Over seven days at the end of June 2026, AI bots visited this website 604 times. Not Googlebot. Not Bingbot. AI systems: OpenAI's crawlers, Anthropic's ClaudeBot, PerplexityBot, and their siblings.
That number surprised us, because getgeofix.com is exactly the kind of site that "shouldn't matter" to AI companies yet: small, young, a modest number of pages, no big brand, no huge backlink profile. And still — 604 requests in a week works out to an AI system asking for one of our pages roughly every 17 minutes, around the clock.
The more interesting part is hidden inside that total. About one visit in five was not a crawler stockpiling pages for later. It was an AI assistant opening a page live, in the middle of a conversation, because a real person had just asked it something. Below is the full breakdown, the exact method we used to count, and how to see the same picture for your own site.
Here is every AI user-agent that requested pages from getgeofix.com between 26 June and 2 July 2026, counted at the CDN level (method described below):
User-agent
Operator
What it does
Visits
Share
ClaudeBot
Anthropic
Crawls pages to train and update Claude models
179
29.6%
OAI-SearchBot
OpenAI
Builds the search index behind ChatGPT search
159
26.3%
GPTBot
OpenAI
Collects content for training OpenAI models
107
17.7%
ChatGPT-User
OpenAI
Fetches a page live when a ChatGPT user needs it
99
16.4%
PerplexityBot
Perplexity
Indexes pages for the Perplexity answer engine
36
6.0%
Claude-User
Anthropic
Fetches a page live for a Claude user
15
2.5%
Perplexity-User
Perplexity
Fetches a page live for a Perplexity user
4
0.7%
meta-externalagent
Meta
Crawls content for Meta AI models
3
0.5%
Amazonbot
Amazon
Crawls pages for Alexa and Amazon AI answers
2
0.3%
A few things stand out. ClaudeBot was the single most active visitor — nearly a third of all AI requests. But if you group by company, OpenAI dominates: its three user-agents (OAI-SearchBot, GPTBot, ChatGPT-User) add up to 365 requests, about 60% of the total. Anthropic accounts for roughly 32%, Perplexity for under 7%, and Meta and Amazon barely showed up.
The daily rhythm was anything but flat:
Day
AI bot requests
Friday, 26 June
182
Saturday, 27 June
53
Sunday, 28 June
127
Monday, 29 June
75
Tuesday, 30 June
54
Wednesday, 1 July
62
Thursday, 2 July
51
The spike on Friday 26 June (182 requests) and the second wave on Sunday 28 June (127) were mostly index and training crawlers working through the site in batches — that is how these bots typically operate. But even the quietest day still logged 51 AI requests. On a site this size, that is not an anomaly. That is the new baseline.
Training crawlers vs live reads: two different things are happening
Lumping all of this together as "bot traffic" misses the story. The nine user-agents in the table do two fundamentally different jobs.
Most of the volume — GPTBot, ClaudeBot, OAI-SearchBot, PerplexityBot, meta-externalagent, Amazonbot — is crawling in the classic sense: reading pages on the operator's own schedule to train future models or to keep a search index fresh. Useful, but speculative from your point of view. Your page goes into a warehouse; maybe it gets used, maybe not.
The other kind is different. ChatGPT-User, Claude-User, and Perplexity-User are user-triggered agents: they fetch a page at the exact moment a person asks the assistant a question that the page might answer, or clicks through to a source. Nobody schedules these visits. Each one is an AI assistant reading your site on behalf of a live human.
On their own schedule, building datasets and search indexes for later
The moment a real person asks the assistant something
What one visit means
Your page may enter a model or a search index at some point
An assistant is reading your page for a live conversation right now
Share of our 604 visits
486 requests (about 80%)
118 requests (about 20%)
To be clear about what we can and cannot see: we observe the fetch, not the conversation. We do not know what the person asked, and a fetch does not tell us how the assistant used the page — whether it was quoted, summarised, or discarded. But the direction is unambiguous. Assistants are not just archiving small sites for some future model. They are actively pulling pages into live answers, today.
Why this matters for small websites
Most published AI-crawler statistics come from large publishers and CDN-wide aggregates — sites with millions of pages. The quiet assumption is that AI reading is a big-site phenomenon, and small business sites can safely ignore it for a few more years.
Our week of data points the other way. A small, young site with no special authority received 604 AI requests in seven days, from three major AI companies plus Meta and Amazon — and a fifth of those requests were tied to live user sessions. If AI systems read a site like this one every 17 minutes, they are almost certainly reading yours too. The difference is only whether you have looked.
One honest caveat: visits are the beginning of the pipeline, not the end. An AI assistant fetching your page does not mean your business gets named in answers — that depends on what the bot finds when it arrives: whether the page is readable without JavaScript, whether key facts are stated in plain text, whether robots.txt is quietly blocking half the crawlers while allowing the rest. Access is the precondition you control.
Want the same picture for your own site?
The free Express Check shows what AI bots can read on your pages — takes about two minutes.
This section exists so the numbers above can be checked, reproduced, or cited. Nothing here requires paid tooling.
Source: Cloudflare zone analytics for getgeofix.com, queried through the Cloudflare GraphQL API (the httpRequestsAdaptiveGroups dataset), grouped by user-agent and by day.
