TL;DR
- The first question to ask about any AI visibility number: was it measured or estimated? Almost always, it’s estimated.
- Semrush estimates how much people ask about a topic, using clickstream data and machine learning. Semrush itself calls its metrics “directional signals”.
- Ahrefs doesn’t measure demand. It asks AI platforms its own questions and records what they answer.
- In Search Console, AI Overviews and AI Mode impressions come as one number, with no clicks, positions or queries.
- Outside English-speaking markets, the samples are thin. My method starts with the business and the buyer, not with a tool.
In classic SEO, keyword tools act as a reality check. It’s worth paying extra for Semrush, Ahrefs or Serpstat, because they tell you who searches for what and how often.
In AI search, there’s no such certainty. There’s no guarantee that the way we think people ask AI is right. There’s no guarantee that the way AI thinks people ask is right either.
A viewer asked me how I do prompt research, so here is my approach. But first, let’s separate the myths from what these tools can actually measure.
Measured or estimated?
This is the simple question I ask every time I look at a number in an SEO tool. Is it measured, or is it calculated?
The difference is huge, and it misleads a lot of people. I’ve talked to business owners who told me: “Semrush says there’s no search volume for my products. Serpstat says 20 searches. I thought there were at least 2,000.” They take the numbers at face value. That’s a gap in understanding what these tools actually show.
How Semrush does it
According to Semrush’s own documentation, its AI Visibility Toolkit is built on more than 317 million prompts and responses from ChatGPT, Gemini, Google AI Overviews and AI Mode, across 117 regional markets, updated daily. It’s also not cheap, and I expect the price to go up.
Where do the prompts come from? Semrush says: clickstream data from AI search, plus Google’s keyword set for AI Overviews.
Clickstream, simply put: some companies collect data on what people do in their browsers. Usually through an extension you install because it does something useful, while the terms say your browsing data is collected anonymously. Or through panels where people agree, often for payment, to share their behavior. That data is aggregated and sold.
The strength of this method is that these are real people and real prompts. The weakness is the sample. Who installs these extensions? Which market do they live in? Which language do they write in? That’s why coverage for a market like Bulgaria is far weaker than for the US or the UK. Not because Semrush does a bad job, but because the sample is what it is. Anyone who has studied sociology knows how much a survey depends on who ends up in it.
The second detail: Semrush doesn’t give you the volume of a single prompt. Individual prompts are too unique to measure directly. So it groups similar prompts into topics and gives you a volume per topic, estimated with machine learning.
The number next to a topic is not a count. It’s an estimate.
And Semrush says it openly: AI answers change fast and are highly personalized, so “no platform can provide exact numbers”. It calls its own metrics “reliable directional signals”. Not measurements. That’s not my opinion. That’s the company’s own wording.
How Ahrefs does it
Most people think Semrush and Ahrefs do the same thing. They don’t.
Ahrefs doesn’t measure what people ask. It asks the questions itself. According to Ahrefs, Brand Radar is seeded with questions from its keyword database and Google’s People Also Ask, paired with search volume. Ahrefs itself calls these prompts synthetic. It runs them through AI platforms and stores the answers. Then you search that archive to see whether your brand is mentioned and your pages are cited.
So: Semrush estimates what people ask and how much. Ahrefs shows what machines answer to questions Ahrefs asked. Two different approaches.
The same goes for Ahrefs’ impressions metric. According to its help center, impressions are calculated by adding up the Google search volumes of the prompts where your brand appears. That’s a model. It assumes that what people search in Google reflects what they ask in AI. Nobody outside the AI companies knows real AI usage data, because it never leaves their servers.
This is a visibility audit, not an audience measurement. It shows what appears, not who saw it.
Ahrefs also offers a few free tools, including an AI Overviews tracker, an AI Mode tracker and an AI visibility checker. They give a limited view, and I wouldn’t call them brilliant for smaller languages, but they’re useful to understand how this works.
A quick reminder while we’re here: AI Overviews and AI Mode are not the same. AI Overviews is the summary at the top, with the blue links still below it. AI Mode is a separate conversational experience with far fewer sources. You can win in one and be invisible in the other.
AI answers keep changing
One Ahrefs study I find very telling: across 43,000+ keywords, around 70% of AI Overview content changed between two consecutive captures. Content lasted about two days on average, and 45.5% of the cited sources were replaced at each refresh.
That’s the global picture. In smaller markets, some businesses sit in niches with very few players and very few sites doing GEO. There, a brand can stay the top AI recommendation for a long time, simply because there’s no competition.
What Google gives you for free
Search Console. In June 2026, Google introduced generative AI performance reports, and by the end of August they were available for all sites worldwide. The report shows impressions in AI Overviews and AI Mode, broken down by page, country, device and date. Discover has a separate generative AI report.
What it doesn’t show matters just as much: no clicks, no positions, no queries. It doesn’t tell you what people asked. And AI Overviews and AI Mode come as one combined number. You can’t separate them yet.
Google Analytics. In May 2026, GA4 added an AI Assistant channel. When someone arrives from a recognized AI assistant, like ChatGPT, Gemini or Claude, GA4 labels it automatically. You’ll find it under Reports, Acquisition, Traffic acquisition, grouped by channel.
Three things to keep in mind:
- In some of the accounts I manage, a setting needed to be switched on first. Check yours.
