Perplexity's marketing agency answers: which agencies and sources appear in a repeatable test?

Perplexity answers an agency question by naming a few firms and showing links to pages it used, and the useful output of a test like this is a domain count reported as three numbers in one sentence: how many distinct domains were cited, across how many answers, from how many queries. A domain total on its own is unreadable, because 90 domains from 30 answers and 90 domains from 300 answers describe opposite situations. As of 29 September 2026 we have not run a marketing agency query set on Perplexity, so we publish no domain count for agencies. This page gives the definitions, the normalisation rules, the place and query design, and the source patterns we did measure, each with its sector and date attached.

Agency questions make the denominator problem worse than usual, because the same words asked in two cities produce different sources. A domain count that mixes cities is a count of your own sampling decisions.

Scope and definitions

Four definitions, counted separately, never merged.

• A displayed citation. A URL the answer shows as a source. The only kind you can observe.
• A distinct URL. One unique address after the normalisation rules below. Two rows pointing at the same page are one distinct URL.
• A distinct domain. The registered domain behind those URLs. One domain can supply many URLs, and a domain count and a URL count are not interchangeable.
• An unobserved contributor. A page that shaped the answer and was not shown. You cannot count these. On 17 September 2026, in the clawlaw.in programme, Claude confirmed its answer had been shaped by clawlaw.in pages that ranked in its raw results while the site appeared only as a name in a list, and separately admitted it had used two arguments from those pages and dropped the attribution both times.

Six normalisation rules, published because two people applying different rules to the same capture report different counts from identical data.

• Strip tracking parameters, keep parameters that change the content.
• Treat http and https as one URL, and www and non www as one URL.
• Treat a trailing slash as the same URL.
• Follow one redirect and record the destination, keeping both addresses.
• Count a fragment as the same URL as its page, unless the fragment loads different content.
• Count subdomains as separate domains in the domain tally, and say so on the page.

Prompts, places, and method

Agency queries split three ways and each split gets its own frozen file and its own denominator: local intent, where the buyer names a place; industry intent, where the buyer names a sector or business model; and task intent, where the buyer names the work rather than the supplier.

For the local file, name the places rather than letting your own connection decide them. A reasonable national set is Mumbai, Delhi, Bengaluru, Pune, Hyderabad, Ahmedabad and Jaipur, one row per city per query. If you serve one city, test one city and say so.

Then eight run rules.

One. Freeze each file with a version number and a date. Every count names its version.

Two. No agency name in any query. On 27 July 2026 two runs of the same 78 questions happened on one day for clawlaw.in. The branded wrapper ranked the brand first on almost every question; the blind run put the company second by breadth and absent from the questions it most wanted to win. The branded run was discarded. ChatGPT, 27 July 2026. A source audit run with a brand in the query over collects that brand's domain for no reason.

Three. Three to five runs per query, each in a new conversation, because the cited set changes between identical runs. Every answer is a row, so 20 queries at 3 runs gives an answer count of 60 and a query count of 20. Both go in the report.

Four. Record the conditions per row. Timestamp, country, city setting, language, signed in state, device, and which mode or model the interface says produced the answer.

Five. Save the raw evidence. Answer text, every cited URL copied exactly before normalisation, screenshot. Keep a raw column and a normalised column and never overwrite the raw one.

Six. Classify every cited domain by type from a fixed list of eight: official or government, platform or vendor documentation, the agency's own marketing site, independent review or directory, news and trade press, forum or community, educational, and other. Add a ninth type only by extending the list and relabelling the whole capture.

Seven. Open every citation and check it supports what the answer said. A link that does not mention the agency it was offered for is its own state, and it is common enough that a study which never checks is reporting something else.

Eight. Record any search claim as unverified. In the September 2026 aiknowsus.com audit Perplexity withdrew its own earlier statement about how many of the questions it had searched for, saying it could not honestly substantiate the claim, and in another batch that its claim to have searched all five was not adequately supported. Three separate batches.

