What we can and cannot measure about which businesses Google AI Overviews name

You can measure how often a business is named inside a sample of AI Overviews that you captured yourself, on a query set you wrote down, on dates you recorded. You cannot measure how often Google names a business in general, and nobody outside Google can. So the honest form of this measurement is a share with four things printed next to it: the numerator, the denominator, the exact query set, and the capture dates. As of 29 September 2026 we have not published an AI Overviews capture of our own, so this page gives the protocol in full and prints only the counts we hold, which come from ChatGPT, Claude and Perplexity runs.

That is a less satisfying answer than a percentage, and it is the only one that survives being checked. A page that says a certain share of AI Overviews name a business, with no query set and no dates behind it, is describing a sample of unknown size taken at an unknown time in an unknown country. We have watched assistants discard exactly that kind of page in our own runs, and we have watched one of them withdraw its own claim about how much work it had done. Both are recorded below with dates.

What counts as a business being named

Before any counting starts you need one definition, written down, that a second person could apply to the same screenshot and get the same answer. Otherwise two people audit the same hundred Overviews and report different numbers, and the difference is the definition rather than the web.

We use four separate labels, and they are not interchangeable. Most published visibility numbers quietly mix them, which is why they cannot be compared with each other.

• Named in the answer text. The business name appears in the words of the Overview itself. No link required.
• Linked as a source. A URL on the business's own domain appears in the Overview's supporting links, whether or not the name appears in the text.
• Named and linked. Both at once. This is the only state that sends a reader to you and also tells them who you are.
• Present but not attributed. The answer clearly uses material from your page, and neither your name nor your URL appears. This one is invisible to every automated tracker we know of, and it happens.

The fourth label is not theory. On 17 September 2026, in the clawlaw.in programme, Claude admitted it had used two specific arguments drawn from clawlaw.in pages and had dropped the attribution in both places, describing it as a citation lapse rather than a ranking judgement. If your definition of being named does not have a slot for that, your audit will score it as a zero and you will conclude the pages did nothing.

Two more rules keep the labelling stable. A brand named inside a third party listicle that the Overview links to is not the business being named, it is the listicle being named, so it gets its own label. And a business named in a follow up answer after the reader asked a second question is a different observation from the first answer, so it is recorded as a separate row with its own prompt text.

What Google publicly documents

Google's public help material describes AI Overviews in general terms: that they appear for some searches and not others, that they include links out to pages that support the answer, and that a reader can send feedback on an Overview from the answer itself. Those three facts are the ones worth building a method on, because they are stable and they are checkable by you in a browser today.

We deliberately do not restate Google's documentation as dated quotations on this page. The wording changes, and a paraphrase of a help page with no date next to it is exactly the kind of source we are asking you not to trust. Open the current help page on the day you read this and check it yourself. If what you find contradicts anything on this page, tell us and we will change the page.

What Google has not publicly specified

This list is the reason a clean percentage is not available, and it is worth being concrete about. The following six things are not published in a form you can measure against.

• Which queries trigger an Overview. There is no published rule, and triggering changes over time for the same words.
• How the supporting links are selected from the pages the system considered.
• Whether a page was read live or came from an index at the moment your Overview was generated.
• How much your location, language and signed in state change the answer. They change it. The size of the change is not published.
• How long any given Overview is cached or how often it is regenerated for the same query.
• What weight commercial intent carries in whether a business gets named at all.

None of that makes measurement pointless. It makes the sample the subject of the measurement. You are not measuring Google. You are measuring what Google showed to a particular browser, in a particular place, on particular days, for a list of queries you wrote down.

Our observation method

This is the protocol in full. It is written so that somebody with no access to our files can run it, and so that two people running it separately can compare their results.

One. Build the query set and freeze it. Write between 20 and 100 queries in the words a buyer types, save them to a file, and do not edit that file again for the length of the study. If you add a query later, the file gets a version number and a date, and every rate you publish states which version it came from.

Two. Apply the blind rule. No query may contain your own brand name. This is not a style preference. On 27 July 2026 we ran the same 78 questions twice on the same day for clawlaw.in. The first run used a wrapper that named the brand and came back ranking it first on almost every question. The second named nothing, and put the company second by breadth and absent altogether from the litigation due diligence questions it most wanted to win. We discarded the first run, because the only thing it had measured was our own prompt. Any audit with the brand in the prompt has the same defect.

Three. Fix the capture conditions and record them on every row. Browser and version, signed out or signed in, country and city setting, language, device type, and the exact timestamp. An Overview captured on a phone in Pune signed in is a different observation from one captured on a desktop in Delhi signed out, and mixing them silently is the most common way an audit becomes uncomparable with itself.

Four. Capture the evidence, not the conclusion. For each query save the full page screenshot, the visible Overview text as text, and every supporting link URL exactly as given. A spreadsheet row saying yes or no is not evidence. If somebody later disputes a number you need the screenshot.

Five. Label with the four states above, by a person, twice. Have a second person label a sample of at least 20 rows without seeing the first labels, and publish how often the two agreed. A published share with no agreement check behind it is one person's reading.

Six. Repeat the whole set on a fixed interval, weekly or monthly, with nothing else changed. A single pass gives you a photograph. Only the repeat tells you whether anything moved.

Seven. Record search behaviour as a claim, never as a fact. Where the interface tells you what it consulted, save that. Where you have to ask the system, treat the answer as unverified. In the aiknowsus.com audit of September 2026, Perplexity withdrew its own earlier statement when asked how many questions it had actually searched for, saying it could not honestly substantiate the claim that it had run a live search for each one, and in another batch that its claim to have searched all five was not adequately supported. Three separate batches produced that retraction.

