Google AI Overviews for marketing agency queries: a dated local and industry search audit

For agency queries, whether an AI Overview appears at all depends on the exact words, the city, the device and the day, so the only measurable result is a snapshot with all four recorded: how many queries you tested, how many of those showed an Overview, and how many of those Overviews linked an agency domain, with the city and the dates in the same sentence. As of 29 September 2026 we have not run an AI Overviews capture, so the number of agency queries we have tested is zero and no proportion is published here. What is published is the audit in full, the way to keep location honest, and the counts we do hold from assistant answers, each with its sector and date.

Location is why this page is separate from the SaaS version of the same audit. A national software query can be captured once. An agency query captured in one city and reported as a national rate is a measurement of where the person doing the audit happened to be sitting.

What Google documents, and what this study observes

We do not restate Google's documentation on this page. Help pages are edited without notice, and an undated paraphrase becomes a false claim that our page cannot detect. The cost of undated information is measured: on 17 September 2026 Claude checked two clawlaw.in comparison pages stating competitors' prices with no link, no date and no source, found the figures correct when it checked them independently, and used official sources instead.

So read these three yourself on the day you read this, and write the URL and the date next to anything you take from them.

• Google's current help material on AI Overviews, for what it says about when they appear and how links are shown.
• Google Search Central documentation, for how your pages are crawled, indexed and made eligible, which is a separate system from the answer.
• Google Business Profile help pages, because local results and map listings are another separate system, and an audit that confuses a local pack with an Overview is counting two things as one.

What this study observes is narrower than any of that documentation. It observes what one browser, in one city, on one date, was shown for a list of queries somebody wrote down. The following five things are not published in a form you can measure against, which is why the sample is the subject and the platform is not: which queries trigger an Overview, how the displayed links are chosen from the pages considered, whether a page was read live or came from an index, how much location and signed in state change the result, and how long an Overview is cached before regeneration.

Query design and method

Three query types, three frozen files, three denominators. Local intent, where the buyer names a place. Industry intent, where the buyer names a sector or business model. Task intent, where the buyer names the work to be done rather than the supplier.

For the local file, name the cities rather than letting your connection choose them. A national starting set is Mumbai, Delhi, Bengaluru, Pune, Hyderabad, Ahmedabad and Jaipur, with one row per city per query. Run only the cities you sell in if that is fewer.

Eleven steps.

One. Write 20 to 40 queries per file in buyers' words, taken from sales calls, repeated enquiry emails and the questions a new client asks in the first meeting. Save with a version number and a date.

Two. Put your agency name in none of them. On 27 July 2026 two runs of the same 78 questions happened on one day in the clawlaw.in programme. The branded wrapper ranked the brand first on almost every question; the blind run put the company second by breadth and absent from the litigation due diligence questions it most wanted to win. The branded run was discarded, because it had measured our own prompt rather than the market. ChatGPT, 27 July 2026.

Three. Use a clean browser profile, signed out, no extensions, and record the browser and its version.

Four. Set the location explicitly and record it. Setting a location in the search settings is not the same as being in that city, and neither is a VPN. Record which method you used, on every row, because a reader cannot judge a local result without knowing how the location was produced.

Five. Capture desktop and mobile separately. They are two observations, not one, and merging them makes a trigger rate meaningless.

Six. Handle near me queries as their own subset. A query containing near me depends on device location in a way a city name does not, so it belongs in a separate file with its own denominator.

Seven. Let the page settle before capturing, so an Overview that loads a second late is not recorded as absent. This one mistake makes the trigger rate look smaller than it is.

Eight. Record the local pack and the Overview in separate columns. They are different features with different selection systems, and an agency can be in one and absent from the other.

Nine. Fill one row per query with: query text, file version, city, location method, device, signed in state, timestamp, Overview present yes or no, number of links shown, own domain linked yes or no, agency named in the text yes or no, competitor domains linked, competitor names present, local pack present yes or no, and the screenshot filename.

Ten. Repeat each query three times in the same pass where you can, treating each as its own observation, because an Overview can differ between two searches minutes apart.

Eleven. Label twice and rerun on a fixed interval, with a second person labelling 20 rows blind and the agreement rate published, and the frozen files unchanged between passes.

The arithmetic, printed so our reporting can be checked. 30 local queries across 7 cities is 210 observations. If 128 of those 210 show an Overview, the trigger line is 128 of 210 in the named cities on the stated dates, and every Overview based metric after that uses 128 as its denominator, not 210. If 9 of the 128 link an agency domain, that is 9 of 128 and it is written that way. Merging the two gates into 9 of 210 explains nothing. This is arithmetic, not an observation.

Alongside the capture, run three checks in Search Console. Whether each page you expect to be used is indexed at all. What Google actually received, by reading the fetched page rather than your published page. And which queries already bring impressions without clicks, which is the best source for your query files. Search Console will not confirm that a particular Overview used your page, so do not build a report on a breakdown you have not verified exists today.

Named agency sample

We publish no agency names, because we have captured no Overviews. A list gathered from directories and printed beside a method reads as a result and is not one.

The sample rules that transfer: generate the list from a first pass over the frozen files rather than from your own view of the market; record each name's first appearance with the query, the city and the date; resolve every name to a working website and keep unresolvable names as rows of their own; write the rule for merging groups, networks and local offices before you label anything; and publish the sample with the file version that produced it.

