Your Server Logs Are the Only Place You Can See AI Crawlers
Analytics can't see AI assistants reading a site, because they never run JavaScript. What a client's access logs showed instead, how I sort the three kinds of AI fetch, and what the log still can't tell you.

A marketing agency I work with wanted to know whether ChatGPT, Claude and Perplexity were reading a client's site, and whether any of it led to real people. Their analytics had nothing to say about it. It can't, as it turns out, so the measuring has to happen one layer down, in the web server's access log.
Analytics never sees them
Google Analytics and every tool like it works by running a script in a browser. An assistant fetching a page makes an HTTP request and reads the HTML. There's no browser, so no script runs and no event fires. As far as analytics knows the visit never happened, and no setting changes that.
The request does land in the access log though, user agent and all. So I wrote a small parser that reads the logs as they come in, sorts each request by who made it, and loads the result into the agency's warehouse next to their other data.
Three kinds of AI fetch
The vendors publish their user agents and are fairly honest about them. The useful part is that the big ones use different agents for different jobs, and those jobs mean very different things for the site.

A training crawl is a model reading the site to learn from it. A search-index crawl is the assistant building the index it answers from, which is what makes a page citable at all. A user-triggered fetch is the interesting one. Somebody asked a question right then, and the assistant went and fetched this page to answer it.
The classifier is an ordered list of patterns, and the first match wins, so the specific agents go before the generic ones:
pythonRULES = [ # OpenAI: three agents, three meanings (r"chatgpt-user", AI_ASSISTANT, "ChatGPT-User", USER_TRIGGERED), (r"oai-searchbot", AI_ASSISTANT, "OAI-SearchBot", SEARCH_INDEX), (r"gptbot", AI_ASSISTANT, "GPTBot", TRAINING_CRAWL), # Anthropic (r"claude-user", AI_ASSISTANT, "Claude-User", USER_TRIGGERED), (r"claude-searchbot", AI_ASSISTANT, "Claude-SearchBot", SEARCH_INDEX), (r"claudebot|anthropic-ai", AI_ASSISTANT, "ClaudeBot", TRAINING_CRAWL), # ... classic search, monitoring, SEO tools ... (r"wordpress/", OTHER_BOT, "WordPress (self)", None), # generic catch-all last, so a named agent always wins (r"\bbot\b|crawler|spider|scraper", OTHER_BOT, "Unnamed bot", None), ] RULESET_VERSION = 2 # bump it, re-run, and the whole archive is reclassified
Every rule was checked against the site's own logs rather than copied from someone's list. That's how the WordPress line got there. The site's own WordPress install calling itself for cron jobs is one of the busiest "visitors" in the whole log, and without its own rule it would have counted as a human. A request with no user agent at all goes in the bot pile too, since no real browser sends that.
If you just want a rough look at your own site, you don't need any of that. On an nginx or Apache log in the default format, this gives you the pages assistants fetched for real people:
bashgrep -iE 'chatgpt-user|claude-user|perplexity-user' access.log \ | awk '$9 == 200 { print $7 }' | sort | uniq -c | sort -rn | head -20
What the log can't tell you
It can't tell you how many people. The request comes from the vendor's servers, so the address in the log is a data centre. One fetch might serve one person, or get cached and serve lots. If someone promises a headcount for AI traffic, I think they're guessing.
It can't tell you whether you were cited either. Whether the answer used the page, quoted it or ignored it isn't in there.
Things that skew the numbers
The odd one I hit: Amazonbot, GPTBot, PetalBot and DuckAssistBot all appeared in the site's logs for the first time on the same day in August. My classifier hadn't changed in two weeks, so it wasn't me. Four unrelated companies don't start crawling a site on the same morning either. Something on the site or in front of it changed that day, most likely a firewall or bot-protection setting that had been turning them away before.
That's the general point. If anything sits in front of the site and blocks bots, the log only shows what got through, so a low number is a floor, not a total. Pages that need JavaScript to show their content get hit hardest, since assistants don't run it. I check what the protection layer is dropping before I read anything into a small number.
The other trap is counting the wrong rows. Sitemaps, feeds, images and 404s sit in the same log. For "pages served to an assistant", filter to successful requests for real pages, or the number looks better than it is.
Even with all that, it's the only view anyone has of this traffic, and most sites don't look at all. Which pages the assistants keep coming back to says something about what they treat as a good source, and I trust that more than a ranking position.