- It works by referrer. If someone comes from a mobile app that doesn’t pass a referrer, or copies the link and opens it manually, the session lands in Direct. The number you see is the minimum, not the full picture.
- Traffic from AI Overviews and AI Mode is not here. It stays in Organic Search, because technically it comes from Google. For that, you use Search Console.
What I learned testing AI visibility tools for two months
Over the past two months, I tested several specialized platforms. I activated features in Ahrefs and Semrush, tried completely new tools, and ran deep research to find anything else worth testing. I now have so many accounts that I’m happy to save you the time. Some of these platforms did nothing except put me on a marketing list.
I won’t name them, because this isn’t an attack, and the market changes every week. But here’s what I learned:
- Small markets get thin data. Almost every tool, big or small, hits the same problem. Big promises on the landing page, a beautiful interface, and very thin data once you log in.
- Many tools do what you can do yourself, on a schedule. They send questions and record the answers. If you have 20 prompts, you don’t need a subscription. And tools that charge per query, with credits, often don’t add up. It’s cheaper to code your own. That’s what I did.
- The results disagreed with each other more than I expected. Same brand, same week, checks days apart, and very different numbers.
My conclusion: for big markets, big budgets and agencies working at scale, automation makes sense. Pick your methodology, and use an external tool if it gives reliable data for your market. Test it first: check a few prompts manually and compare with what the tool shows. If there’s no big deviation, go ahead.
But for smaller-language markets, I wouldn’t invest in these tools yet. For a local business in a niche, the same money usually works better in content and in professionals who know how to track visibility properly.
My method: business first, then the buyer, then the prompts
The biggest problem I saw: when I entered a website into these tools, they didn’t understand what the business sells or which categories matter most. They grabbed a few keywords, as if they only read the homepage or the meta tags, and covered maybe 40% of what the business actually does.
That’s always been a problem in classic SEO too. When I teach keyword research, many people think whatever the tool spits out is the demand. It isn’t.
For me, most of the work isn’t optimization. It’s understanding the business. Who buys, why they buy, what stops them, what they compare, who they ask for advice, how long the decision takes. That’s the foundation for everything else.
Technical optimization is the easiest part today. Almost any good AI tool can speed up a site, compress images and fix tags. You can connect an MCP server to your AI assistant and run the checks. Running a checklist isn’t where the value of SEO is anymore. The value is in understanding the business and the buyer.
Synthetic buyer personas. This is where AI helps me in a way I think is underrated. I build synthetic buyer personas. Not generic ones like “a 35-year-old woman interested in this and that”. Specific ones, built on everything I know about the business: real customers, real objections, real reasons for saying no, real words from sales conversations.
Then I talk to them. I ask the persona: what stage are you at? What worries you? What would you check before paying? What would you ask ChatGPT right now?
That’s not guessing. It’s structured thinking about the buyer’s journey. Every prompt comes from a stage of a decision. If you have the map of the decision, you have the map of the prompts. Search intent, customer journey, sales funnel: everything we learned in marketing theory applies here, just on steroids, because now we have AI.
Compare the two approaches:
- The tool approach: “The tool says this topic has this much volume” or “The tool says people use this prompt, because I generated variations of my main keyword and it sounds right.”
- The buyer approach: “I know that at stage three, the buyer worries about support. So they’ll ask about support. So I need to be there, with a buying guide or a page on exactly that.”
I follow the second one. It doesn’t depend on a sample or on a number in a tool, and it works the same in any language, for every client.
My order is always: the business, then the buyer, then the prompts, and only at the end, the website.
My own tool, and why I still check manually
When I saw that the existing tools didn’t work for smaller markets, I built my own. It runs a prepared set of prompts many times, in several languages, and crawls the results to collect who gets mentioned and who gets cited. For larger prompt sets, that’s the difference between an afternoon of manual checks and a real dataset collected the right way. It simulates search at the moment, instead of relying on archived data and assumptions.
And still, for the most important prompts, I check manually. It may sound silly, but I sit down and test from different devices that aren’t associated with me. For the hottest prompts, the ones closest to a purchase, I don’t trust automation. For informational prompts further up the funnel, automation is fine to get the overall picture.
For a step-by-step guide with examples and a free prompt library template, read What Is Prompt Research? How to Find the Questions Customers Ask AI.
Two traps even experienced agencies fall into
1. Testing in your own account. If you test with memory on and 500 conversations about your brand in your history, the model knows you. It will mention you more often than it would to a stranger. I’ve had a client show me a screenshot from their agency: “We’re number one for everything!” The screenshot came from the account of the person working on the project. It’s like looking in a mirror. Always test with no history, in an incognito session, as isolated as possible.
2. Testing once. A single test is not a measurement. Remember: around 70% of AI Overview content changes between two observations. Instead of asking once, ask the same question repeatedly, on different days, and record in how many cases you appear. The result shouldn’t be yes or no. It should be a percentage, and ideally a sense of how you’re mentioned: positively, negatively, or not at all.
Want to know where you stand in AI search?
I’m not building a product company. But several clients asked for this after seeing the reporting I do, so I now offer it as a service. If you want to know where your business appears in AI answers, which prompts you win and where competitors are recommended instead of you, get in touch. I’ll prepare the research as a report for you.
About The Author
