Brand Monitoring in Claude: The Assistant Marketers Skip

Marketers watch ChatGPT because everyone watches ChatGPT. Then they check Perplexity, because it shows its sources and that feels reassuringly like the SEO they already know. Claude tends to get skipped, on the sensible-sounding grounds that it has fewer consumer users than the others.

That reasoning is fine as far as it goes, and it lands in the wrong place. User counts tell you how loud a platform is. They do not tell you who is in the room. Claude sits inside a lot of work: drafting documents, reviewing contracts, writing code, summarising a research pile, assembling the shortlist that ends up on a slide. If you sell to businesses, the question is not how many people open Claude on a Sunday. It is whether the person building a vendor list on a Tuesday afternoon is using it.

Why this is a B2B problem in particular

Consumer assistants get asked for restaurant recommendations and gift ideas. Work assistants get asked which tools a team should evaluate, which suppliers serve a region, whether a platform supports a standard. Purchase-adjacent questions, asked by people with budget.

And visibility is not evenly distributed across platforms. A brand named routinely by one assistant can be a stranger to another, because the models absorbed overlapping but not identical slices of the web and differ in how readily they search before answering. Treating your ChatGPT result as a proxy for all of them is a guess wearing the costume of a measurement.

Claude leans on what it learned during training

Every large language model is trained on a corpus that stops at some point in time. That boundary is the knowledge cutoff. Everything absorbed before it is available as internal knowledge, recallable without touching the network. Everything after it is invisible unless the model goes and looks.

Claude will answer a great many questions straight out of that internal knowledge. It has web search available and reaches for it when a question obviously demands fresh information, but "what are the leading tools for X" does not look, to a model, like it demands anything fresh. It looks like general knowledge. So the model answers from memory, and memory has a date on it.

The consequence for a brand is blunt. If your company is young, or repositioned recently, or rebranded, or was acquired and renamed, there is a real chance the model has no idea you exist. Not "ranks you poorly". Does not have you. Competitors written about for years sit in that internal knowledge with a head start, and nothing you publish this quarter changes what the model absorbed last year.

Two different failures that look identical

You ask Claude which vendors lead your category. You are not in the answer. That observation has at least two causes, and until you know which one you have, whatever you do next is guesswork.

The memory case. The model answered from training knowledge and you were not in it, or you were in it thinly, described vaguely, filed under a category you have since left. The failure happened long before this conversation started.

The retrieval case. The model searched, read a handful of pages, and built its answer from those. You were not on them. Here the failure is happening now, in public, and it is legible: some roundup is being read aloud by an assistant, and your name is not in it.

The remedies point in opposite directions, which is why the diagnosis is not academic. The retrieval case is about which pages get retrieved: which roundups exist, which rank, whether your site is fetchable and says plainly what you do. The memory case is the slower question of how much the web says about you at all, and no landing page fixes it.

A protocol you can run this week

Write the questions your buyer asks before they know you exist. Not "what is [your brand]", which proves nothing: hand a model a name and it will find something to say about it, occasionally by inventing. Ask what comes earlier. Which tools should a mid-sized company evaluate. Who supplies this in Europe. What are the alternatives to the obvious market leader. Your absence from those answers is your real position.

Ask each question several times, in fresh conversations. These models are probabilistic: the same prompt can produce a different cast of names an hour later. One answer is an anecdote. A pattern across repeated runs is the start of a measurement.

Record, for each answer, whether it searched. This is the step people skip, and it is what makes the rest interpretable. When Claude runs a web search it says so, and shows the pages it consulted. Links attached means it retrieved. Confident prose with no sources means it was almost certainly speaking from training. Two mentions are not the same finding if one came from memory and the other from a page you can read.

Force the other mode and compare. Ask again, this time telling the model explicitly to search the web first. If you appear in the searched version and vanish in the unsearched one, the live web knows you and the model's internal knowledge does not. That is a specific diagnosis: your problem is age and coverage, not a broken robots file.

Check that you are fetchable. Open yourdomain.com/robots.txt and look for ClaudeBot. A Disallow rule under it means you opted out of Anthropic's crawler, quite possibly during the scraping panic, without anyone deciding it on purpose. Whichever side you want to be on, be there deliberately.

Test the old name, if you have one. Rebranded and acquired companies routinely find the model knows them under the name they abandoned and not at all under the current one. That is a knowledge cutoff sitting in plain sight.

Run it in every language you sell in. Visibility does not travel. A brand described fluently in English can be absent from the same question in German, because the German-language web says less about it.

What this method cannot tell you

When a mention arrives without citations, you do not know where it came from. The model produced your name; it did not produce a receipt. It might be reflecting a review site, a forum thread, your documentation, or a comparison article written by someone who dislikes you. Without sources in the answer, the provenance is closed to you, and any account of why you were mentioned is a story you are telling yourself.

Asking the model to explain itself does not fix this. A model's description of its own reasoning is generated text like any other, not an audit log, and it will name influences it never consulted. Versions change too: a cutoff moves when a new one ships, so any reading you take is a snapshot.

When the manual version stops scaling

You can do this by hand once. You cannot do it across dozens of prompts, four assistants and several languages, repeatedly, and still have a job. That is where measurement has to be automated, and it is what PSentry does: it runs your prompt set across ChatGPT, Claude, Gemini and Perplexity, in each language you sell in, and reports where you are named, where you are cited, which sources came up, and who was recommended in your place.

Two limits worth stating outright. A model's internal knowledge does not refresh by the hour, so the reading does not pretend to either: it happens twice a month, matching the pace at which a cutoff-bound memory actually shifts, not the pace of a dashboard. And the only thing on offer is an honest account of what is already there. Nobody manipulates what a model outputs, no improvement is promised, and what moves your standing is the same slow accumulation of pages that moves the model's, not a subscription.

Frequently Asked Questions

Does Claude read my website when someone asks about my brand?

Only if it searches during that conversation, and often it does not. Answering from training knowledge means drawing on whatever it absorbed before its cutoff, including pages that no longer exist and descriptions you have since changed. An outdated third-party profile can outlive your own homepage.

If I block ClaudeBot, does Claude stop mentioning me?

No. Blocking the crawler affects what can be collected from your site going forward. It has no effect on what the model already learned, and none at all on what other people's pages say about you, which is a large part of what the model is repeating.

Claude described my product incorrectly. Can I correct it?

Not directly. There is no submission form, no verified profile, no correction queue. You can make the accurate version easy to find and hard to misread, and accept that the effect is indirect and slow.

Why does Claude name a competitor of our size and not us?

Size in a market and presence in a training corpus are different quantities. Models cluster brands by how much the web writes about them and how clearly, not by revenue. A smaller rival with heavy public documentation and a place in every roundup can be far more present to a model than a larger company that sells through sales calls.

Is Claude worth monitoring if my customers are consumers?

Less so, but not zero: people carry their work tools into personal decisions. The stronger reason is diagnostic. Because Claude leans harder on training knowledge, its answers read as a fairly direct report on how thoroughly the web described you before the cutoff.

My brand appears in some runs and not others. Which is true?

Both. A probabilistic system has a distribution, not one answer waiting to be revealed. State a finding as a frequency across repeated runs: a screenshot of one good answer is evidence of nothing.