Marketing teams now get a monthly question: how much traffic do we get from AI? The honest answer is that some of it is visible, some is misfiled under direct or organic, and some cannot be seen at all. This article is for analytics owners, web developers and marketing leads at healthcare organisations who can read server logs or work with someone who can. It shows which signals are reliable, which are partial and what to stop claiming.
It is deliberately separate from the question of what AI says about you. For the wider effect of generative search on healthcare traffic, see AI Overviews and healthcare SEO. Here we concentrate on identification.
Three different things called "AI traffic"
Keeping these apart prevents most reporting errors.
- Crawler requests. Automated systems fetch your pages to train a model, build a search index or retrieve content. They are not visitors.
- User-initiated fetches. A person asks an assistant a question, and the assistant fetches your page in real time to answer. The request comes from the vendor's infrastructure, not from the person's browser.
- Human referrals. A person clicks a link in an AI answer and lands on your site, usually carrying a referrer or a tracking parameter.
Logs show all three. Analytics tags, which run in the browser, show only the third, and only when the browser supplies enough information.
Read the vendors' documentation first
Vendors publish user-agent names, and they have different purposes. As documented at the time of writing:
- OpenAI's bots documentation lists OAI-SearchBot, which surfaces sites in ChatGPT search features; GPTBot, which crawls content that may be used to train models; and ChatGPT-User, which handles user-initiated actions and, because a user triggers it, may not follow robots.txt rules.
- Perplexity's bots guide describes PerplexityBot, used to surface and link sites in search results, and Perplexity-User, which supports user actions and generally ignores robots.txt because a user requested the fetch.
- Anthropic's crawler help page names three bots: ClaudeBot for training content, Claude-User for user-directed retrieval and Claude-SearchBot for improving search results.
- Google says that AI features in Search use no dedicated crawler. Googlebot governs access for Search, including AI features, and Google-Extended is a separate control for some other Google AI uses.
The purposes matter for policy. Blocking a training crawler does not remove you from a vendor's search results, and blocking a search crawler can. Decide separately for each, with legal and clinical communications input, and re-read the vendor pages before you act, because these names and behaviours change.
Finding crawlers in server logs
Pull a thirty-day access log from your server, CDN or reverse proxy. If the site runs behind a CDN, use the CDN logs, since the origin may only see the CDN's addresses.
- Filter the user-agent field for the bot names above. Match on the name token, not the full string, because version text changes.
- Count requests per bot per day and the top fifty URLs per bot.
- Check response codes. A high share of 403, 404 or 5xx responses to a search bot means your content may not be retrievable.
- Compare against your sitemap. Which service and doctor pages are never fetched?
- Verify identity. User-agent strings can be faked. Where a vendor publishes IP ranges or verification methods, check that requests claiming to be that bot come from them, and treat the rest as unverified.
A hypothetical illustration: a hospital's log shows ChatGPT-User fetching one doctor-profile page repeatedly on the same afternoon. That suggests someone, or several people, asked an assistant a question that involved that doctor. It does not tell you who, what they asked or whether they later booked. It is a sign that the page is a retrieval source, which is useful for content planning.
Finding human referrals
When a person clicks a citation, the visit may arrive with a referrer such as the assistant's domain, or with a tracking parameter appended to the link. Do not assume the exact strings. Look in your own data.
In GA4:
- Open the traffic acquisition report with session source and medium.
- Search sources for the names of the assistants you care about, and for unfamiliar sources that appear after you test.
- Create a custom channel group, for example "AI assistants", with rules that match the source values you actually observe. Keep the rule list in a document and revisit it quarterly, since new products and domains appear.
- Break out landing pages. AI referrals tend to land on deep informational pages, not the homepage.
In server logs, filter the referrer header for the same domains, and inspect the query string of the landing URL for source parameters. Logs catch visits that blocked scripts hide from GA4, so compare the two counts. A large gap tells you how much your analytics is missing.
What Search Console can and cannot show
Google states that sites shown in AI Overviews and AI Mode count toward overall search traffic and appear in the Performance report under the Web search type. In other words, clicks from Google's AI features are folded into your ordinary Search Console numbers. At the time of writing, that documentation does not describe a separate breakout for them, so you cannot isolate them by filter. Check the page again before telling stakeholders this, in case Google has added a report.
You can still infer. Look at query types: long, conversational queries are more likely to trigger AI features. Track impressions and click-through for that query group against a short-keyword group over time, and annotate the dates when AI features appeared in your results. That is inference, and should be labelled as such.




