If you want to monitor Claude mentions, you are really trying to answer two questions that sound alike but are not the same. The first is whether Claude names your brand when buyers ask it for help. The second is whether any of those answers sent people to your site. This guide walks you through both, in order, so you end up with a repeatable way to check your Claude brand visibility and a clear way to track Claude traffic in your analytics.
You will need access to Claude, a Google Analytics 4 (GA4) property for your site, and a shared sheet or doc where you can save results. Nothing here needs a developer, though someone with Editor access in GA4 will help in the last steps.
A quick note before the steps. A mention in an answer is not proof that someone clicked, and a visit in your analytics is not a full count of how much Claude influenced a buyer. These are two separate measurements, and the steps below keep them separate on purpose.
Step 1: Define what counts as a Claude mention
Start by writing down what you are looking for, because it is very easy to count the wrong thing later. List your brand's exact name, any shorthand people use, your product names, and any words that could be confused with another company.
Then set two rules and keep them apart:
- A mention is an explicit brand or product name in Claude's answer.
- A citation is a link in the answer to a page on your domain.
An answer can mention you without linking to you, cite one of your pages without really recommending you, do both, or do neither. Those are four different results, and each one tells you something different about your Claude brand visibility.
You are done when your team has one rule for spotting a mention and a separate rule for spotting a citation.
Common mistake: treating "Claude named us" and "Claude cited our website" as the same outcome. If you blend them, you will not be able to tell whether you have an awareness gap or a sourcing gap.
Step 2: Build a small, stable set of buyer prompts
Next, pick the questions you will ask. Use the kinds of things a buyer would really type, rather than only asking Claude about your company by name. A mix works well: some discovery questions, some category comparisons, and some problem-solving questions. If you want a deeper look at how to sort them, the four types of prompts worth tracking is a good place to start.
Here are a few shapes you can adapt:
- "What does [brand] offer, and who is it for?"
- "What are good [category] options for a [target buyer]?"
- "How can a [target buyer] solve [problem]?"
Keep the wording fixed from one round of checks to the next. For every prompt, save the exact text, the date and time, the Claude model or experience if it is shown, whether web search was on, and any location or account detail that could matter. There is no magic number of prompts or magic schedule here. Pick a number you can actually keep up with, write it down, and stick to it.
You are done when each question maps to a buyer intent and someone else on your team could rerun the same set without asking you what you meant.
Common mistake: changing the wording between checks and then blaming the difference in answers on a change in visibility. Small wording changes can change what Claude returns, so you would be measuring your own edits.
Step 3: Run comparable Claude checks and save the evidence
Now run each prompt in Claude and record what comes back. This is the core of how you monitor Claude mentions by hand.
When you want to test web-grounded answers, ask Claude to use web search and check that the feature is switched on where the interface shows a toggle. Claude's help guidance explains that it searches when it decides current web information would be useful, so a prompt typed into Claude is not automatically a web search. The details of enabling and using web search are worth a read, because the interface can differ by plan. On Team or Enterprise, an Owner or Primary Owner may need to turn on web search for the workspace under Organization settings, then members can turn it on for a chat from the plus menu. The newer Claude experience may not show a web search toggle at all, and availability can also change with connectivity. Claude may also use a location inferred from your IP address for localized results, so note where you ran the check from.
For each answer, record these things:
- Whether your brand was named, and where in the answer it appeared.
- Whether your domain was cited, and which page.
- The full answer text and the source links, saved with the run.
- Whether web search was used, and the test conditions.
Keep the whole answer and its citations, not only a yes or no. When a result changes later, the saved answer is how you work out why. It also helps to open the pages Claude cites and check that they say what Claude says they say, especially for results you plan to report to someone else.
For a cleaner repeat check, you can use an incognito chat outside of a project. Claude says incognito chats are not saved to chat history or memory and do not use existing memory, and that they are available on all plan types, though not inside projects. That helps reduce memory carryover. It does not make every other variable identical, so keep writing down your conditions anyway.
You are done when each run has the exact answer, its citations, the search setting, and the conditions saved in one place.
Pro tip: do your checks in one sitting, using the same account and the same settings, so you are not mixing in differences that have nothing to do with your brand.
Step 4: Turn repeated checks into a visibility trend
One answer is a snapshot. It is not a trend, and it is easy to read too much into it. The trend comes from running the same prompt set again and again, comparing like with like: the same prompt version, a comparable Claude experience, and the same definitions of a mention and a citation.
For each prompt and each run, your record should hold:
- The exact prompt and its version.
- The date.
- The answer.
- Mention status and position.
- Citation status and the cited page.
- Notes about any search or model changes.
If you do this by hand, a spreadsheet is enough to start. The work is mostly the discipline of repeating it on the same schedule. Many teams find that the manual version gets tiring after a few weeks, which is a normal point to start thinking about a tool. If you are comparing options, a roundup of AI visibility tools is a sensible next read.