For comparison, the two generated samples we do hold, with sectors named: ProVakil, CLAW and Legistify in Indian legal research software, on 28, 21 and 18 of 78 blind questions in the ChatGPT baseline of 27 July 2026; and Semrush, then Profound, then Peec, then Otterly, then Scrunch in AI visibility tooling, by frequency of mention across the aiknowsus.com capture of September 2026, counted by Perplexity over 24 batches and 72 conversations.

Results

Agency queries tested for AI Overviews by us as of 29 September 2026: zero. Proportion showing an Overview: not measured. Proportion linking an agency domain: not measured. Those cells are empty on purpose and no borrowed figure will fill them. When the capture runs, this section will carry the query count, the Overview count, the linked count, the cities, the location method and the dates in one sentence.

The counts we hold, from assistant answers rather than Overviews, with denominators and sectors.

• 0 of 6. Asked six questions about its own category with no brand named, Perplexity reported afterwards that it had not cited or recommended aiknowsus.com in any of the six answers. September 2026, AI visibility tooling.
• 1 of 18. Across eighteen blind commercial questions on 18 August 2026, ChatGPT made clawlaw.in the top source on exactly one. Indian legal research software.
• 161 across 24 batches and 72 conversations. The phrase recording that we were not cited appears 161 times in the engines' own self audits across the aiknowsus.com capture of September 2026. Perplexity.
• 0 of 6 for third party sources. Across six questions on 17 September 2026, no third party review or directory source reached any answer, and one well known review site was discarded from the raw results because its list was of American products. Claude, Indian legal research software.
• 0 of 78 searched. On 27 July 2026 ChatGPT confirmed it had run no live web search for any of the 78 blind questions, which is what a visibility reading looks like when it is produced entirely from what the model absorbed.

None of those are Overviews. They come from ChatGPT, Claude and Perplexity, which select sources differently, and they are the nearest real evidence we hold rather than a stand in.

Two dated records explain classes of absence an Overviews audit would otherwise record as bad luck. On 6 August 2026 the clawlaw.in pricing page rendered prices only after scripts ran, so a crawler received a page with no prices, and the prices the assistants quoted came from an app store listing instead. On the same date the website and the app store listing carried different plan names and prices for the same product, and ChatGPT named the contradiction in its answer. Also on 6 August 2026, ChatGPT cited clawlaw.in/blog/how-to-check-a-companys-court-cases-in-india fourteen days after that page was published, in a zone where the same question set had named the company nowhere at the 27 July baseline. Fourteen days is one observation. It is not a median, not an average and not a timetable.

Limitations, raw data, and update schedule

Seven limits.

• A snapshot is not a rate. 128 of 210 in seven cities on three days is a fact about those queries, those cities and those days.
• Triggering is unstable, so a fall between passes is not necessarily a loss of visibility.
• Location method changes the answer and the size of that effect is not published. A city setting, a VPN and a physical device in the city are three different observations.
• The displayed links are not the full source set. On 17 September 2026 Claude confirmed it had used two arguments from clawlaw.in pages and dropped the attribution in both places. A page can be used and not shown.
• Local packs and Overviews are different systems. Counting them together produces a number about neither.
• Our figures are not Overviews figures, and they come from two sectors that are not agencies.
• No audit can promise a position in an Overview. Nobody can sell you that.

Raw data available now: the three query types, the seven named cities, the eleven steps, the fifteen row fields, the arithmetic and the three Search Console checks. Not available: our own Overviews screenshots, because the capture has not happened, and the capture files behind the assistant counts, which are being prepared for publication. We restate no competitor's prices anywhere on this site; open the vendor's own pricing page and record price, date and URL together in one row.

Update schedule: reviewed when our first Overviews capture completes, when any figure here is corrected, and when the capture files are published. Version: 29 September 2026, first publication.

Common questions

Do agency searches show AI Overviews as often as other searches?

We do not know and will not guess. There is no published trigger rate by query type, and any figure you see without a query file and a city attached cannot be rerun by anybody. Count it in your own cities on a recorded date.

Is setting a location in search settings good enough for a local audit?

It is usable if you record that this is what you did, on every row. It is not the same observation as a device physically in that city, and the difference is not published, so the method column matters as much as the result column.

My competitor appears and I do not. What do I check first?

In this order. Is your page indexed, checked with URL Inspection. Is the fact a buyer needs present in the page a crawler receives, rather than added by a script after load, which was the clawlaw.in pricing failure of 6 August 2026. Then compare what the competitor's linked page states plainly that yours does not: on 17 September 2026 Claude ranked two enterprise vendors above clawlaw.in specifically because they published explicit coverage lists, and said its ordering reflected price transparency and source authority rather than product quality.

Should near me queries be in the same file as city queries?

No. They depend on device location differently and they will make your trigger rate for city queries wrong in both directions. Separate file, separate denominator, separate reporting line.

How do I record an Overview that names my agency but links a directory?

As two columns: named in the text yes, own domain linked no, with the directory domain recorded. That is a real and weaker outcome than a link to your own site, and it depends on a page you do not control.

How often should the audit be rerun?

On an interval you can actually keep, monthly for most agencies, with the query files frozen, the same cities, the same location method and the same devices. An irregular rerun with an edited query file produces trends that are artefacts of your own process rather than changes in the world.

What to do first

Write 15 local queries naming only the cities you actually serve, with your agency name in none of them, and save the file with today's date and the location method you will use. Capture all 15 today, signed out, on desktop and on mobile as separate rows, recording Overview present, links shown, own domain linked, agency named and local pack present. Then run URL Inspection on your three most important pages and read what Google actually received. That gives you a dated baseline and usually the reason for your absences, in one afternoon.