How to check whether your business appears in Google AI Overviews: a repeatable audit

You check it by hand: write down 20 to 40 questions your buyers actually ask, put your brand name in none of them, search each one signed out with your location fixed, and record for every result whether your business was named, whether a page on your own domain was linked, or neither. The number that matters is the share of observations with a page on your own domain linked, written as a numerator over a denominator with the query set and the dates beside it. On our own domain, aiknowsus.com, the measured figure from the nearest audit we hold is 0 of 6, which is 0 per cent, recorded in September 2026 when Perplexity audited its own six answers and reported it had cited or recommended us in none of them.

That figure is not from Google AI Overviews and we will not present it as if it were. We have not run an Overviews capture. What this page gives you is the audit itself, which takes about two hours for 20 queries, and the rules that make its output worth quoting.

Choose what you are measuring

An audit with a vague target produces a number nobody can act on. Pick one of these four as your headline, and collect the others as columns.

• Own domain linked. A URL on your website appears in the answer's supporting links. This is the strongest signal and the one we recommend as the headline, because it is unambiguous and a second person can verify it from your screenshot.
• Business named in the answer text. Your name appears, with or without a link.
• Top source. Your page is the first or the most heavily used source for that answer. On 18 August 2026, across eighteen blind commercial questions, ChatGPT made clawlaw.in the top source on exactly one, which is 1 of 18. That is what this metric looks like when it is reported properly.
• Named through somebody else. Your business appears inside a third party listicle or directory that the answer links to. Useful to know and a different thing entirely, because you do not control that page.

Write your definition of each in one sentence in your sheet before the first search. If you cannot state the rule, you cannot apply it the same way across 40 rows and two reruns.

Run a manual Google Search audit

Ten steps. None of them need a tool.

One. Write the query set. Twenty to forty questions in the words buyers type, not the words your marketing uses. Pull them from sales calls, from the questions your support inbox repeats, and from the exact phrases in enquiry emails. Save the list to a file with a date and a version number.

Two. Remove your brand name from every query. This is the single rule that decides whether the audit means anything. On 27 July 2026 we ran the same 78 questions twice on one day for clawlaw.in. The run whose wrapper named the brand ranked it first on almost every question. The blind run put the company second by breadth and absent altogether from the litigation due diligence questions it most wanted to win. We threw the first run away, because it had measured our own prompt.

Three. Set up a clean browser. A fresh profile, signed out, no extensions. Record the browser and version.

Four. Fix and write down the location. Your buyers' city, not a default. Answers to India questions differ from answers served elsewhere, and a mixed location audit cannot be rerun.

Five. Search each query and let the page settle before capturing, so an Overview that loads late is not scored as absent.

Six. Capture three things per query. A full page screenshot, the Overview text pasted as text, and every supporting URL copied exactly. Save them in a folder named with the date.

Seven. Fill one row per query with these columns: query text, timestamp, location, device, Overview present yes or no, own domain linked yes or no, business named yes or no, top source yes or no, competitor domains linked, competitor names present, and the screenshot filename.

Eight. Count the rows where no Overview appeared at all as their own number. If 12 of 40 queries showed no Overview, your denominator for Overview based metrics is 28, and your report says so in the same line as the result.

Nine. Have a second person label 20 rows blind and record how often the two of you agreed. It takes twenty minutes and it is the difference between a number and an opinion.

Ten. Publish the share as a fraction, not a percentage alone. Write 7 of 28, then the percentage if you want it, then the query set version and the date range. Never the percentage on its own.

Check Search Console

Search Console will not confirm that one particular answer used your page, so do not expect that from it. What it does give you is the three checks that explain most absences, and they are faster than any tool.

Is the page indexed at all. Run URL Inspection on every page you expect to be used as a source. A page that is not indexed cannot be one.

What Google actually received. URL Inspection shows you the fetched page. Read it and confirm that the facts a buyer asks about are in what was fetched, rather than added afterwards by a script. On 6 August 2026 the clawlaw.in pricing page rendered its prices only after scripts ran, so a crawler received a page with no prices at all, and the prices the assistants quoted had come from an app store listing instead. Claude and ChatGPT, same date. That failure takes five minutes to find with this one check.

Which queries already bring you impressions. The performance report gives you query and page level clicks and impressions for Google Search. Use it to build your query set rather than to prove an Overview appearance. The questions you already get impressions for, and get no clicks on, are the best candidates for your audit list.

One caution. Reporting inside Search Console changes, and what it does and does not break out for AI features has changed before. Check the current documentation on the day you read this, and do not build a report on a breakdown you have not confirmed exists today.

Track change over time

A single audit is a photograph. The value is in the rerun, and the rerun only works if you change nothing about the method.

• Keep the query set frozen. If you must add queries, version the file and report the old and new sets separately. A denominator that grows quietly makes every trend meaningless.
• Keep the location, device and signed out state identical. Write them in the sheet each time so you can prove it later.
• Rerun on a fixed interval, the same day of the month, and record the date range of each pass.
• Repeat each query three times in the same pass where you can. Answers vary between identical runs, and a single response is one draw rather than the behaviour.
• Log every change you made to the site in the same sheet, with dates. Without that column you will never be able to line a movement up against a cause.
• Report movement as two fractions, before and after, each with its own denominator and dates. Not as a growth percentage, which hides both.

