
From working on client sites, we know four recurring patterns. They're not a result of this measurement, but experience from analysis practice.
The page doesn't answer a question. Product pages list features. What's citable is a passage that conclusively answers a question.
The content only loads via JavaScript afterward. Many crawlers then see an empty page. What isn't in the server-side source simply doesn't exist for them.
The information is stuck in a PDF. Spec sheets and warranty terms as PDFs are considerably harder for models to use than a normal HTML page.
There's no claim you could actually cite. "Leading quality" can't be cited, because it doesn't claim anything specific. "Cuts branches up to 45 millimeters in diameter" can.
Ask ten questions typical for your category without your brand name, and note the cited sources for each answer. If your domain shows up in none of them, you have the core problem this article is about.
Load the page with JavaScript disabled, or look at the delivered source code. What's missing there, many crawlers never see. This is one of the most common and least noticeable mistakes.
A heading phrased as a question, the answer right underneath in two to four sentences. That structure extracts cleanly as a passage without losing context.
Dimensions, standards, timeframes, price ranges, test results with a source and date. Anything carrying a number is citable; anything that's just a judgment isn't.
At least the important ones: spec sheets, warranty terms, manuals. The PDF can stay as a download, but the information also belongs on a normal page.
Schema.org as JSON-LD for products, FAQ sections and the company itself. It doesn't replace good content, but it makes that content unambiguously attributable.
A look at robots.txt and your firewall rules. Some companies block AI bots without marketing even knowing, then wonder why their visibility is missing.
No, and we'd advise against relying on it. A study by Ahrefs from summer 2026 found that 97 percent of these files receive zero requests from AI crawlers, and Google has confirmed it has no ranking effect. As technical housekeeping it's harmless, as a visibility measure it's not an argument.
Well-designed and well-citable are two different things. The most common causes: content loads in afterward via JavaScript, the actual information sits in PDFs or videos with no transcript, or the copy does a lot of judging and states little that's concrete. The source-code test is usually the fastest way to diagnose it.
That's a business tradeoff, not a purely technical question. Blocking protects content from being used, and simultaneously gives up visibility in exactly the answers this is all about. What matters is that the decision gets made deliberately, not as a side effect of a firewall rule.
We don't have measurement data of our own on that. It depends on whether the given system uses live web search or answers from training data. With web search it can move fast; with pure training data, it waits for the next training cycle, which can mean months.
Share of answers with a brand mention, grouped by whether the brand's own domain was among that answer's cited sources. Survey July 26, 2026: 92.9 versus 0.0 percent. Survey August 3, 2026: 87.7 versus 1.6 percent. Both with gpt-5.6-terra. International parent-company domains were counted separately and not treated as the German brand domain.
A note on this article: research, analysis and writing were produced with the support of AI systems (the Ex Tenebris agent team) and reviewed editorially by Felix Zeh. Every figure cited comes from a real LUX/GEO analysis; the analyzed brand is anonymized to protect the client relationship.
The methodology behind this article is the same one we use for every assessment we run, tailored to your category, your competitors and your need states.