How Gemini answers 'best marketing agency' searches: a repeatable local and industry prompt study

There is no fixed list of agencies Gemini recommends. The same question asked twice can return different names, and an agency can be named in an answer with no source behind it at all, which is worth almost nothing to the reader. So the measurement worth publishing is one fraction per prompt type: out of the repeated runs of that prompt, how many named an agency and also supplied a citation that supports the agency named. As of 29 September 2026 we have not run a marketing agency prompt set on Gemini, so this page publishes no agency share. It publishes the full protocol, the prompt type split, the denominator arithmetic, and the nearest real fractions we hold, which come from Indian legal research software and from AI visibility tooling rather than from agencies.

The separation by prompt type is not a refinement, it is the finding. A local prompt and an industry prompt behave differently enough that one number covering both describes neither, and the person paying for the study usually cares about only one of them.

What recommended agency means in this study

An agency name in an answer is not one outcome, it is four, and they have different value to the agency and different fixes when they go wrong.

• Named with no source. The name appears and nothing supports it. The reader has a name they cannot check and no link to click.
• Named and sourced to a third party listing. The supporting link goes to a directory or a listicle that includes the agency. Useful, and it is a page the agency does not control and can be dropped from without notice.
• Named and sourced to the agency's own site. The supporting link is on the agency's domain. This is the state most agencies think they are measuring.
• Named and sourced to an independent, checkable page. A client's own site, a trade publication, an award body, a case record. The strongest state, and the hardest to manufacture.

Two more labelling rules stop the count drifting over hundreds of rows. A named agency whose supporting link is broken, or goes to a page that does not mention the agency, is recorded as named with no source, not as sourced. And an agency named only in a follow up answer, after the reader asked a second question, is a separate observation with its own prompt text.

The fifth state you cannot see is worth stating because it is documented. 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. Material can be used without the source appearing anywhere. No prompt study will catch that, and a study which claims to have measured everything is already wrong.

Test design

Three prompt types, kept in three separate frozen files with three separate denominators.

• Local intent. The buyer names a place. Test these across named cities so the place is a variable rather than an accident: Mumbai, Delhi, Bengaluru, Pune, Hyderabad, Ahmedabad and Jaipur. One row per city per prompt.
• Industry intent. The buyer names a sector or a business model rather than a place, for example a B2B software company, a manufacturer selling to distributors, a hospital group, or a company that sells only through retail.
• Task intent. The buyer names the work, not the supplier: who can run performance marketing, who can fix organic traffic that has fallen, who can produce a year of content in one language and then a second.

Then the run rules. Nine of them.

One. Freeze each prompt file with a version number and a date, and never edit it mid study. If you add a prompt, the version changes and every fraction states its version.

Two. No agency name and no client name inside any prompt. This is the rule with measured consequences. On 27 July 2026 two runs of the same 78 questions happened on the same day in the clawlaw.in programme. The first 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. The first was discarded, because the only thing it had measured was our own prompt. ChatGPT, 27 July 2026.

Three. Five independent runs per prompt as a floor, decided before you see any result. A single run cannot distinguish an agency that appears every time from one that appeared once.

Four. A fresh conversation for every run, with memory and personalisation off and recorded as off, and runs spread across the day rather than fired within a minute.

Five. Fix and record the conditions on every row. Surface used, model version as reported by the interface, signed in state, country, city setting, language, device, timestamp.

Six. Save the answer text and every link exactly as given, plus a screenshot. Also open every supporting link and record whether the destination page actually mentions the agency, because a citation that does not support the claim is a different state.

Seven. Log any search claim as a claim. In the aiknowsus.com audit of September 2026 Perplexity withdrew its own earlier statement about how many questions it had 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. Treat any assistant's account of its own work the same way.

Eight. Label twice. A second person labels at least 20 rows blind using the four states above, and you publish how often you agreed.

Nine. Rerun each frozen file on a fixed interval, reporting each pass as its own fraction rather than as a growth percentage.

The denominator arithmetic by prompt type

Worked through so that any figure we publish can be checked for shape. Eight local prompts across seven cities at five runs each is 280 runs, and that is the local denominator. Six industry prompts at five runs is 30 runs. Six task prompts at five runs is 30 runs. Those three numbers never get added together, because a study wide fraction over 340 runs is dominated by whichever type you happened to test most.

Inside the local set, city is its own breakdown: 8 prompts at 5 runs in Pune is 40 runs, and the Pune fraction is reported separately from the Mumbai fraction. An agency that is named and cited in 12 of 40 Pune runs and 0 of 40 Jaipur runs is a real finding. The same agency reported as 12 of 280 across all cities looks like noise. This is arithmetic, not an observation.

Named agency sample and selection method

We publish no list of agency names, and the reason is the point of the page. We have not run this study, so any named sample here would be a list we assembled from directories and presented next to a protocol, which reads as a finding and is not one. The engines in our own audit asked for measured results. An unmeasured list dressed as a sample is the thing they discount.

What we can give you is the selection method, which is the part that determines whether your sample is honest.

• Derive the sample from the runs, not from your own opinion. Run each frozen file once, write down every agency named anywhere in any answer, and that becomes your tested sample. Adding an agency because you consider it a competitor measures your view of the market rather than the assistant's.
• Record the first appearance of every new name, with the prompt id, run number and date. Names that enter late in a study are interesting and must not be back filled into earlier passes.
• Keep a separate list of names that appear and cannot be verified as real agencies. Check each one against a working website. A name you cannot resolve is its own row and its own finding, not a deletion.
• Do not merge a group and its subsidiaries, or a network and a local office, unless you write the merging rule down first and apply it to every row.
• Publish the sample with the prompt file version it came from, so a reader can see that the sample was generated rather than chosen.