Window: seven full days, 26 June to 2 July 2026 (UTC).
Definition of a "visit": one HTTP request whose User-Agent string contains one of the tracked AI bot identifiers. Cached responses served by the CDN are included; the count reflects requests reaching the zone, not unique pages or sessions.
Six of the tracked identifiers — Google-Extended, Applebot, Bytespider, CCBot, MistralAI, and Cohere — recorded no matched requests during the window; the nine bots in the table are the ones that actually showed up.
Known limits: user-agent strings can be spoofed, so a small share of matched requests may be impostors; conversely, AI fetchers not on the list (and any that hide behind browser user-agents) are not counted. Treat 604 as a floor for identified AI traffic, not an exact census.
If you reference this data, please cite the window (26 June – 2 July 2026), the site (getgeofix.com, a small SaaS site), and the method (CDN-level user-agent matching). A different week on a different site will produce different numbers — that is rather the point of publishing the method.
How to check AI bot traffic on your own site
One thing that will not work: Google Analytics. AI crawlers and fetchers do not execute the JavaScript that analytics tags rely on, so GA4 is effectively blind to them. You need something that sees raw HTTP requests. Three practical options, from easiest to most manual:
Cloudflare (free plan included). If your site is behind Cloudflare, open Security → Bots, or query the GraphQL analytics API the way we did and group requests by user-agent. Cloudflare also labels many verified AI crawlers for you.
Server access logs. Download a week of access logs from your host and search for the bot names listed in the methodology above. Our guide to verifying AI bots in server logs walks through the exact patterns.
Your host's or CDN's built-in analytics. Netlify, Vercel, Fastly, and most managed hosts expose request-level or user-agent-level views that show bot traffic that never reaches your analytics tag.
Whichever route you take, separate the user-agents into the two groups from this article: training and index crawlers (GPTBot, ClaudeBot, OAI-SearchBot, PerplexityBot, CCBot, meta-externalagent) versus live, user-triggered readers (ChatGPT-User, Claude-User, Perplexity-User). The second group is the one that tells you real people are already meeting your site through an assistant.
What to do once you know they are visiting
Counting is diagnosis. The treatment is a set of deliberate decisions rather than one big switch:
Decide your policy per bot type, not per company. Blocking training crawlers while keeping search and user-triggered agents open is a coherent strategy; blanket-blocking everything with one wildcard rule usually is not. Our robots.txt guide for AI bots covers the trade-offs.
Make the pages AI reads most actually readable: key facts in plain HTML text, not only in images or JavaScript-rendered widgets; clear headings; one page per question where you can manage it.
Add machine-readable context: an llms.txt file that points AI systems at your most important pages, and structured data that states who you are and what you offer.
Re-check after any security or CDN change. Firewall rules and bot-protection features can silently start rejecting AI user-agents that your robots.txt happily allows.
None of this guarantees that assistants will cite you — no honest tool or consultant can promise that. What it does is remove the failure mode where an assistant arrives seventeen times a day and finds a page it cannot use.
FAQ
Check a source that sees raw HTTP requests: Cloudflare or another CDN dashboard, your server access logs, or your host's request analytics. Search for user-agent strings such as GPTBot, ClaudeBot, OAI-SearchBot, ChatGPT-User, and PerplexityBot. Web analytics tools like GA4 will not show these visits because bots do not run the analytics JavaScript.
All three belong to OpenAI but do different jobs. GPTBot collects content for model training. OAI-SearchBot builds and refreshes the index behind ChatGPT search. ChatGPT-User fetches a specific page in real time when a ChatGPT user asks about it or opens a cited source. In our week of data they behaved very differently: 107, 159, and 99 requests respectively.
Not by themselves. A visit means an AI system requested your page; whether the content is then used in an answer depends on what the bot could read, how relevant it was, and choices made by the platform. Access is necessary but not sufficient — treat crawl data as evidence of opportunity, not proof of visibility.
It depends on what you are protecting and what you want to grow. Publishers protecting original content often block training crawlers. Businesses that want to be findable through assistants usually keep search and user-triggered agents open. The common mistake is blocking by habit with one wildcard rule and losing live reads you actually wanted.
GA4 and similar tools count visitors by executing a JavaScript tag in the browser. Crawlers and AI fetchers request the raw HTML and do not run that tag, so they never appear in your analytics reports. This is why AI bot traffic has to be measured at the CDN or server-log level.
There is no reliable public benchmark for small sites yet — most published numbers come from large publishers and network-wide CDN reports. Our single data point: a small SaaS site logged 604 identified AI requests in seven days. The only way to know your number is to check your own logs, which takes a few minutes with the steps above.
AI bots already read small websites: 604 identified requests hit this small site in one week — roughly one every 17 minutes.
About one visit in five was a live, user-triggered read (ChatGPT-User, Claude-User, Perplexity-User) — an assistant reading a page for a real person mid-conversation.
Measure at the CDN or server-log level (analytics tags cannot see bots), then decide your robots.txt policy per bot type and make key pages readable in plain HTML.
Free · 2 minutes · no card
AI bots are already visiting. What can they actually read?
The free Express Check reads your site the way AI crawlers do — robots.txt, llms.txt, structured data — and shows what is open, blocked, or missing.