DeepSmith's AI visibility module handles the scheduled part of this for you when Claude is enabled for your workspace. Your tracked prompts run on a schedule, the answers are captured, and mention rate and citation rate are tracked separately. The prompts view keeps per-prompt rates and the full answer history, and Discover Prompts can suggest candidates from your product and persona context. Claude tracking is part of the Enterprise and Custom plans, which cover all ten engines, so it is not included in every plan. Check what is enabled in your workspace before you plan around it.
It helps to be clear on what this does and does not tell you. DeepSmith measures Claude visibility. It does not tell you that a cited answer led to a visit, because that is a question for your web analytics, which is what the next steps cover.
You are done when you can compare prompt-level mention and citation results across runs and open the answer behind any change.
Common mistake: reporting one manual answer as if it were a stable result. If a number is going into a slide, it should come from repeated runs.
Step 5: Confirm GA4 is collecting usable website data
Before you go looking for Claude traffic, make sure your analytics is working at all. A report that looks empty is sometimes just a tag that is not firing.
- Confirm the Google tag is firing on your site and that GA4 is receiving events. Tag Assistant or GA4 DebugView can show you the events on your own device.
- If you use DebugView, turn on debug mode for your test device, since Google recommends this so your device can be identified. Then check each event and its parameters.
- Open Realtime and confirm your test activity starts to appear within a few minutes.
- Check your consent setup. If required consent has not been granted, or a privacy control on the browser blocks things, events may not show up.
- Filter out developer traffic so your own tests do not distort the regular reports.
Timing matters here, so it helps to know what to expect. DebugView events can appear within seconds, Realtime within minutes, and standard reports can take 24 to 48 hours to process fully. DebugView and Realtime give you only limited attribution analysis, so when you want an accurate view of where traffic came from, use the Acquisition reports.
You are done when your test page views and key events show up, you understand how your site's consent setup affects what is collected, and developer traffic is excluded or at least accounted for.
Common mistake: checking a standard report an hour after a test and deciding nothing is being tracked. Give it the time it needs, and use Realtime for the quick check.
Step 6: Find Claude-attributed sessions in Traffic acquisition
Now you can look for Claude traffic. In GA4, go to Reports, then Acquisition, then Traffic acquisition. This is the main way to track Claude traffic once you know your data is clean. Use the session-scoped dimension called Session source / medium, and search or filter for claude.ai / referral.
That label comes from a third-party analytics explainer, which documents source claude.ai and medium referral. Treat it as a good value to look for, and not as a promise that every visit that started in Claude will carry it. Look at the actual source and medium values in your own property. If you see another Claude hostname, check it before you add it to any rule.
The scope matters, so here is how to think about it. Use the session-scoped view to answer "where did this visit come from?" The first-user dimensions answer a different question, which is where a person was first acquired. Event-scoped dimensions are for event attribution, and they are not interchangeable with session source. The Traffic acquisition report covers new and returning users, and a session ends after 30 minutes of inactivity by default. If someone arrives from a new source in the middle of an existing session, GA4 does not start a new session for it.
You are done when you can pick out the Claude-attributed sessions in the source and medium report and explain why you are using the session view, not first-user or event attribution.
Common mistake: mixing up the scopes, for example by reading a first-user source as if it were the source of this particular visit.
Step 7: Inspect landing pages, engagement and outcomes
Knowing that Claude sessions exist is useful, and knowing which pages they land on is more useful. In GA4, open Reports, then Engagement, then Landing page. Add Session source / medium as a secondary dimension, and you will see the sources for each landing page.
You can also work the other way. In Traffic acquisition, use Session source / medium as the primary dimension, then add Landing page + query string as the secondary one, which shows landing pages by source. The landing page report measures the page path and query string tied to the first page view in a session. If the report is missing from your left navigation, it may have been removed from the default view, and someone with Editor access or higher can add it back.
Then look past the visit count. Sessions on their own are not success. The Traffic acquisition report also includes engaged sessions and engagement rate, and if you already have events set up, you can look at key events and revenue for these sessions too. To mark an existing event as a key event, an Editor or higher goes to Admin, then Events, and selects the star beside the event. A newly marked key event can take up to 24 hours to show in standard reports, and marking it does not change your historic data.
There is one limit to know about with UTM tags. GA4 leaves UTM parameters out of the Landing page + query string and Page path + query string dimensions, and puts them in Page location instead. So if you do not see a UTM in those landing page views, that alone does not mean your tagging failed.
You are done when the report shows which session source led to which first page, and you can look at engagement or configured outcomes for those sessions.
Common mistake: celebrating a spike in Claude visits without checking whether those people did anything on the page.
Step 8: Separate observed Claude referrals from direct and unknown traffic
This step is where careful reporting pays off. In GA4, (direct) / (none) means GA4 has no clear referral source for that traffic. It can be people who typed a URL or followed an offline link, and it can also include visits whose source information was lost along the way. Missing campaign tags, redirects that strip parameters, URL shorteners, and referrer suppression can all get in the way. GA4's own guide to direct and none traffic also notes that a direct visit that comes after a referred visit will not override the existing referrer.