One name pattern to handle before you start: many agency names are common words or initials. Write the string matching rule into the labelling sheet with an example of a match and a non match, and have the second labeller apply the identical rule. Most disputes about a visibility number turn out to be disputes about string matching.

Results

Marketing agency prompt runs completed on Gemini as of 29 September 2026: zero. There is no local fraction, no industry fraction, no task fraction, no estimate and no range on this page. The cells stay empty until the protocol above has been run and the logs published.

The nearest real fractions we hold are printed below with the sector each one belongs to, so nobody reads a legal technology result as an agency result.

• 0 of 6. Asked six questions about its own category with no brand named, Perplexity audited itself afterwards and reported it had not cited or recommended aiknowsus.com in any of the six answers. September 2026, AI visibility and generative engine optimisation tooling.
• 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. Indian legal research software.
• 28, 21 and 18 out of 78. Blind breadth order on 27 July 2026, ChatGPT: ProVakil on 28 questions, CLAW on 21, Legistify on 18. Same sector.
• 0 of 78 searched. Same run and date. ChatGPT confirmed it had run no live web search for any of the 78 questions, so those breadth counts describe what the model had absorbed rather than the web that week.
• 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.

Two things in that list are directly relevant to an agency, even though none of it is agency data. The 0 of 78 line shows that a visibility reading can be produced with no live search at all, in which case it is a measurement of what the model absorbed rather than of who is good this quarter. And the 1 of 18 line shows what happens to a business whose strongest page is its own comparison page: the engine named it and labelled it as the vendor's own editorial page in the same breath.

One more finding, from the engine rather than from us. In batch after batch of the September 2026 capture Perplexity declined to name a competitor as winning most often, saying its own previous answers had not produced a comparable live tested result to support such a claim. If an assistant will not rank agencies from a single pass, a page that does so from a single pass is not more rigorous than the assistant.

Google documentation and limitations

We do not restate Google's documentation here. Help pages are edited, and an undated paraphrase becomes a false claim that the page cannot notice. We have watched correct but undated information get discounted: on 17 September 2026 Claude checked two clawlaw.in comparison pages whose competitor prices had no link, no date and no source, found the figures correct on independent checking, and used official sources instead.

Read these three on the day you read this, and record the URL and the date beside anything you take from them: Google's Gemini Apps help pages, Google's developer documentation for the Gemini models if you are testing through an interface rather than the app, and Google Business Profile's own help pages if any part of your study touches local results, which are a separate system from the assistant.

Six limits on the study itself.

• No Gemini agency data exists behind this page. Everything printed above came from ChatGPT, Claude or Perplexity in two other sectors.
• A fraction of runs is not a probability. 3 of 5 on one prompt in one city on one day is five draws.
• Local answers are tied to the location you set, and the size of that effect is not published, so a Pune result is not a Nagpur result.
• The citation you can see is not the full source set. Documented above with a date.
• A change after a change is not a cause. The engine can change its own behaviour in the same month you publish pages.
• No study here can promise you a position in an answer. We do not sell that and would not believe anybody who did.

Reproduction materials

Available to you today, without us: the four named and sourced states, the three prompt types, the seven named cities for the local set, the nine run rules, the denominator arithmetic, and the labelling rules for names and mergers. That is the whole study apart from the running of it.

Not available: our own Gemini logs, because they do not exist, and the capture files behind the dated fractions above, which are being prepared for publication. Of the observations on this page, the citation of a named clawlaw.in URL by ChatGPT on 6 August 2026 can be checked from outside today. The counts over our capture folders cannot be, until the captures go up.

Version: 29 September 2026, first publication. The page updates when the first Gemini agency set is run, when any figure is corrected, and when the capture files are published.

Common questions

Why separate local prompts from industry prompts at all?

Because they are answered from different sources and a merged number hides which one you are losing. An agency can be named and cited on every industry prompt and invisible on every local one, and the fix for those two situations is not the same. Report them as two fractions with two denominators, always.

How many cities should a local study cover?

As many as you sell in, and no more. Testing eleven cities you do not serve produces a bigger table and no decision. The seven named above are a starting set for a national study. If you serve one city, run one city and report one denominator.

Is being named without a source any use to an agency?

Some. The reader hears the name and can search for it later. It sends no visit, so it will not appear in your analytics, and an agency measuring only referrals will conclude nothing happened. Keep it as its own state and report it separately rather than adding it to your cited count.

What should an agency publish if it wants to be citable?

Facts an answer can quote without needing to trust you. The nearest measured evidence we have for what that means came from a different sector: on 17 September 2026 Claude ranked two enterprise vendors above clawlaw.in specifically because they publish explicit court and tribunal coverage lists, and said its ordering reflected price transparency and source authority rather than product quality. For an agency the equivalent is a named list of services, the named sectors you have worked in, named channels and platforms with certification dates, and prices or ranges with a date on them.

Our results pages have client numbers on them. Is that enough?

Only if each number has a date, a source and a way for a reader to check it. On 6 August 2026 a headline figure on clawlaw.in could not be true, and when Claude checked it against public numbers it advised a buyer against the product, after which the company's accurate claims stopped counting for that answer. One unverifiable number can cost you the ones that were true.

Can I run this with a tool instead?

Yes, and get five things from the tool before quoting it: the prompt list, the denominator, the location and device settings, the dates, and whether the labelling was done by a person or a model. We sell a tool and we would still rather you asked us those five questions.

What to do first

Write six local prompts naming only the cities you actually sell in, and six industry prompts naming only sectors you have delivered work in, with your agency name in none of them. Save both files with today's date. Run each prompt five times in separate conversations, log the conditions, and record for every agency named which of the four states applies. Then open every supporting link and check the destination really mentions the agency. You will finish with two dated fractions that belong to you, which is more than any published agency ranking can offer.