Findings, with counts and examples

Here is the part most pages on this subject get wrong, so it is stated flatly. We hold no AI Overviews capture, so we publish no share of Overviews that name a business. Not an estimate, not a range, not a figure borrowed from somebody else's study. The cell is empty and it stays empty until we have run the protocol above and published the captures.

What we do hold are counts from assistant answers, from two programmes, and they are printed here with their denominators because they show what the same measurement looks like when it is done properly.

• 0 of 6. Asked six questions about its own category with no brand named, Perplexity audited itself afterwards and reported that it had not cited or recommended aiknowsus.com in any of the six answers. September 2026, our own domain. There was no position for us to hold.
• 161 statements across 24 batches and 72 conversations. Across the whole aiknowsus.com capture, the phrase recording that we were not cited appears 161 times in the engines' own self audits of their answers. September 2026, Perplexity.
• 1 of 18. Across eighteen blind commercial questions on 18 August 2026, ChatGPT made clawlaw.in the top source on exactly one, and named the company's own comparison page in the answer as still being the vendor's own editorial page.
• 28, 21 and 18 out of 78. In the blind baseline of 27 July 2026, ChatGPT's breadth order was ProVakil on 28 questions, CLAW on 21 and Legistify on 18. That is a ranking of what the model had absorbed, not of what was true that week, because on the same run it confirmed it had run no live web search for any of the 78.
• 0 of 78 searched. Same run, same date. Every one of those breadth counts came from memory.

The last two lines together are the most useful thing in this section. A visibility number produced without live search is a measurement of training data, and it is still worth having, as long as nobody reads it as a measurement of the web this week.

What the findings do not prove

Five limits, and each one has cost somebody a wrong conclusion.

• None of the counts above are AI Overviews. They are ChatGPT, Claude and Perplexity. The systems select sources differently and a result from one is not evidence about another.
• A sample describes itself. 0 of 6 is a fact about six questions asked in September 2026. It is not a rate for our domain and it is not a rate for the category.
• Counts over our own capture files cannot be checked by you yet. Of the observations behind this page, the citation of a named URL by ChatGPT on 6 August 2026 can be verified from outside. The counts over our capture folders cannot, until we publish the captures.
• A change after a change is not a cause. If a business starts being named after you publish pages, the pages are one candidate explanation among several, including the engine changing its own behaviour that month.
• No method here can promise you a position in an answer. We do not sell that and we would not believe anybody who did.

Sources and change log

Every number on this page comes from one of two recorded bodies of work. The clawlaw.in programme from July 2026, recorded in GEO_BASELINE_RESULTS_2026-07-27.md, GEO_GAP_ANALYSIS_2026-08-18.md and the 17 September 2026 assistant audit files. And the aiknowsus.com audit of September 2026, 24 batches and 72 conversations, captured outside this repository. Counts were produced by a script over those files rather than from memory.

Version of this page: 29 September 2026, first publication. Changes that will trigger an update: our first AI Overviews capture being completed, in which case the empty cell above gets a numerator, a denominator, a query set version and dates. A correction to any figure above. Publication of the capture files, which will move several of these observations from unverifiable to checkable.

Common questions

Why not just publish an estimate with a caveat under it?

Because the number gets quoted and the caveat does not. A figure lifted out of a page travels without its footnote, and in our own runs we have seen the opposite of a caveat happen too: on 17 September 2026 Claude checked two clawlaw.in comparison pages that stated competitors' prices with no link, no date and no source. The figures were correct when it checked them independently, and it still treated the pages as advocacy and used official sources instead. An unsourced number does not gain trust from being hedged.

How many queries do I need before a share means anything?

There is no threshold that turns a small sample into a general truth. What changes with size is how much the number moves when you rerun it. Twenty queries is enough to find out what your site is missing. It is not enough to publish a rate and defend it. Whatever size you use, print the denominator beside the number every single time.

Can a tool do this capture for me?

Tools can collect at a scale you cannot match by hand, and that is a real advantage. What you still have to obtain from the tool is the denominator, the exact prompt list, the location and device settings, the dates, and whether the labelling was done by a person or a model. If a tool will not give you those, its percentage is not auditable, and we are the vendor of a tool and still say so.

Does being named without a link have any value?

It has value and it is hard to bank. The reader hears your name and can search for it later. What it does not do is send the visit, and it will not appear in your analytics, so a business measuring only referrals will conclude nothing happened. Count it as its own label, as above, and report it separately.

Our brand name is also a common word. How do we count it?

Write the disambiguation rule into your labelling sheet before you start, with examples of a match and a non match, and have the second labeller apply the same rule. Then publish the rule. Most disputes about a visibility number are really disputes about string matching.

Download the data and report an error

There is nothing to download for the Overviews cell, because it has not been run. What is available now is the protocol above, which you can copy and run this week, and the observation record behind the counts, which we are preparing for publication so that the unverifiable rows become checkable.

If you find an error on this page, including a Google help page that now says something different from what we describe, write to us through the contact page on this site and say which sentence is wrong. We correct the page and note the change in the log above rather than editing quietly.

What to do first

Write 20 queries in your buyers' words, with your brand name in none of them. Capture the Overviews for all 20 today, signed out, with your location fixed and recorded. Label each one with the four states above. That gives you a dated baseline of your own by this evening, which is more than any published percentage can give you, because it is about your business.