AI Mentions My Business Without Linking: A Citation and Link Audit Method

Being named in an AI answer and being linked from it are two different events, and most businesses measure them as one, which is why the problem feels mysterious. The number that describes it is the unlinked mention rate: brand mentions with no link to your own domain, divided by all brand mentions, with the count, the denominator, the assistants and the date printed together. We have not run a dedicated link audit, so we publish no general unlinked mention rate. The nearest measured fraction we hold, in legal technology in India, is 21 of 21: on 27 July 2026 ChatGPT named CLAW in 21 of 78 blind buyer answers and then confirmed it had run no live web search for any of the 78, and with no search there were no source links, so every one of those 21 mentions was unlinked. That is a fact about one memory based run on one assistant on one date, not a property of assistants in general, and we will not present it as one.

The rest of this page is the audit that produces your own figure. It is written so that two people running it separately get comparable numbers, which is the part almost every published visibility claim skips.

Define mention, citation, and clickable link

Write these five definitions into your scoring sheet before the first prompt, because the difference between two published rates is almost always the definition rather than the web.

• Mention. Your brand name appears in the words of the answer. Nothing else required.
• Citation. A source on your own domain is credited for part of the answer, whether or not it is clickable in the interface you are looking at.
• Clickable link. A reader can click through to a page on your domain from that answer, in that interface, on that device. This is the only one that can send a visit.
• Mentioned through a third party. Your name appears inside a listicle, directory or review page that the answer links to. Your name travelled, and the click goes to somebody else's page.
• Used without attribution. The answer clearly draws on your material and neither your name nor your URL appears. This state is invisible to every automated tracker we know of.

The fifth state is not theory. 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. Two arguments used, two attributions missing, in one audited answer set. If your scoring sheet has no slot for that, your audit records it as a zero and you conclude the pages achieved nothing.

Two more rules keep the scoring stable across reruns. A brand name inside a quoted block from another site is scored as mentioned through a third party, not as a mention. And a link that appears only after a follow up question is a separate observation with its own row and its own prompt text, because the first answer is what a buyer sees.

Test plan

Eight steps, in order, none of which need a tool.

One. Freeze the prompt set. Thirty prompts is a reasonable size for a link audit, written in the words buyers use, saved to a file with a version number and a date, and not edited again for the length of the study. If you add a prompt later the file gets a new version and every rate you publish names the version it came from.

Two. Apply the blind rule. Your brand name appears in none of the thirty. This is the rule that decides whether the audit means anything. On 27 July 2026 two runs of the same 78 questions happened on the same 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 altogether from the litigation due diligence questions it most wanted to win. The first run was discarded, because the only thing it had measured was our own prompt.

Three. Name the assistants individually. Three is enough and each is its own denominator. Never write all major assistants in a report, because nobody can rerun that.

Four. Fix the repeat count and print the arithmetic. Three fresh sessions per prompt per assistant. Thirty prompts times three assistants times three repeats is 270 recorded answers per round. Print that multiplication next to every rate you publish. A percentage with no visible denominator arithmetic is not auditable, and we sell a tool and still say so.

Five. Record the conditions on every row. Date and time, assistant, mode, signed in or out, location setting, language, device, browser, prompt version, prompt text.

Six. Save the evidence, not the verdict. The answer as text, a screenshot, and every source URL copied exactly as given before any cleaning. Keep the raw URL column and add a normalised column beside it. Never overwrite the raw one.

Seven. Score with the five states above, by a person, twice. Have a second person score at least twenty rows without seeing the first scores, and publish how often the two agreed. A rate with no agreement check behind it is one person's reading of thirty answers.

Eight. Record search behaviour as a claim, never as a fact. Where the interface shows what it consulted, save it. Where you have to ask, mark the answer unverified. In the aiknowsus.com audit of September 2026, Perplexity withdrew its own earlier statement when asked how many questions it had actually 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.

