Sometimes they influence a person. Sometimes they send the click. Sometimes they complete the signup themselves. We are starting to see four distinct agent funnels.
Most website analytics assumes the visitor and the customer are the same person.
Someone lands on a page, looks around, and signs up or buys. The company connects the traffic source, website session, and conversion into one funnel.
AI agents break that assumption.
The person with the need might never visit the website. The agent visiting the website might hand the task back to a person. The conversion might happen inside ChatGPT. Increasingly, the agent may create the account and use the product itself.
Calling all of this “AI traffic” hides what is actually happening.
The four agent funnels
1. The agent influences a person
Someone asks an AI assistant for help. The agent researches the options, compares products, and recommends a company.
The person does not click immediately. They might search for the company later, type in its URL, or hear about it again somewhere else before converting.
The agent influenced the decision, but analytics attributes the conversion to search, direct traffic, or another channel.
This is probably the largest agent funnel and the hardest to measure. It is similar to word of mouth: the recommendation matters, but the eventual visit carries little evidence of where it came from.
2. The agent refers a person
This is the agent funnel companies can see today.
Someone asks ChatGPT, Claude, Gemini, or Perplexity for a recommendation. The agent links to a company, the person visits its website, and the rest of the conversion looks relatively normal.
The difference is what happened before the click. The agent may have already compared the options, answered the person's questions, and narrowed the decision. The visitor arrives with more context and stronger intent.
The data is starting to show this. In March 2026, Adobe found that AI-referred visitors converted 42% better on retail websites and 54% better on travel websites than visitors from other sources. They also spent more time on the site and viewed more pages. (Adobe retail data, Adobe travel data)
AI referrals are still small relative to search and other major channels. But the people arriving through them increasingly look like some of the highest-intent visitors on the internet.
3. The agent starts and the person finishes
Many agent conversions will be shared between an agent and a person.
The agent might research the options, configure an order, fill out a form, or create an account. The person appears only when approval, identity verification, or payment is required.
This is already happening in commerce. ChatGPT lets people find and purchase products inside the conversation. The merchant receives the order and owns the customer relationship, but the buyer never enters its normal website checkout. (OpenAI)
Google's Universal Commerce Protocol is enabling similar experiences across AI Mode and Gemini, including native checkout and account linking. (Google UCP)
The agent does most of the work. The person approves the final action.
For these conversions, the most important experience may not be the homepage or product page. It may be the handoff between the agent and the person.
4. The agent converts itself
Across the websites Sherpa watches, we are starting to see agents complete signup flows themselves.
The agent may be working for a person or company. It might create an account, connect a service, make an API call, or purchase something it needs to finish a larger task.
The infrastructure for this is being built quickly. Microsoft now gives agents their own identities, sponsors, and permissions. Google's Agent Payments Protocol supports transactions completed without the human present. Coinbase already lets agents discover an API, pay for a request, and use it without creating an account, managing an API key, or asking a human to approve each transaction. (Microsoft, Google, Coinbase)
In this funnel, the agent is not a traffic source. It is the user.
Not every agent visit is demand
An automated request does not necessarily mean an agent is considering the product.
Some agents are researching for a person. Others are collecting training data, indexing pages, checking prices, or scraping content.
Cloudflare found that nearly 80% of AI bot activity was related to training. Some AI platforms crawled thousands of pages for every human referral they sent. (Cloudflare)
Counting every agent request as interest would be like counting every Googlebot crawl as a potential customer.
Companies need to distinguish between crawlers collecting information, agents researching with intent, agents representing existing customers, and agents trying to complete an action.
The useful question is not how much agent traffic a website receives. It is what those agents are trying to do and what happens afterward.
The existing funnel cannot measure this
Traditional analytics connects a source, a session, and an outcome.
Agent conversion can involve several actors across several surfaces:
- An agent researches the product.
- A person receives the recommendation.
- Another agent visits the website.
- The person approves an action.
- The agent creates the account or completes the purchase.
- A person or agent uses the product afterward.
That chain can create revenue without a referral, a website session, or even a human signup.
Companies will need to measure agent influence, agent referrals, agent-assisted actions, and agent-completed actions separately. They will also need to connect agents to the people or organizations they represent.
Otherwise, their best new acquisition channel may appear as direct traffic, bot traffic, or nothing at all.
The agent might be the top of the funnel or the customer
Most companies currently think about agents as a discoverability problem: make sure ChatGPT can find, understand, and recommend the product.
That is only the first funnel.
Some agents will influence people without sending a click. Some will refer highly qualified visitors. Some will complete most of the conversion before asking for approval. Others will sign up and use the product themselves.
AI agents do not convert like humans because they do not play a single role.
They can be the researcher, the referral source, the buyer, or the customer.