Ask an assistant for an electrician in Austin and it hands you a list of businesses. Ask it for a payroll provider and it hands you nobody - 98 readings, not one list.

Key findings

• Across 98 readings of 14 national, product, software, franchise and nonprofit questions, an AI assistant rendered a business list in none of them and named no company in any of the 14.
• In 96 of those 98 readings the assistant ran a web search before answering, so the absence of a business list is not the absence of a lookup.
• The same instrument, on the same days, produced a business list in 16 of 34 readings of plain local-service questions and named at least one business for 8 of those 10 questions.
• Inside the non-local set, the one question that was a local service - an electrician in Austin - produced a business list in 12 of 12 readings and named businesses for 2 of 2 question forms.
• Making a question narrower did not make a business list more likely: 7 of 14 control readings rendered one and 7 of 14 narrowed readings rendered one.

How this was measured

When

Readings taken 4 September 2026 - the local set that morning, the non-local set that same afternoon, on the same rig and the same account.

Instrument

Our own gate rig: a real Chrome instance driven over the DevTools protocol against a logged-in free ChatGPT account. Every answer is stored whole, with its business list and its citation pills, before anything is counted.

What counts as naming somebody

A business name read out of the list the assistant renders above its prose. The prose underneath sometimes names companies the list does not carry; that channel is measured separately and is not counted here. So “named nobody” on this page means “rendered no business list”, not “the answer was empty”.

The two sets

Ten local-service questions in ten trades outside the 46 we publish pages for, all in Austin. Fourteen questions in seven kinds of thing that are not a local service at all, each asked in a plain and a narrowed form, with no city in them.

Repetition

Every question was asked between three and seven times, in a fresh chat each time, so a single unlucky answer cannot carry a row.

What we measured

Everything else we publish measures how an assistant answers a question about a local business: which businesses it names, how stable that is, whose pages it reads. All of it assumes the answer has businesses in it to begin with. This measurement tests that assumption, because if it fails for your kind of business then nothing else we sell can help you and we would rather find that out on our own data than on your money.

So we asked two sets of questions with the same rig on the same day. The first set was ten ordinary local-service questions in ten trades we have never published a page for. The second set was fourteen questions that are not local services at all: inventory management software, trail running shoes, a payroll provider, a conveyor belt manufacturer, an SEO agency, a franchise to buy, an environmental nonprofit - each asked plainly and then again with the kind of detail a real buyer would add.

The result

The split is not a tendency. It is a wall.

A business list is the block of named businesses an assistant renders above its prose answer. Every question was asked in a fresh chat, three to seven times.

It is not that the assistant did not look

The obvious reading of a zero is that the assistant answered out of its own memory and never went to the web at all. That is not what happened. In 96 of the 98 non-local readings the assistant ran a web search before it answered, and we have the citation pills to show for it. It looked, it read pages, and it still wrote prose about what to consider rather than a list of companies to consider.

The clean control sits inside the same set. One of the sixteen non-local questions was not non-local at all: we asked for an electrician in Austin, and then for a panel upgrade on a 1960s house in Austin. Same rig, same session, same account, same day as the payroll and conveyor-belt questions. Both forms produced a business list in 12 of 12 readings. Whatever decides this is in the question, not in the weather.

Narrowing the question changed nothing

The second thing we set out to test was the advice everyone in this category gives: that assistants reward specificity, so a business should write for the narrow question rather than the broad one. We built the wave to test it - every question in a plain form and a narrowed one, and a separate set of four local trades asked both ways.

It did not survive. Across the four local trades asked both ways, 7 of 14 control readings rendered a business list and 7 of 14 narrowed readings rendered one. The same number, twice. And across the non-local questions, narrowing moved nothing either, because there was nothing to move: zero both ways.

We are reporting a null result on our own idea, and it is worth saying what it does not mean. It does not mean specificity is useless - our other pages show it changes which businesses are named. It means it does not conjure a business list where the question does not call for one.

The finding we had to throw away

The first version of this page carried a different, better-looking result: a table showing the number of businesses named falling to zero as questions narrowed. It was an artefact. Our raw file holds one row per reading, and the first pass read one row per observation - so a question asked seven times was counted once, from whichever reading happened to be first in the file.

Recomputed across all readings, that table collapsed and the two findings above are what was left standing. We publish this because a measurement you cannot check is not a measurement, and because the version of this page you would have read a day earlier would have been wrong in a way you could not have detected from the page.

What this means if you are buying

• If your business is a local service that customers ask for by place - a trade, a clinic, a venue, a firm people visit - the mechanism we sell readings of exists for you, and the rest of our research is about how to be in that list.
• If you sell software, a physical product, a national service with no city attached, or a franchise, an assistant answering your buyer’s question does not render a list of companies for us to measure your place in. Buying an AI visibility reading for that question buys you a measurement of nothing. Anyone selling you one should be asked for their own version of this page.
• Two of the ten local trades never produced a list either - commercial real estate and wedding venues - so being local is necessary here and is not by itself sufficient. That is the reason we measure a business’s own trade and city before we sell anything else.

Honest limits of this measurement

• We count a business only from the list the assistant renders. The prose underneath it sometimes names companies the list does not carry, so a zero here means no business list, not an empty answer. That is the honest claim and it is also the commercially useful one: the list is the thing a business is in or out of.
• One assistant, one account type, one day. A free account cannot pin the model tier, so the tier is recorded per reading rather than chosen.
• The non-local questions carry no city because that is what makes them non-local. We therefore cannot separate 'not a local service' from 'no place in the question' - they are the same variable here, and pulling them apart needs a wave we have not run.
• Fourteen questions in seven kinds of thing is a small set. It is enough to say that the list does not appear for any of them and not enough to rank them against each other.
• The ten local trades are ten trades in one city. A trade that was silent in Austin may not be silent elsewhere.

Questions about this research

We run this measurement for a living. If you want it run on your own category and your own city, the free scan is the same instrument on one question, and the report is the full set. Nothing here is behind a form.

More measurements from the same instrument

• People ask about AI visibility on Google, not inside ChatGPT
• How many companies can an AI assistant name at all?
• All eight measurements →

Every category we measure has its own page, with the questions customers put to an assistant in that category and what a reading costs. See all industries we measure →

Published 31 August 2026. Every figure on this page is recomputed from raw observation files by a single script, so any of them can be traced back to the readings behind it. If you are named here and believe a reading is wrong, tell us: we will run the measurement again and publish what it returns, whichever way it goes.