Semrush now tracks AI Overviews (the summaries Google generates above organic results) and increasingly, practitioners are using it as a way to measure something broader: how visible a brand or project is to AI systems. I use it regularly as part of the audit process, and it earns its place there. But over time I’ve noticed a pattern in how people interpret what the data is telling them, and the gap between what the tool measures and what actually needs to change is worth being explicit about.
This piece is about how I use Semrush’s AI-related features in practice, what the data is genuinely good for, and where the diagnostic stops and the real work begins.
What Semrush is actually tracking
The most relevant features for AI visibility work are the AI Overview tracking within Position Tracking, the Keyword Magic Tool’s ability to filter for keywords that trigger AI Overviews, and the broader visibility metrics that show you how often your content appears inside AI-generated responses.
What this tells you, in practical terms: whether your content is being cited in AI Overviews for given keywords, which queries are generating AI Overviews at all, and roughly how your share of AI citations compares to competitors on a keyword-by-keyword basis.
That’s genuinely useful data. If you’re running an audit and want to show a client exactly how absent they are from AI-generated responses in their category, Semrush gives you a clean, documentable way to do that. It anchors the conversation in something concrete rather than anecdotal.
The scale of what’s being tracked matters, too. Semrush’s own 2025 study analysed over 10 million keywords to map how AI Overviews behave across query types and industries. What it found wasn’t a stable, predictable feature but a volatile one: AI Overviews appeared on just 6.49% of queries in January 2025, peaked at nearly 25% by July, and settled back to around 16% by November. More significantly, the query types triggering AI Overviews shifted dramatically over the same period. In January, 91.3% of AI Overview queries were informational. By October, that share had dropped to 57.1%, as commercial and transactional queries began generating AI summaries at much higher rates. The implication is that AI-generated responses are increasingly appearing at points in the search journey where decisions get made, not just at the awareness stage.
Where the data stops being the answer
Here’s what the data doesn’t tell you: why a project is being cited, or ignored, or described inaccurately when it is cited.
AI Overviews pull from sources that the underlying model has determined to be clear, credible, and relevant to the query. Being absent from those citations can mean several things — your content doesn’t exist in the right format, the underlying explanation of your product is unclear or inconsistent, you don’t have enough external validation, or all three. Semrush shows you the gap. It doesn’t diagnose which of those problems is driving it.
The accuracy issue is also worth keeping in mind. Research from the Tow Center for Digital Journalism at Columbia University, published in early 2025, tested eight major AI search platforms and found that collectively they gave incorrect answers to more than 60% of queries when asked to cite specific sources. The problem wasn’t just absence, it was confident misrepresentation. Across ChatGPT’s incorrect responses in that study, hedging language appeared in fewer than one in nine cases. So the question for any brand isn’t only “am I being cited?” but “when I am cited, is what’s being said accurate?”, and that second question requires a different kind of investigation than Semrush can provide.
For Web3 projects specifically, this distinction matters more than in most categories. A DeFi protocol or L2 infrastructure project might have extensive documentation, active community channels, and decent domain authority, and still be invisible in AI Overviews, or described incorrectly when it does appear. The reason is almost always that the available information is technically fragmented: too much assumed context, inconsistency between how the product is described across different sources, or a core explanation that was written for insiders rather than for a reader (or model) encountering the project cold.
None of that shows up in the keyword-level data. You can see that you’re absent; you can’t see that you’re absent because no clear, accessible, consensus description of your product exists anywhere on the web.
How I use it in the audit process
I use Semrush early, as a mapping tool. The first thing it’s useful for is understanding which queries in a client’s category are generating AI Overviews at all, because not every query does, and the ones that do tend to be the queries where authoritative, structured content is most likely to influence AI representation. That shapes where the content architecture work should focus.
The second thing I use it for is competitor benchmarking. Looking at which projects in a client’s category are being cited, and then examining the content that’s actually being pulled, tells you a lot about what AI systems are treating as reliable. In most Web3 categories, the projects that get cited aren’t necessarily the market leaders, they’re the ones whose content is structured clearly enough that a model can extract and summarise it without distortion.
This matters because of where AI citations actually come from. Research published by Yext in October 2025, based on analysis of 6.8 million AI citations across ChatGPT, Gemini, and Perplexity, found that 86% of citations came from sources that brands already control, their own websites and listings. That finding cuts against the assumption that AI visibility is primarily about what third parties say about you. It suggests that the quality and structure of your own content is doing the heavy lifting, which is exactly where the structural work is concentrated.
That’s the diagnostic that matters. Once you’ve identified which competitors are getting cited and looked at why, you have a concrete picture of what the content environment needs to look like, not just a visibility score to improve.
The limit of any tool in this category
The broader point is that AI visibility is not a metric problem. Improving the number of AI Overview citations is a downstream outcome of doing structural work on how a project is explained, what sources describe it consistently, and whether the content architecture supports reliable extraction. If you focus on the metric before that foundation is in place, you’re optimising a symptom.
Semrush is a strong tool for making the symptom visible and for tracking whether the structural work is having an effect over time. Where it leaves a gap is in telling you what structural work to do, and in a category as fragmented and terminology-dense as Web3, that gap is usually where most of the actual problem lives.
What I look for in Semrush, when it’s working well as a diagnostic tool, is not just absence but pattern of absence. Which query types produce no AI Overviews at all? Where are competitors appearing that a client isn’t? Which keywords produce AI Overviews that cite generic explainer content rather than project-specific sources? Those patterns point toward the structural issues — and the structural issues are what the audit is actually for.
If you’re working through how AI systems are representing your Web3 project, the AI visibility audit is where that work starts.