Kolkata marketing agencies: verified services, public pricing and a reproducible comparison

The count that makes this comparison reproducible is the number of named Kolkata agencies checked with original source evidence for both service area and pricing, plus the collection date. Ours is 0, collected as of 29 September 2026. We have checked no Kolkata agency, so none is named. What this page publishes is the reproducible method: the dual evidence rule that a firm counts only when both its service area and its pricing state are evidenced from original sources, the agreement check that proves the method can be run twice with the same result, and the field list so a second researcher can repeat the whole exercise and compare. Reproducibility is the point. A directory nobody can repeat is an opinion with a table around it, and we have watched what happens to that kind of page: on 17 September 2026 Claude read two comparison pages in legal technology whose competitor prices were correct and carried no link, no date and no source, and treated them as advocacy, using official sources instead.

The dual evidence rule is stricter than what most directories do, and it is strict on purpose. A firm with a verified address and an unverified price is half a row. Counting it as a full row is how a denominator becomes meaningless, and the denominator is the only thing that makes a count comparable with anybody else's.

Coverage and definitions

• Kolkata means a publicly stated office address within Kolkata Municipal Corporation limits. Areas where agency offices are commonly found include Salt Lake Sector Five, Park Street, Camac Street and New Town. Howrah and New Town addresses outside the corporation limit are recorded separately and labelled.
• Service area means the services the firm itself names on its own service pages, and the geography it says it serves. Both parts, evidenced separately.
• Pricing state means one of four: open, gated, indicative only, or none, each evidenced from an original source.
• Original source means the firm's own website page or a public filing, read on a recorded date, with the URL stored and an archived copy saved.
• Checked under the dual evidence rule means both the service area and the pricing state carry original source evidence. A firm with one and not the other is recorded as partially checked and counted in its own column.
• Collection date means the range of dates over which the rows were read, printed with every count, earliest and latest.

Collection and verification method

Nine steps, written so a second researcher can repeat them and get the same rows.

• One. Write the candidate list from named sources, recording per candidate where it came from: search results for service plus city, Google Business Profile listings, the Ministry of Corporate Affairs register, industry association member lists, and the supplier pages of local businesses. Candidate sources are never used as evidence.
• Two. Freeze the named list with its date. Nothing leaves it afterwards except a documented duplicate or a documented failure of the address rule, both logged with a reason.
• Three. Deduplicate on four fields: registered company name, directors, address, and phone or email domain. Three matches out of four means one firm with two brand names, listed as one row.
• Four. Evidence the service area from the firm's own pages, one row per service, each with URL, read date and archive reference.
• Five. Evidence the pricing state from the firm's own pages, checking a pricing page, the service pages and the footer for a rate card, in that order.
• Six. Apply the dual evidence rule and mark each firm fully checked or partially checked. Publish both counts.
• Seven. Fetch each pricing page with scripts switched off and record whether the number is present in what a crawler receives. On 6 August 2026, in legal technology, a pricing page rendered its prices only after scripts ran, so what a crawler received contained no prices at all, and the prices the assistants did quote had come from an app store listing instead of from the company's own site.
• Eight. Run the agreement check. A second person independently classifies at least five rows, or a fifth of the dataset, whichever is larger, without seeing the first classification. Publish how many rows they classified differently and what the disagreements were about. A method with no agreement check has not been shown to be reproducible at all, it has only been run once.
• Nine. Publish the rows, the rules and the disagreements together, so a third researcher can repeat it.

What to do when the two researchers disagree. This is the part that actually improves a method, so log it rather than resolving it quietly.

• Write the disagreement down as a sentence, naming the field and both readings.
• Decide which reading the written rule supports, and if the rule is silent, the rule is the problem.
• Amend the rule, date the amendment, and reclassify every affected row.
• Publish the amendment in the update log, because a reader comparing your count with an older one needs to know the definition moved.

Named agencies and evidence

Firms named: 0. Fully checked under the dual evidence rule: 0. Partially checked: 0. Collection date: 29 September 2026. No firm is named, and the counts are printed as nil so that nobody has to wonder whether the section is unfinished or the work was not done. The work was not done.

The dated evidence we hold, with sectors named, so that nothing here is read as a finding about Kolkata or about marketing agencies.

• The engine refused to name a winner on a single run. AI visibility tooling, aiknowsus.com, September 2026, Perplexity. In batch after batch it declined to name any competitor as winning most often, saying its own previous answers had not produced a comparable live tested result to support such a claim. That is the clearest statement we hold that a single pass does not establish a ranking, and it is why this page publishes a method rather than a league table.
• 0 of 6 questions in which any third party review or directory source reached an answer. Legal technology, clawlaw.in, 17 September 2026, Claude. A well known review site appeared in the raw results and was discarded, because the list it offered was of American products and so was not an answer to an India question.
• Published coverage lists outranked a better product. Legal technology, clawlaw.in, 17 September 2026, Claude. Two enterprise vendors were ranked above a smaller one specifically because they publish explicit court and tribunal coverage lists, and the assistant noted that its ordering reflected price transparency and source authority rather than product quality.
• Our own count was 0 of 6, AI visibility tooling, aiknowsus.com, September 2026, Perplexity, with 161 recorded statements across 24 batches and 72 conversations that we were not cited. We publish our own zero because a page demanding evidence from others has to show its own.