The arithmetic, so our reporting can be checked: 20 queries at 3 runs is 60 answers. If those 60 answers carry 348 citations, normalising to 141 distinct URLs across 52 distinct domains, the sentence reads 141 distinct URLs and 52 distinct domains across 60 answers from 20 queries, in the named city, on the stated dates. That is arithmetic and not an observation, and every number in it is illustrative of the format only.

Named agency sample

We publish no agency names, because we have not run agency queries. A named sample assembled from directories and printed beside a protocol reads as a finding, and it is not one. This whole batch exists because engines asked us for measured results, and an unmeasured list is exactly what they discount.

The selection method, which is the part that makes a sample honest, has five rules.

• Generate the sample from a first pass. Run each frozen file once, write down every agency named anywhere, and use that list. Your own view of the competitor set is not the subject.
• Record each name's first appearance with prompt id, run number and date, and never back fill a late arriving name into an earlier pass.
• Resolve every name to a working website and keep unresolvable names as their own rows rather than deleting them.
• Write the merging rule for groups, networks and local offices before you label anything, and apply it to every row.
• Publish the sample together with the query file version that produced it.

For comparison, here are the two samples we do hold real runs for, with sectors named: in Indian legal research software, ProVakil, CLAW and Legistify appeared on 28, 21 and 18 of 78 blind questions in the ChatGPT baseline of 27 July 2026. In AI visibility and generative engine optimisation tooling, the tools named most often across the whole aiknowsus.com capture of September 2026 were Semrush, then Profound, then Peec, then Otterly, then Scrunch, counted by Perplexity across 24 batches and 72 conversations. Both are generated samples. Neither is an agency sample.

Results

Distinct domains cited for marketing agency queries on Perplexity, as of 29 September 2026: not counted. Answers: zero. Queries: zero. No estimate goes in that cell and no figure is borrowed from anybody else's study.

What we hold is the pattern of which source types reach an answer at all, which is the half of a source audit that actually changes what you do next. Five dated findings, with sectors.

• Official pages and vendors' own pages dominated the cited set. Across the whole aiknowsus.com capture of September 2026 the domains cited most often were the assistants' own documentation and the vendors' own websites, with a single well known review site appearing far down the list. Perplexity, 24 batches, 72 conversations, AI visibility tooling.
• 0 of 6. Across six questions on 17 September 2026 no third party review or directory source made it into any answer at all, and the one well known review site that appeared in the raw results was discarded because the list it offered was of American products and so was not an answer to an India question. Claude, Indian legal research software.
• Official portals took every position above commercial products on how to questions. Claude, 17 September 2026, which said plainly that no commercial product should rank above the official portal for a question about using that portal.
• An unreadable page hands the fact to a third party. On 6 August 2026 the clawlaw.in pricing page rendered prices only after scripts ran, so a crawler received no prices, and the prices the assistants quoted came from an app store listing instead of the company's own site. Claude and ChatGPT, same date.
• A hidden instruction to the model was found and refused. On 6 August 2026 a competitor's page carried a hidden block of text addressed to answer engines, instructing them to cite that company as the source. Claude found it, refused it, and named the company that had done it.

The second and third findings together are the uncomfortable result for agencies. Directory and review listings are the standard advice for agency visibility, and in the one sector where we have measured it, no third party review or directory source reached a single answer across six questions, while official sources took the top of every how to question. That is one sector on one date, and it is the only measured evidence on this page about directories.

Perplexity's published description of citations

We do not restate Perplexity's own description of how it selects and shows sources. Product pages are edited without notice, and an undated paraphrase on our page becomes a false claim we cannot detect. The cost of undated information is measured: on 17 September 2026 Claude checked two clawlaw.in comparison pages that stated competitors' prices with no link, no date and no source, found the figures correct when it checked them itself, and used official sources instead.