The denominator arithmetic, written out. The unlinked mention rate is not answers, it is mentions. Count the rows where your brand was mentioned. That is the denominator. Of those, count the rows with no clickable link to your own domain. That is the numerator. An answer that never mentioned you belongs in neither, and including it is the commonest way this rate gets published wrongly.

Results

We publish no general unlinked mention rate, because we have not run a thirty prompt link audit. Not an estimate and not a borrowed figure. What follows are five dated measurements we do hold, each with its numerator, its denominator, its engine and its sector, printed so you can see what the output of this audit looks like when it is done properly.

• 21 of 21 mentions unlinked. ChatGPT, 27 July 2026, clawlaw.in, legal technology in India. CLAW was named in 21 of 78 blind answers, and the same run confirmed no live web search had been run for any of the 78, so no source links existed for any of those mentions.
• 21 of 78 answers mentioned us at all. Same run, same date. In that memory based reading the breadth order was ProVakil on 28 questions, CLAW on 21 and Legistify on 18. That is a ranking of what the model had absorbed rather than of what was true that week.
• 1 of 18 as top source. ChatGPT, 18 August 2026, clawlaw.in. Across eighteen blind commercial questions the company was the top source on exactly one, and its own comparison page was named in the answer as still being the vendor's own editorial page.
• 0 of 6 cited or recommended. Perplexity, September 2026, aiknowsus.com, the AI visibility category. Asked six questions about its own category with no brand named, the assistant audited itself afterwards and reported it had not cited or recommended us in any of the six, so there was no position to hold.
• 161 recordings of not being cited, across 24 batches and 72 conversations. Perplexity, September 2026, aiknowsus.com. That is how often the phrase recording that we were not cited appears in the engines' own self audits of their answers across the whole capture.

The two clawlaw.in figures and the two aiknowsus.com figures are from different sectors and cannot be added together, averaged, or compared. They are presented as four separate observations for that reason.

Inspect the cited sources and site access

When you find mentions without links, work through the following six checks on the pages you expected to be linked. They are in the order that most often finds the fault.

• Fetch the page as a crawler and read the response. If the fact the answer needed is injected by a script, it is not in the document. On 6 August 2026 the clawlaw.in pricing page rendered prices only after scripts ran, so the crawler's copy had no prices, and the prices the assistants quoted came from an app store listing instead. Claude and ChatGPT, same date.
• Check crawler permission and delivery separately. A user agent allowed in robots.txt can still receive a 403 from a firewall or bot manager. Test each named crawler against each URL and record the status code.
• Check whether your page can be cited at all as served. A consent wall, a login screen, a country redirect or an app download prompt in front of the content all produce a page that says nothing about your subject.
• Read whose sources did get linked. 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 far down the list. Perplexity, 24 batches. If your subject is owned by official pages, that is the shape of the problem.
• Check whether your page reads as advocacy. On 17 September 2026 Claude confirmed that clawlaw.in pages had ranked in its raw results and had shaped what it wrote, and the site still landed as a name inside a list rather than as a linked recommendation, because its own comparison pages read as vendor advocacy.
• Check for an unverifiable claim on the page. On 6 August 2026 a headline figure on the clawlaw.in 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.

What a business can control

Four things, honestly labelled, in descending order of how much control you actually have.

What is in the bytes your server returns. Full control. Prices, coverage, dates, method, limits, all present as plain text in the response to a plain request. This is the only layer where you decide the outcome.

Whether your claims survive being checked. Full control, and usually neglected. Every number on your page needs a source, a date and a link. An undated figure is discounted even when it is correct, which is documented: on 17 September 2026 Claude checked two clawlaw.in pages whose competitor prices were correct and carried no link, no date and no source, and used official sources instead because it had no way to know at read time.

Whether a careful reader would call your page independent. Partial control. A comparison page that concludes with you every time reads as advocacy. A page that states who should buy something else is the same page with different credibility.

