Claude web search and law firm citations: a reproducible test of what it finds and cites

As of 29 September 2026 the number of Claude answers we have tested for a citation to aiknowsus.com is zero, so that count is 0 out of 0 and no rate follows from it. The nearest Claude measurement we hold is a set of six blind questions in Indian legal research software on 17 September 2026, run for clawlaw.in, and in those six answers the number that contained any third party review or directory source was 0 of 6. In the same run Claude confirmed that clawlaw.in pages had shaped what it wrote while the site appeared only as a name in a list, and admitted it had used two specific arguments from those pages and dropped the attribution in both places. This page publishes the test, the log fields, the arithmetic and those dated findings, and it publishes no citation rate we have not measured.

Legal research software is not a law firm, and the six questions were not law firm questions. We name the sector on every figure below for that reason. What does transfer is the mechanism: Claude read the pages, used them, and did not link them, and it said why.

The answer, first

Four numbers.

• Claude answers tested for a citation to aiknowsus.com, as of 29 September 2026: 0. No runs exist, so there is no numerator and no denominator.
• 0 of 6 answers contained any third party review or directory source. Claude, 17 September 2026, six blind questions in Indian legal research software. One well known review site did appear 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.
• 2 arguments used with the attribution dropped. Same run. Claude described it as a citation lapse rather than a ranking judgement.
• 0 of 6 on the nearest measurement of our own domain. Perplexity, September 2026, six blind questions in AI visibility tooling: not cited or recommended in any of the six. A different engine, so it tells you nothing about Claude.

If you came for a share of Claude answers that cite law firms, no such figure exists here, and we have not seen one published anywhere with a query list, a repeat count and dates attached.

How it was measured

Ten steps. Steps one to six describe how the 17 September 2026 findings were produced. Steps seven to ten are parts of the method that were not run, named as such.

One. Freeze a blind question set. Six questions in the words a buyer types, about the category and not about the brand, saved with a date. For a law firm, build the set from a grid of markets and practice areas: Mumbai, Delhi, Bengaluru, Hyderabad, Chennai, Pune and Ahmedabad, against company and commercial litigation, insolvency under the Insolvency and Bankruptcy Code, employment and labour, intellectual property, real estate and RERA, and tax.

Two. Keep every brand name out of the questions. Measured reason: on 27 July 2026 two runs of the same 78 questions happened on one day in the clawlaw.in programme. The wrapper that named the brand ranked it 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 first run was discarded, because it had measured our own prompt. ChatGPT, 27 July 2026.

Three. Capture the full answer, every cited link, and the raw results where the interface exposes them. The 17 September 2026 findings exist because the raw results were captured, not only the final answers. That is how it became visible that pages had ranked and shaped the answer while not being linked.

Four. Record whether a web search ran, as a claim rather than a fact. Claude's web search is a tool that may not be invoked for a given question, so an answer with no search behind it is a different observation. Where the interface shows what it consulted, save that. Where you have to ask, mark it unverified: in the aiknowsus.com capture of September 2026 Perplexity withdrew its own statement about how many questions it had searched for, in three separate batches.

Five. Label every mention with one of five states. Named and cited to your own domain. Named and cited to a third party. Named with no support, including where the link offered does not mention you when you open it. Cited without being named. Used without attribution. The fifth state is the one that needs the raw results to detect, and it is documented in this very run.

Six. Log these thirteen fields per run: question set version, question id, exact question text, run number, timestamp, surface, model version as the interface reports it, whether search was shown as having run, country, city setting, language, device, and the answer text filename. Then one row per business named, with the name, the state, the supporting domain, and whether the destination page mentions the business.

Seven. Repeat every question five times in separate conversations. Not done on 17 September 2026, and it is the biggest gap in that run. Six single answers are six draws.

Eight. Open every citation and check that the destination supports the claim. A link that does not mention the business it was offered for is recorded as named with no support.

Nine. Label twice, with a second person on at least 20 rows blind and the agreement rate published. Not done.

Ten. Rerun the frozen set on a fixed interval. The 17 September 2026 run is a single pass and carries no trend.

The arithmetic, so the figures can be checked for shape: six questions asked once is a denominator of 6, which is what 0 of 6 means. The same six asked five times each would give 30, and the count would be reported out of 30. A firm testing 42 market and practice area pairs at five runs each would have 210 runs, with a five run denominator for every pair as well. Report the pair first and the study wide figure second. This is arithmetic, not an observation.

What the numbers were

Claude answers tested for aiknowsus.com: 0. Claude law firm queries tested: 0. Both as of 29 September 2026, and neither cell will be filled by an estimate.

The measured Claude findings, all from Indian legal research software, on two dates.

• 0 of 6 answers carried a third party review or directory source. 17 September 2026. The single well known review site in the raw results was dropped because its list was of American products.
• Two enterprise vendors ranked above the product for a stated reason. Same run, on a question about finding every case against a company. Claude placed them above clawlaw.in specifically because they publish explicit court and tribunal coverage lists, and noted that its ordering reflected price transparency and source authority rather than product quality.
• Named, not linked. Same run. The site's pages ranked in the raw results and shaped the answer, and it still landed as a name inside a list rather than as a linked recommendation, because its own comparison pages read as vendor advocacy.
• Used, attribution dropped, in two places. Same run, Claude's own description being a citation lapse rather than a ranking judgement.
• Official portals took every position above commercial products on how to questions. Same run. Claude said plainly that no commercial product should rank above the official portal for a question about using that portal.
• Correct but unsourced figures were discounted anyway. Same run. Two comparison pages stated competitors' prices with no link, no date and no source; the figures were correct when Claude checked them independently, and it treated the pages as advocacy and used official sources instead.
• An unreadable page hands the fact to somebody else. 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. Claude and ChatGPT, same date.
• A claim that could not be true cost the recommendation. 6 August 2026. A headline figure on the main site 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.
• A hidden instruction to the model was found and refused. 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.