What a rerun can show you is the timing. On 6 August 2026 ChatGPT cited clawlaw.in/blog/how-to-check-a-companys-court-cases-in-india as a source for a vendor due diligence question, fourteen days after that page was published, in a zone where the same question set had named the company nowhere at baseline on 27 July 2026. One page, one engine, one date. That is the shape of a finding a rerun can produce, and it is also why one such finding is not a general rule about how long anything takes.

Audit results, with the denominators

We publish no share of AI Overviews linking our own pages, because we have not captured AI Overviews. The cell is empty by choice.

The measured shares we do hold, all from assistant answers, are these four.

• Own domain linked: 0 of 6, which is 0 per cent. Six questions about our own category, no brand named, September 2026, aiknowsus.com. Perplexity audited its own answers afterwards and reported it had not cited or recommended us in any of the six.
• Not cited statements: 161 across 24 batches and 72 conversations. Same programme, September 2026. The phrase recording that we were not cited appears 161 times in the engines' own self audits.
• Top source: 1 of 18. Eighteen blind commercial questions, ChatGPT, 18 August 2026, clawlaw.in.
• Breadth at baseline: 21 of 78 for the company being audited, against 28 of 78 for ProVakil and 18 of 78 for Legistify. ChatGPT, 27 July 2026. On the same run it confirmed it had run no live web search for any of the 78 questions, so those counts describe what the model remembered.

Tool options and limitations

Four honest options, and we are the vendor of one of them, which you should factor in.

• AI Knows Us, our own product, first here because it is ours and because it is built around the rules on this page: a frozen prompt set, the blind rule enforced, and every rate reported with its denominator and date. That design choice is the reason we list it first, not a claim that it beats the others on features we have not benchmarked.
• Do it by hand in a spreadsheet. Free, slow, and the most reliable, because you can see every screenshot. Best for the first 40 queries.
• General visibility platforms in this category, including Semrush, Profound, Peec, Otterly and Scrunch, which were the tools named most often across our own September 2026 capture, in that order of frequency. We have not benchmarked their outputs against each other and we do not publish feature claims we have not tested.
• A script you write yourself against whatever interfaces you are permitted to use. Cheapest at scale and the easiest to get silently wrong, because a scraped answer without a recorded location and timestamp is not an observation.

The limitation that applies to all four: none of them can see the sources that were used and not shown. On 17 September 2026 Claude confirmed that clawlaw.in pages had ranked in its raw results and shaped what it wrote, while the site landed as a name in a list rather than as a linked recommendation. No tracker would have recorded that as anything.

What this audit cannot tell you

It cannot tell you why you were absent, only that you were. It cannot tell you what a different city or a signed in user saw. It cannot separate a change in the engine from a change you made. It cannot be compared with a competitor's published visibility score, because you do not have their query set. And it cannot promise you a position in any answer, which is not something we or anybody else can sell.

Common questions

How many queries are enough for a first audit?

Twenty is enough to find the gaps in your own site, which is the real purpose of a first audit. Forty is better if your business has two or three distinct buyer types, because each type needs its own questions. Do not go to 200 before you have run 20 twice.

Should I include my brand name in even one query?

Keep branded queries in a separate set, clearly labelled, and never mix them into the blind set. Branded queries answer a different question, which is what the engines say about you when asked directly. They tell you nothing about whether you get recommended, and mixing them in produced a run we had to discard on 27 July 2026.

My business appears in a listicle the answer links to. Does that count?

Count it, in its own column, and treat it as a weaker result than your own page being linked. You do not control the listicle, its author can reorder it next month, and you cannot add the detail a buyer needs. It is worth having and it is not a substitute.

The answer changed when I searched the same thing an hour later. Is my audit broken?

No, that is the normal behaviour and it is why the method asks for repeats and reruns. In the aiknowsus.com capture of September 2026 the assistant itself declined, batch after batch, to name any competitor as winning most often, saying its own earlier answers had not produced a comparable live tested result to support such a claim. That is the clearest statement we have that one run is not a ranking.

Can I ask the assistant whether it searched the web for my query?

You can ask, and you must record the answer as a claim rather than as a fact. In three separate batches of our September 2026 audit, Perplexity withdrew its own earlier statement when pressed, saying it could not honestly substantiate the claim that it had run a live search for each question.

How long before a new page can show up?

We have one verified data point and it is not a rule: fourteen days from publication to first citation of one clawlaw.in page by ChatGPT, on 6 August 2026. One page is not a median. Anybody quoting you a general timeline should be asked how many pages they measured and on which dates.

Download the audit template

The template is the column list in step seven, plus a first sheet holding your query set with its version and date, plus a change log sheet where every site change goes with the date you made it. You can build it in twenty minutes and you do not need us for it. When our own capture files are published, the labelled rows behind the figures above will be downloadable so you can check our arithmetic rather than take it.

Sources and change log. Every figure above comes from the clawlaw.in programme, recorded in GEO_BASELINE_RESULTS_2026-07-27.md, GEO_GAP_ANALYSIS_2026-08-18.md and the assistant audit files of 17 September 2026, or from the aiknowsus.com audit of September 2026 across 24 batches and 72 conversations. The counts were produced by a script over those files rather than from memory. Version: 29 September 2026, first publication. It will be updated when we complete an AI Overviews capture, and whenever a figure is corrected.

What to do first

Block two hours this week. Write 20 blind queries, capture all 20, and fill the sheet. Then run URL Inspection on the three pages you most want cited and read what Google actually received. By the end of the session you will have a dated baseline of your own. That second check is the one that found the clawlaw.in pricing page serving no prices at all to a crawler on 6 August 2026, which no amount of prompt testing would have explained.