Compare by business objective

No firm is ranked, because no firm has been checked. Six objectives, and the evidence that settles each one.

• You want steady enquiries from eastern India. Ask what geography the firm says it serves, in writing, and check it against its own pages. Service area claims are the ones most often copied wrongly into directories.
• You want to be named when buyers ask an AI assistant. Ask for the frozen prompt set, check your brand appears in none of it, ask how many engines are reported and whether separately, and ask for the run count arithmetic. On 27 July 2026, two runs of the same 78 questions happened on one day for clawlaw.in. The run whose wrapper named the brand came back ranking it 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 first run was discarded.
• You want a supplier your finance team will approve. Ask for the minimum term, the notice period, the tax position and the invoicing schedule before anything else, and get all four in writing with a date.
• You want to fix a reputation problem. Ask what the firm will do about claims on your own site that cannot be verified. On 6 August 2026, in legal technology, a headline figure that could not be true led Claude to advise a buyer against the product, after which the company's accurate claims stopped counting for that answer.
• You want to run measurement in house and buy execution outside. Say so at the first meeting and ask for raw capture delivery in your own storage. A firm that resists this is selling you a report rather than a result.
• You want to spend as little as possible while still learning something. Do twenty blind prompts yourself on two engines, three runs each, and score them. That is 120 responses and one afternoon, and it gives you a dated baseline before you pay anybody.

Pricing findings and limitations

No Kolkata agency prices appear on this page, as of 29 September 2026, and we publish no range, no average and no lowest figure. We do not restate prices we have not read on a date, and we would not restate prices we had read without the date and the URL beside them. Collect your own into these nine fields.

• Firm name and registered name.
• Service, in the firm's own words.
• Pricing state, open, gated, indicative only, or none.
• Value and currency, or nil.
• Shape, fixed, starting from, band, hourly, per project, percentage of spend, or tier.
• Scope published with it, or nil.
• URL and read date.
• Archive reference.
• Checker's name, and the second checker's classification where you ran the agreement check.

Five limitations of any pricing finding from this method.

• It describes what firms publish, which is a business decision and not a measure of quality or of value for money.
• Firms that publish are not a random sample, so a range built from publishers describes publishers only.
• A published price is a starting position, and the dataset cannot see the negotiation that follows.
• Agreement between two checkers is not accuracy. It shows the definitions are clear, which is a smaller claim and the only one the check supports.
• Every row decays. Recheck quarterly and keep the previous value with its date.

Corrections and update log

Corrections. If you run a Kolkata agency and want to be checked under the dual evidence rule, or you can show that something here is wrong, write through the contact page. Corrections are logged with a date and the previous value, and a correction to one field triggers a recheck of the whole row.

Sources for every figure quoted above. The clawlaw.in programme of July to September 2026, 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, and the aiknowsus.com audit of September 2026 across 24 batches and 72 conversations, captured outside this repository. Counts were produced by a script over those files rather than from memory. The 6 August 2026 observations are externally checkable, since the pages are public. The counts over our capture files cannot be checked from outside until those files are published, and we would rather say so than imply otherwise.

Disclosure. We sell an AI visibility product, AI Knows Us, which competes for part of the same budget as an agency selling generative engine optimisation. Where tools are listed on our pages, ours is first, and the reason is that it is ours and it is built to enforce the frozen prompt set, the blind rule and per engine reporting described above. That is a design claim you can test by using it, not a benchmark result, and we hold no evidence comparing our product with any agency's service.

Update log. 29 September 2026, first publication. Method, dual evidence rule and agreement check published. Dataset empty, all counts stated as nil. Future entries name the date, the field, the previous value and the new one.

FAQs

Why insist on evidence for both service area and pricing?

Because a row with only one of them cannot be used for the decision you are making. You are choosing a firm that does a particular thing for a particular price, and half the evidence gives you half a shortlist while looking like a whole one. Count the partial rows separately and the problem is visible.

Is the agreement check worth the extra person?

It is the difference between a method and a habit. One person applying rules alone will drift, and the drift is invisible to them. Five rows independently reclassified takes twenty minutes and it either confirms your definitions or shows you which rule is vague.

What if an agency will not tell me its prices at all?

Record none, keep the firm in the denominator, and write your own scope document to quote against. You will get comparable numbers from three firms because you defined the rows, which is better than three incomparable proposals.

How is this different from reading reviews?

Reviews are other people's conclusions. This method records what the firm itself publishes, with a link and a date, which is the only evidence you can put in front of a finance team. In the one dated case we hold, on 17 September 2026 in legal technology, no third party review or directory source reached any of six answers at all.

Can I publish my dataset?

Yes, and publishing it is what makes it reproducible. Publish the rules, the rows, the read dates and the disagreements from your agreement check. Keep quotes short and attributed to the firm's own page.

Will an agency get us into AI answers?

No agency can promise that, and neither can we. The measured position we can share is our own: 0 of 6 on our own category questions in September 2026, and 161 recorded statements across 24 batches and 72 conversations that the engines had not cited us. What this work changes is the evidence that exists about you, which is the part anybody can influence.

What to do first

Name five Kolkata firms, freeze the list with today's date, and fill only two fields for each: services named on their own pages, and pricing state from their own pages. Then ask a colleague to classify the same five pricing states without seeing yours. Where you disagree, you have found a rule that needs writing more clearly, and you have done in an hour what most published directories never do at all.