For a law firm the second, fifth and sixth findings are the ones worth acting on. Claude ranked above a product two vendors who published explicit, named scope lists and clearer prices. It put official portals above every commercial page on questions about using those portals. And it discounted correct figures that carried no source. The firm equivalent is a page that names the courts and tribunals you appear before, the practice areas you take instructions in, your fee basis stated as clearly as your regulator allows, and a source and a date on every factual claim about anybody else.

What this cannot tell you

Eight limits.

• There is no Claude measurement of our own domain. Zero answers tested. Anything else on this page comes from another domain or another engine.
• None of it is a law firm measurement. Every Claude figure is from legal research software, a supplier to firms rather than a firm.
• Six questions asked once is six draws. 0 of 6 cannot separate never from not this time.
• Used without attribution puts a floor under every citation count. Documented above with a date. Your true influence is at least your cited count and cannot be measured exactly from outside.
• One observation is not a median or an average. The nearest timing figure we hold is one citation of a named clawlaw.in URL by ChatGPT on 6 August 2026, fourteen days after publication. Fourteen days is one observation and we will not call it a typical time.
• The app and the API are separate surfaces. A fraction from one does not transfer to the other, so the log must say which was used.
• Most counts here cannot be checked by you yet. The 6 August 2026 citation of a named URL can be. The counts over our capture folders cannot, until they are published.
• No test here promises a position in an answer. We do not sell that.

One sector limit for law firms in India: what you may publish about yourself is governed by the Bar Council of India's rules on advertising and solicitation. Read the current rules, and take advice where you are unsure, before publishing claims about results or clients. Nothing on this page is legal advice.

Sources and change log

Every Claude figure above comes from the clawlaw.in programme: GEO_BASELINE_RESULTS_2026-07-27.md for the 6 August 2026 records, and claude_response_17_09_audit.md for the 17 September 2026 records. The aiknowsus.com comparisons come from the September 2026 audit of 24 batches and 72 conversations, captured outside the site repository. Counts were produced by a script over those files rather than from memory.

We restate no Anthropic documentation on this page. Read Anthropic's help centre pages for the Claude apps, its developer documentation for the web search tool, and its published model version notes yourself, and record the URL and the date beside anything you take from them, because whether and how a search runs is exactly the sort of behaviour that changes between versions. We restate no competitor prices either: open the vendor's own pricing page and record price, date and URL together, because on 17 September 2026 correct undated prices on our pages were discounted.

Version: 29 September 2026, first publication. Updates will be triggered by the first Claude run set on aiknowsus.com, by a repetition pass that turns the denominator of 6 into 30, by any correction, and by publication of the capture files, which will make these counts checkable line by line. Corrections are noted here rather than made quietly.

Common questions

Does Claude search the web for every question?

Not necessarily. Web search is a tool that may or may not be invoked, so an answer can be produced without one, and your log has to record which happened. Do not rely on asking the assistant afterwards: in September 2026 Perplexity withdrew exactly that kind of statement in three separate batches.

Why did Claude use our pages and not cite them?

We can only tell you what it said when it did that to us. On 17 September 2026 it described dropping the attribution in two places as a citation lapse rather than a ranking judgement, and separately said the site landed as a name in a list rather than a linked recommendation because its own comparison pages read as vendor advocacy. The second reason is fixable by dating and sourcing every competitor fact and linking to the competitor's own page.

What kind of page did Claude prefer in that run?

Official portals on questions about using those portals, and vendors who published explicit coverage lists on questions about scope. It said the ordering reflected price transparency and source authority rather than product quality. That is a statement about what a page makes checkable, not about who is better.

Should a law firm publish fees to be citable?

Publish whatever your regulator allows, as specifically as it allows, with a date on it. The measured finding behind that advice is about price transparency in a software comparison on 17 September 2026, not about legal fees, and the Bar Council of India's rules decide what is permissible for you. Read them first.

Is one engine's result a guide to the others?

No. This page holds Claude findings and a Perplexity figure and they cannot be substituted for each other. Run each engine you care about, and label every number with the engine that produced it.

How do we detect being used without being cited?

Capture the raw results where the interface exposes them, not only the final answer, and read the answer for arguments and phrasings that exist on your pages and nowhere else. That is how the 17 September 2026 finding was made. It is manual work and no automated tracker we know of reports it.

What to do first

Write six blind questions from your own markets and practice areas, run each five times in separate conversations today, and capture the raw results as well as the final answers. Label every mention with the five states. Then take the two pages you most want cited and add three things to each: the named scope of what you do, a source and a date on every factual claim about anybody else, and a link to the original for each of those claims. The first job gives you a dated baseline. The second addresses the two failures Claude named out loud on 17 September 2026.