So it is reasonable to suspect that some Claude-origin visits end up unlabeled. It is not reasonable to move the whole direct row over to Claude. Do not relabel all direct traffic as Claude traffic, and do not claim that a zero in the Claude referral row proves Claude had no influence. Report what you can observe, and call the rest unknown.
A note on UTM tags
You cannot reliably add a UTM tag to an organic Claude citation, because you do not control that link. UTMs are for links you do control, such as links in your own emails or campaign materials. Google recommends utm_source, utm_medium and utm_campaign, and the values are case-sensitive, so pick consistent lowercase names. Any redirect has to keep the parameters in place. A UTM tag cannot bring back a referrer that was never sent, and a tagged test link you click yourself is not an organic Claude referral.
Create a Claude channel group
If it helps your reporting, make a custom channel group so Claude shows up on its own line. Go to Admin, then Data display, then Channel groups, and create a group from the default one. Add a channel and set a Source condition that matches a regular expression. If the source you observed is exactly claude.ai, then ^claude\.ai$ is a narrow rule that is a fair place to start. Check the source value in your report first, and add other hostnames only if your own data shows them.
Put the Claude channel above Referrals, since GA4 assigns traffic to the first channel whose rule matches. Custom channel groups can be used in reports and applied back to existing report data. They only sort what was collected, so they cannot repair a source that was never recorded. An Editor or higher is needed to create or edit them. Standard properties support two custom groups on top of the predefined one, and each group can hold up to 50 channels.
One shortcut to skip: do not add Claude to GA4's unwanted referrals list. That setting tells Analytics to ignore a matching referrer, which hides the source and does the opposite of what you want here. It is meant for self-referrals and other noise.
You are done when Claude-attributed visits can be told apart from the default Referral and Direct channels, and your team reports sessions with no clear source as unknown instead of assigning them to Claude.
Common mistake: treating server logs from Claude's bots as human visits. Anthropic describes ClaudeBot as collecting content that may contribute to model training, Claude-User as web access that supports a user's request, and Claude-SearchBot as improving search result quality. Requests from these are not proof that a person visited, and they are not referral sessions.
Read the two measurements side by side
Once both sides are running, put them next to each other, and keep them next to each other without joining them. Your prompt checks tell you whether Claude names you and cites you. Your GA4 reports tell you what recorded sessions came from Claude and what they did. Neither one proves the other, and there is no reliable public figure for what share of Claude clicks GA4 will identify, so avoid inventing a ratio between them.
DeepSmith does not replace GA4 for web sessions, landing pages or conversions, and the two systems answer different questions. Your page-level citation view can show which of your pages were cited and which prompts drove those citations. Your GA4 landing page report can show whether any of those same pages received visits attributed to Claude. Looking at the two together is a good way to form a question, such as "Claude cites this page often, so is it getting visits?" Treat the answer as a lead to follow, not as a joined attribution report.
If Claude is only one of several engines you watch, each engine has its own habits and its own guide. Perplexity has one on how to track brand visibility in Perplexity, and there is a comparison of Perplexity share of voice tools. Teams that watch ChatGPT can start with the roundup of ChatGPT visibility tools. Gemini has a similar roundup of Gemini tracking tools. Copilot has one for Microsoft Copilot tools. For Google's own surfaces, see how AI Overviews and AI Mode differ for tracking, and read how AI Overviews tracking works. The roundup of AI Overviews tracking tools is the hub for that group. Those are different jobs from this one, and this guide stays on Claude.
What to do next
Save your prompt set, your definitions of a mention and a citation, and a copy of your Source / medium and Landing page reports. Then repeat the same checks on the same schedule without changing the rules between rounds. After a month or two you will have a much better sense of where Claude names you, where it cites you, and which pages are picking up visits.
If you would rather not run the prompt checks by hand, you can start a DeepSmith free trial and set up scheduled prompt tracking, then keep GA4 as the place you go for sessions and outcomes.
Related guides
- AI Brand Monitoring vs Social Listening: What Each One Covers: See what AI brand monitoring and social listening each cover, such as AI answers versus public posts, where their data and goals overlap, and when a team needs.
- AI Reputation Management: What It Is and Who Should Own It: Define AI reputation management and decide which team should own it.
- Featured Snippets and AI Overviews: How to Analyze Your Performance in Both: Analyze both surfaces in one workflow: list queries where you hold a snippet or overview citation, compare impressions and clicks in Search Console, and note where each wins.
- Buying an AEO Platform: Contract Terms, Onboarding and Multi-Year Commitments: Review the contract terms that matter when buying an AEO platform: renewal and exit clauses, data ownership, onboarding commitments and multi-year pricing, so you can negotiate with a.
- Prompt Volume: Can Anyone Know How Often People Ask AI a Question?: Learn why no one has true prompt volume for AI engines, what estimates exist and how to prioritize without it.


