Claude and AI search: what brands can actually measure
Claude shows up in AI visibility work in two ways: as a model people ask, and as a set of crawlers hitting your site. What to measure for each, and what to leave alone.
Claude gets mentioned in client calls more than it used to, usually in one of two forms. Someone on the team uses it daily and wonders whether it recommends the brand. Or a developer spots a Claude-related user agent in the logs and asks whether it's a problem.
Both are fair questions, and they need different tools. Here is how we separate them and what we measure for each.
Two different things called "Claude"
First, Claude as an answer engine: a person types a question and gets a recommendation. Second, Claude-related bots that fetch web pages: ClaudeBot, Claude-User and Claude-SearchBot all appear in the list of crawlers that Promptwatch tracks.
They aren't interchangeable. The bots tell you who visits your site. The answers tell you what gets said about you. A site can be visited often and described badly, or described well without ever being fetched because the model relies on other sources.
What the three Claude bots suggest
We're careful here, because bot behavior is documented by the vendor and changes. From the product's own bot list, we know all three names are logged. Check Anthropic's current documentation for what each one does before writing rules around it, since we haven't verified their purposes for this post.
What we can say is practical:
- Treat each bot name as a separate line in your robots.txt thinking, not a single "Claude" toggle.
- Don't allow or block on the basis of a user agent string alone, because strings can be faked. IP verification is the safer evidence.
- Decide per bot, on purpose, and write the decision down so the next person knows why.
Our post on robots.txt and CDN rules for AI crawlers shows how to verify what's reaching you.
How Claude is tracked in a visibility tool
This is a point where vendors blur things, so it's worth being clear. In Promptwatch, Claude (Haiku 5.5) is tracked via API. The live-search platforms monitored through their real interfaces are ChatGPT Search, Google AI Overviews, Google AI Mode, Copilot, Gemini, Perplexity, Alexa and Meta AI, plus the coding agents Codex and Claude Code.
That distinction matters for how you read the data. An API answer reflects the model's response to a prompt through the programmatic route. It can differ from what a person sees in a consumer app. Use API-tracked results as a useful signal about how a model talks about your category, but don't present them as "what Claude users see."
What to measure for Claude as an answer engine
Keep it modest. Run your standard prompt panel and record:
- Does the model name your brand for category questions?
- How does it describe you? Check the facts.
- Which competitors does it name instead?
- How does that compare with the other models you track?
Look at that last point especially. If Claude's answers differ sharply from ChatGPT's for the same prompts, the cause could be training data, retrieval, or both. You can't see inside either, so don't invent a theory. Note the difference and test what changes when you fix factual gaps on your own site and in third-party sources.
Claude Code and developer brands
There's a corner of this that matters for software companies. Promptwatch tracks what coding agents recommend, including Codex, Claude Code and OpenCode. If you sell a developer tool, library or API, developers increasingly ask an agent which package to use. Appearing in those recommendations is a separate goal from appearing in a chatbot answer.
For those brands, check the documentation. Agents read docs. Clear installation steps, accurate examples and a plain statement of what the tool does and doesn't do are the raw material they work from. Our post on llms.txt is worth reading too, mainly for what the file doesn't guarantee.
What not to measure
Don't count Claude mentions as a headline KPI for most brands. The volume is hard to size and we have no sourced number to give you for how many people use it. Don't compare a Claude API panel against a live ChatGPT panel as if they're the same kind of measurement. And don't read a single answer as a trend.
Reading the logs without panic
Seeing Claude-related bots in your logs isn't a crisis. It's a prompt to look at which pages they request, whether they succeed and whether those same pages are cited anywhere you track. Promptwatch's Agent Analytics shows the crawl-to-citation path with a citation rate per page, error tracking and stackable path filters, plus CSV export. If a bot fetches a page repeatedly and the page never appears in answers, the page has an access or quality problem worth investigating.
If a bot hits errors, fix those first. A fetch that returns a 403 or a timeout can't turn into a citation.
What we'd set up
Add Claude to your model list, keep it labeled as API-tracked in your reports, and connect crawler logs so the bot activity is visible. Then review it with the rest of your data, weekly, not as a special project.
Promptwatch covers both halves: the API-tracked Claude answers and the real crawler logs, with Agent Analytics on Professional and above for brands, and on every agency plan. One platform lets you check whether Claude bots visit and whether the model says anything useful, which is a better position than guessing from a single log line.
Keep the claims small and the records dated. That's what keeps this useful when someone asks you in six months what changed.