The tooling that measures AI search cannot see a local business
A metric sold as AI search volume returned 0 for every local commercial term we put to it, including one where our own stored transcripts show two of four engines naming the business by name. The zero is a fact about the instrument, not about the market.
Tools for measuring AI search are being sold before they work. One of them reports how often people ask an AI about a given topic. We put it to five terms we already had recordings for. It said nothing was being asked, on the same topics where we can play you the recording of an assistant answering. Anyone planning a website around that number is being told to ignore the local questions their customers are asking right now.
Macdara Ó Murchú - Runs the monthly AI citation tracking behind Solaisoft's client reporting.
Published - the work it reports ran
Numbers fromlib/aeo-evidence.tslib/tools/opportunity-sizing.ts
- AEO
- Measurement validity
- Keyword data
Somebody built a counter for a thing that cannot be counted.
For twenty years, being found online rested on one convenience: people typed a few words into a search box, and those words could be counted. If ten thousand people a month typed gym perth, you knew what to put on your website. The counting is what made the whole industry work.
Assistants broke the counting. Nobody types three words at ChatGPT. They ask where is a good gym near Claremont that does small group strength work, the next person asks it a different way, and neither of them will ever type the same phrase twice. There is no word left to count.
So when a company sells you a number for how often people ask an AI about a business like yours, the fair question is what they counted. We asked. For every local business term we tried, their answer was that nobody is asking, while we hold recordings of assistants answering those same questions and naming businesses by name.
The number is not measuring your customers. It is measuring what a keyword counter can see, which turns out to be almost nothing outside a handful of generic national terms. The rest of this post is that test, with the figures and the recordings.
Zero for every local term, and one number for the term nobody can win.
DataForSEO sells an AI keyword data endpoint that reports search volume inside AI assistants. It is on our account, it returns real-looking numbers, and that is what makes it tempting. We put five terms to it on 8 August 2026, four of them the kind of thing a local business wants to be found for.
| Term asked about | Reported volume |
|---|---|
| gym claremont | 0 |
| handyman clonakilty | 0 |
| best gym in perth western suburbs | 0 |
| azure certification practice exam | 0 |
| gym | 77,314 |
Measured live 8 August 2026 and recorded in lib/tools/opportunity-sizing.ts. The last row is the bare generic head, included as the control.
- The control is the finding
- Only
gymreturned a number, andgymis the one term no business can rank for. Every term carrying a place name, a qualifier or a buying intent came back empty. - A zero is a claim
- Read plainly, these rows claim that nobody asks an AI assistant about gyms in Claremont or handymen in Clonakilty. That is a strong claim about human behaviour, and the next section shows a transcript of it happening.
We already had a measurement of the thing it says is not happening.
One of those five terms is a topic we track monthly. On 4 August 2026 a panel of 20 questions went to four engines for AzurePrep, our own product, with no brand name in the wording. The answers are stored verbatim.
azure certification practice examAzurePrep Claims the Top Spot for Largest Azure Practice Exam Question Bank
Google AI Overviews, 4 August 2026
- Named, by two of four
- Asked which Azure practice exam site has the largest question bank, Google AI Overviews and Perplexity both named AzurePrep in the answer text. ChatGPT and Claude cited the page as a source without naming it.
- Quoted, with a figure
- AI Overviews opened with AzurePrep claiming the top spot for largest Azure practice exam question bank. Perplexity gave a number and a domain. Neither answer is ambiguous about whether the topic gets asked.
- The same topic scores zero
- The reported AI search volume for
azure certification practice examis 0. Two measurements of the same question disagree completely, and one of them is a stored transcript of what the engines said. - Which one is wrong
- When a model and an observation disagree that hard, the observation wins. The transcript is what four engines produced on a dated run. The zero is a keyword model's opinion about a medium it cannot read.
A keyword model pointed at a medium with no keywords.
The failure is structural rather than a bug in anyone's pipeline, which is why waiting for the number to improve is the wrong response.
- Nobody types like that
- A search box trained people to type
gym claremont. An assistant getswhere is a good gym near Claremont that does small group strength work, phrased differently by every person who asks. There is no short string left to count. - The zero means blind, not absent
- A metric that can only count short keyword strings returns 0 for conversational demand whether that demand exists or not. The reading is the metric cannot see this, and it is being sold as there is nothing here.
- It ranks generic above local
- Used as a content signal, it scores national heads and blanks every specific local topic. It steers a business away from the specificity that earns citations, toward the terms it has no chance of winning.
How to test a metric for a medium it may not be able to see.
- Ask it about something you already measured
- The whole test here was one term we held a transcript for. Any metric can be checked this way, and it costs one call against data you already have.
- Watch which direction the zeroes fall
- Scattered zeroes look like thin demand. Zeroes that line up with one property, in this case anything local or specific, describe the instrument rather than the market.
- If you are the business, not the agency
- The demand is there. People ask assistants for a gym in Claremont and a handyman in Clonakilty, and the assistants answer with somebody. What does not exist is a keyword-shaped number to put next to it, which is why we report the answers themselves: the question, the engine, the date, and whether it named you.
- Narrow the use rather than banning the tool
- We kept one use for it: a tiebreaker between two comparable generic topics. Any local against generic comparison is out, and so is dropping a topic because it scored zero. What we do not do is put the figure in front of a client, because it would contradict the method we publish.
Have something to build?
Tell us what you are working on. We reply within a day, in plain language, with a clear next step.
Perth, Western Australia