Zero of forty-six

This week an answer-engine optimisation company raised 180 million dollars at a 1.8 billion valuation, seven months after its previous round. The same week I was writing up the only visibility measurement I have run with any care, for a leadership-development provider in the German-speaking market. Across 46 real web searches by three AI assistants, answering the questions a buyer in that market would actually ask, the firm was cited zero times. The number is correct. What it means took three rounds of measurement to understand, and the first round was partly measuring my own setup.

The number, and how it was earned

The setup was plain. Twenty-two questions: eighteen of the kind a buyer asks before they know any vendor's name — who helps with leadership development for a particular role, what a board should look for, how a family business approaches succession in management — and four brand questions that name the firm directly. Each question went to four assistants, all with web search switched on, for 88 measurements.

On the eighteen demand questions, three assistants actually went to the web: one in all 18 cases, one in 16, one in 12. That is 46 real searches that returned real sources. The firm appeared in none of them. Other names appeared repeatedly — large consultancies, established academies, a handful of specialists — often ten times or more across the set.

On the brand questions the picture reversed: 13 of 16 answers cited the firm correctly. Ask for it by name and the assistants find it. Ask for what it does and they find someone else.

The first round was partly measuring me

That result is the third version. The first run had asked three of the four assistants through a plain chat interface without a search tool, and they returned "no sources available" on every single question. That is not a market finding. A model without a search tool cannot cite anything; I had paid for twelve calls that measured the absence of a tool.

The second run found two bugs in the scoring. An answer that cited the right domain but wrote a shorter form of the name was counted as "not cited", which moved brand questions from two of four to four of four once fixed. Another provider returned redirect links in place of source addresses, with the real domain hidden in the link title. Neither bug changed the demand-question zero, but both would have changed a slide.

And the fourth assistant in the final run never searched at all — zero of eighteen. It answered from memory every time. The first draft of the summary said "all four assistants agree". They did not. One of them had said nothing that could be checked, and silence is not agreement.

Invisible, or unaddressable

The more useful discovery was in the brand questions that went wrong. The programme under test is a sub-brand without its own web address; it lives inside the parent company's site, and the person who is the face of the company is found reliably, with a personal domain cited again and again. The programme name, however, is an ordinary word, and it is also a registered trademark of an unrelated company in a completely different industry. Asked what the programme is, the assistants filled the gap with a dictionary entry and the other company's product.

It is tempting to call that hijacking. It is not. Nobody took anything. There was no entity to find — no page with a fixed address that says, in a form a machine can quote, what this programme is, who it is for, and who runs it. The assistants did what they always do with a gap: they filled it with the nearest thing that had an address.

That changes the repair. A ranking problem is solved with more content, more mentions, more reach. An address problem is solved with one page that did not exist before, and structured data that tells a crawler what the page is about. The first is a budget line. The second is an afternoon.

What the funding round does not tell you

The venture money confirms that the category is real. It does not tell a mid-sized firm which of its problems it has. The same week, an infrastructure provider introduced a label that lets site owners opt out of AI training without losing search ranking — a genuine lever, and one that solves a different problem. Opting out of training is not opting out of being summarised. The answer still gets written; the only question is whether your page is one of its sources.

Three questions are worth answering before buying any visibility tool or service:

• Does the measurement actually search? A chat interface without a search tool will always report that nobody cites you, and that report costs money.
• Are you absent, or are you unaddressable? Ask for yourself by name. If the answer describes someone else, the first fix is a page, not a campaign.
• Which of the numbers are measurements and which are silence? Count only the answers where the assistant demonstrably went looking.