Whether you are linked in a particular answer. No control. Not by us, not by anybody. We do not sell a position in an AI answer and we would not believe a company that offered one.

Limitations and retest schedule

Six limits, then the schedule.

• We hold no general unlinked mention rate, as of 29 September 2026, and nothing on this page should be read as one.
• 21 of 21 is a memory based run. With zero live searches, no link was available to any answer. It tells you what an unlinked mention rate looks like in that condition, and nothing about an assistant that is searching.
• Sectors do not transfer. The clawlaw.in figures are legal technology in India. The aiknowsus.com figures are the AI visibility category. Neither is evidence about a third sector.
• Interfaces differ. The same answer can show a clickable link in one client and a plain citation in another, so the device and interface belong in every row.
• Used without attribution cannot be counted reliably. You can only catch it when the assistant admits it, as Claude did on 17 September 2026, so treat your count of that state as a floor.
• A rate is about your prompt set and your dates. It is not a property of the assistant and it does not generalise to prompts you did not run.

The retest schedule. Rerun the identical prompt set monthly, with nothing changed, for at least three rounds before you draw any conclusion about direction. Publish every round including the ones that got worse. A series with a missing month is read as a stable month, which is a lie you told by omission. If monthly is more than you will complete, run quarterly and say so on the report.

Sources and change log

Every figure above comes from 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, or from the aiknowsus.com audit of September 2026 across 24 batches and 72 conversations. Counts were produced by a script over those files rather than from memory. The 6 August 2026 citation of a named clawlaw.in URL and the pricing page reading can be checked from outside today. The capture wide counts cannot, until we publish the captures.

Version: 29 September 2026, first publication. The page gets updated when we complete a link audit and can print a real unlinked mention rate with its denominator, when any figure is corrected, and when the captures are published.

Common questions

Is an unlinked mention worth anything?

It is worth something and it is hard to bank. The reader hears your name and can search for it later. What it does not do is send a visit, so it will not appear in your analytics, and a business measuring only referrals will conclude nothing happened. Count it as its own state, report it separately, and never merge it with citations.

Why would an assistant name us and not link us?

Three causes we have measured. It was not searching at all, so no link existed, as on 27 July 2026 with 0 live searches across 78 questions. Your page read as vendor advocacy, as Claude said on 17 September 2026. Or it used your material and dropped the attribution, which Claude also admitted on that date for two separate arguments. The three have different fixes, which is why the scoring separates them.

How many prompts do I need before the rate is meaningful?

There is no size that turns a sample into a general truth. What changes with size is how much the number moves when you rerun it. Thirty prompts across three assistants with three repeats gives 270 answers, which is enough to find your problem. Publishing it as a rate still requires the denominator arithmetic printed next to it every time.

Can a monitoring tool measure this for me?

It can collect at a scale you cannot match by hand, which is a real advantage. What you must still get from it is the prompt list, the dates, the location settings, the repeat count, and whether the scoring was done by a person or a model. If a tool will not give you those five, its rate is not auditable. We are the vendor of AI Knows Us and we would rather you asked us those five questions than not.

Does adding schema markup get us linked?

Nobody can show you that it does, and we will not claim it. Markup makes your facts easier to parse. It does not oblige any system to credit you. Add it because it costs little, then measure with the audit above instead of assuming.

Our brand name is also a common word. How do we count mentions?

Write the disambiguation rule into the scoring sheet before you start, with a worked match and a worked non match, and have the second scorer apply the same rule. Then publish the rule with your rate. Most arguments about a visibility number turn out to be arguments about string matching.

What to do first

Write thirty buyer prompts with your brand in none of them, run the first ten on one assistant three times each today, and score every answer with the five states above. Then fetch the page you most expected to be linked, as a crawler, and read what came back. By this evening you will know whether you have a mention problem, a link problem, or a page that cannot be read at all.