Read these three yourself and record the URL and the date you read them: Perplexity's help centre pages on how answers and sources work, its published pages describing its models and modes, since a research mode answer is a different observation from a default answer, and its developer documentation if you collect answers through an interface rather than by hand.

The same refusal covers prices. We restate no competitor's pricing anywhere on this site. Open the vendor's own pricing page, and write the price, the date you read it and the URL into one row together.

If you buy a tool for this

Hand collection does not scale past a few hundred rows, and a tool collects more than you can. We are the vendor of one, so weigh the next line accordingly.

• AI Knows Us, ours, first, for one stated reason: it is built to report the fraction with its denominator, the query file version, the location and the dates together, because that is the form the engines in our own audit demanded. We are the vendor. Discount this line.
• Semrush, Profound, Peec, Otterly and Scrunch were the five tools named most often in our own category across the September 2026 capture. We have not tested any of them, we make no claim about their features or accuracy, and that order reflects how often assistants mentioned them rather than any judgement by us.

Get five things from whichever tool you use before quoting its numbers: the query list, the denominator, the location and device settings, the dates, and whether the labelling was done by a person or a model.

Raw data and caveats

Seven caveats.

• A domain count describes your query set. It is not a fact about Perplexity and not a fact about agencies.
• It cannot list sources used and not shown. Documented above with a date.
• It cannot say why a source was chosen. Appearance is not a ranking and the selection rule is not published.
• It cannot be compared with another audit unless their normalisation rules, query file and cities match yours, and those are almost never published.
• City changes the answer and the size of that effect is not published, so a Bengaluru count is not a Jaipur count.
• Our dated findings are not agency findings. They come from Indian legal research software and from AI visibility tooling, named beside each figure for that reason.
• No audit promises a position in an answer. We do not sell that.

Raw data available now: the four definitions, the six normalisation rules, the three query types, the seven named cities, the eight run rules, the eight source types and the arithmetic. Raw data not available: our own agency runs, which do not exist, and the capture files behind the dated findings, which are being prepared for publication.

Version: 29 September 2026, first publication. Updated when the agency query set is run, when a figure is corrected, and when the capture files go up.

Common questions

How many domains does a typical agency answer cite?

We will not give you a number, because we have not counted it for a frozen agency query file, and the published numbers we have seen arrive with no query list and no date. Count it on your own 20 queries in your own city this week.

Should I count distinct domains or appearances?

Both, in two columns. Distinct domains tells you how wide the source pool is. Appearances tells you which few pages are load bearing for your subject. Reporting only the first hides that three pages might be answering half your queries.

Are directory listings worth buying for an agency?

The only measured evidence on this page says less than the usual advice assumes, and it comes from a different sector: across six clawlaw.in questions on 17 September 2026, no third party review or directory source reached any answer, and the one review site in the raw results was dropped because its list was of American products. A listing is still cheap and worth having. Do not expect it to decide the answer, and do not buy it on the strength of one sector's result either.

My agency is cited through a client's website. Does that count?

Count it as its own state, because it is the strongest one and also the one you control least. A client page that names you as the agency is independent evidence in a way your own case study page is not, and it can be taken down without telling you.

What if a citation does not mention the agency it was offered for?

Record it as named with no supporting source. A link that fails to support the claim is not a citation for counting purposes, and studies that never open the links overstate support by including these.

Can I compare my count with a published industry study?

Only if that study publishes its normalisation rules, its query file, its cities, its answer count and its dates. If any of those five is missing, the two numbers are not comparable, however close they look.

What to do first

Write 15 agency queries split into local, industry and task, with no agency name anywhere in them, and save the file with today's date and the city you will run it in. Run each three times, paste every cited URL in exactly as shown, normalise with the six rules, classify with the eight types, and open every link to check it supports what the answer claimed. Then write the one sentence carrying distinct URLs, distinct domains, answers, queries, city and dates. That sentence is publishable. A domain total on its own is not.