> Site index: https://usehardal.com/llms.txt
> Every content route also serves markdown at <url>.md

# What Is Agentic Analytics?

Agentic analytics measures what AI agents do, including visits and actions taken with first-party data. Defined events, sources, and consent rules still apply.

Source: https://usehardal.com/glossary/agentic-analytics

---

Agentic analytics measures activity that involves an AI agent. The agent might fetch a page, recommend a product, open a site on someone's behalf, or complete a step such as a search or a purchase. The measurement job is to record those actions as events with a source, a time, and the properties a report needs.

The name describes the actor. It is not a special class of data that skips definitions, consent, or reconciliation. An agent event still needs a name and a decision about which fields leave your system.

## How does it work?

Keep three kinds of activity in separate events. They answer different questions.

- **Crawlers** fetch pages for an assistant's index or answers. They identify themselves with a user agent such as `GPTBot`, `OAI-SearchBot`, `ClaudeBot`, or `PerplexityBot`. They usually do not run JavaScript, so a tag in the page never sees them.
- **Referrals** are visits that follow a link from an assistant. ChatGPT adds `utm_source=chatgpt.com` to many of those links. A visit can also arrive with a referrer host such as `chatgpt.com`, `perplexity.ai`, `gemini.google.com`, or `copilot.microsoft.com` and no campaign parameter.
- **Agent actions** are steps an agent takes for a person, such as requesting a price or placing an order through a protocol your store accepts. The order still lands in your commerce system. The extra fact to store is that an agent started it, not a browser session you can replay.

A crawler hit is not a customer. A referral is a visit. An agent order is a conversion with its own source. One rolled-up "AI traffic" number hides which of the three you actually have.

## A concrete example

One product URL can see all three on the same day.

1. A server log records `user_agent` containing `GPTBot` and the path `/pricing`. No `page_view` event fires in browser analytics, because the crawler did not execute the tag. The log line is the record.
2. A browser session lands on `/pricing` with `utm_source=chatgpt.com` and `utm_medium=referral`, then sends `page_view` and later `purchase` with a `transaction_id`. That is a referral you can group with other tagged sessions.
3. Your checkout API records an order with `source=agent` and a `transaction_id`, and no browser session exists for it. Count the purchase from the order. Do not invent a session to hold it.

If the tagged session and the order API both emit `purchase` for the same `transaction_id`, deduplicate them. Two records of one order are not two customers.

## What are the limits?

Referrers are incomplete. A link copied out of a chat, or opened inside an app, often arrives with no `utm_source` and no referrer. It looks like direct traffic. Treating every direct visit as an agent invents a channel you cannot defend.

A mention in an answer is not a visit, and answers change from one run to the next. A single citation is not a ranking and not a traffic source.

Some agent-driven browsers look like ordinary Chrome. The user agent string will not reliably split them out. Signed agent requests are an emerging convention, not a field you should require before a report is useful.

Server-side collection does not recover a request that was never sent. It does not bypass consent, App Tracking Transparency, or browser controls. When an event identifies a person, purpose limits and permission still apply.

## How Hardal handles it

Hardal measures agent activity on the same first-party pipeline as the website and the app. [AI Visibility](https://usehardal.com/ai-visibility) shows which AI agents request the site and, when a paid channel can be identified, which channel that visit came from. The source is server-side data, not only a browser tag. [Hardal AI](https://usehardal.com/ai) answers questions from read-only aggregates of that data. It does not create, edit, or delete records.

When you send an event on to an ad platform, Hardal forwards the fields you allow. It does not relabel an unidentified direct visit as a confirmed agent, and it does not count a crawler hit as a conversion.

For the practical split between referrals and crawlers, read [Is ChatGPT sending you customers yet?](https://usehardal.com/blog/is-chatgpt-sending-you-customers-yet).

**Related terms:** [First-party data](https://usehardal.com/glossary/first-party-data), [Server-side measurement](https://usehardal.com/glossary/server-side-measurement), [Cookieless analytics](https://usehardal.com/glossary/cookieless-analytics)
