Web analytics fundamentals

The IAB's three AI pathways, filled in with what a small site can count

By Emily Parker · Nov 15, 2026 · 9 min read

What the IAB published

On Nov 12, 2026, the IAB released its AI-Era Attribution Blueprint. Per the release, it classifies each conversion as Human-Owned, AI-Assisted or AI-Executed, and it asks who or what created the intent, made the decision and completed the purchase. The release names the three categories but does not define them in its text. The definitions below are our working reading for a small site, not IAB wording.

The release is direct about the data problem. It says much of the evidence needed to support attribution claims is "either not captured or sits inside AI platforms." It cites the IAB's 2026 Outlook Study: September Update, in which 44% of ad buyers said they rely on branded search and direct traffic as proxies for AI influence they cannot track. Caroline Giegerich, the IAB's VP of AI, is quoted saying the clicks and referrals that attribution depends on are "thinning out or disappearing." MarTech's coverage adds that the blueprint "does not solve the lack of observable data behind many of them."

One recommendation in the release is aimed at brands: estimate the share of your conversions in each pathway. That is the task for a small site too. The useful version of it separates what you counted from what you are guessing.

Why a referrer count is a floor

Start with how rare the click is. Pew Research Center analyzed browsing data from 900 U.S. adults and found that when a Google results page showed an AI summary, users clicked a traditional result link in 8% of visits. Without a summary, they did so in 15% of visits. Clicking a link inside the summary happened in 1% of visits to pages that had one.

8%15%1%Traditional result clicked, AI summary shownTraditional result clicked, AI summary...Traditional result clicked, no AI summaryLink in AI summary clicked

Source: Pew Research Center, Google users are less likely to click on links when an AI summary appears in the results, 2025-07-22. Retrieved Nov 12, 2026.

Read the limits before you borrow the number. It is a study of Google's AI summaries, not of ChatGPT or Claude. It measures clicks, not signups or sales. It says nothing about your site. What it supports is narrow: if an assistant answers a question and the link in that answer is rarely followed, then a count built on followed links will undercount the assistant's influence. A referrer count is a floor by construction.

The three questions against your own records

Here is our working reading of the three pathways, followed by what a small site's records can say about each question.

  • Human-Owned: a person had the idea, made the choice and paid, with no AI step you can observe.
  • AI-Assisted: an AI shaped the idea or the choice, and a person completed the purchase.
  • AI-Executed: software acting for a person completed the purchase.
IAB questionWhat your records can showWhat they cannot show
Who created the intentA visit whose referrer is ChatGPT, Claude, Perplexity, Copilot or GeminiThe question that was asked, and any assistant visit that arrived without a referrer
Who made the decisionA goal reached later by a person who first arrived from an assistantWhether the assistant caused the decision or the person had already made it
Who completed the purchaseA verified sale with its page and sourceWhether a person or an agent clicked pay

Three paths leading to a single shop counter: one clearly lit, one faintly lit, and one fading into fog before it arrives.

The first row is the only one where your records hold direct evidence. The second holds a sign, not proof. The third holds nothing that answers the question. The sections below take them in that order.

AI-Assisted: count the floor

Peeko counts visits from ChatGPT, Claude, Perplexity, Copilot and Gemini from ordinary page views, by their referrer, with no middleware involved. Our post on counting visits from AI assistants by referrer covers that step. Turning that count into an AI-Assisted floor takes two more things.

1. A goal that matches the conversion. A goal is a definition rather than a counter, so a goal you write today counts over the months before you wrote it. You do not have to guess in advance which signups will matter. Five kinds of click are recorded with no code: outbound, phone, email, whatsapp and form. For anything else, add a name to an element or call it from your own code:

<a href="/signup" data-peeko-event="signup_click">Start free</a>
window.peeko.event("signup_complete");

Each takes one name and one optional string, with no properties object.

2. A way to connect the earlier visit to the later goal. A second hash, built from the address and browser string without the date, tells a returning visitor from a new one and strings a person's page views into a journey. It is the identity behind goals, funnels and revenue. The count you want is the number of people whose journey began with an AI assistant referrer and later reached the goal. We walk through the full path in how to follow a visit from an AI answer through to a signup. You can read the same numbers in the dashboard, through npx @usepeeko/cli, or by asking an agent over MCP.

Be exact about what that number means. It is people who arrived from an assistant and later reached the goal. It is not people the assistant persuaded. A person might have been ready to sign up anyway. That is why we call it a floor for AI-Assisted and not a measure of AI influence.

It is also a floor because the identity is built from an address and a browser string. A person who clicks from ChatGPT on a phone and signs up on a laptop is two visitors. The assistant visit is counted, the goal is counted, and the link between them is lost.

The part that lands in Direct

The release also notes that AI-influenced sales are sometimes reported as search or direct traffic. On a small site the mechanism is easy to picture. A person reads an answer, remembers your name, and types it into a browser a day later. No referrer exists to count. Nothing in your records distinguishes that visit from a bookmark.

You cannot recover those visits one by one. You can watch the same proxy the 44% of ad buyers use, and you can be honest that it is a proxy. Two places to look:

  • Your Direct share of visits, before and after a stretch where an assistant started recommending you. Our post on checking whether ChatGPT visits moved to direct traffic shows the comparison.
  • Branded queries in Google Search Console and Bing Webmaster Tools, which connect read only and keep their rows for sixteen months. Google publishes search data about three days late and revises its three most recent days. Bing's query rows are weekly buckets dated to a Friday.

A rising Direct share or branded query count next to a rising assistant referrer count is a hint. It is not a number you can assign to a bucket. If you hand out links yourself, tag them with the free UTM builder, which builds tagged links in the browser and stores nothing, so your own newsletter and social clicks stop landing in Direct and muddying the proxy.

AI-Executed: what a sale record does not say

For revenue, Peeko takes one of two routes: a one line window.peeko.revenue call from your own site, or a verified Stripe webhook. The two totals are never added together, and amounts in different currencies are listed, not converted. Peeko never holds a Stripe API key. The only credential it stores is the webhook signing secret, which proves an event came from your account and does nothing else.

Attribution works without a cookie through a page token. It is minted per page view, good for 48 hours, and passed into the Checkout session as metadata. A verified sale then carries its page and its source. The full setup is in how to attribute a sale to the page and source that produced it.

That record answers where a sale came from. It does not answer who or what clicked pay. Nothing in it says an agent made the purchase, so Peeko cannot report AI-Executed and we will not pretend a number is one. Two traps to avoid:

  • A missing token is not a label. A sale with no page token might come from a payment link, a token older than 48 hours, a blocked script, or a purchase with no recent page view. Each is a reason to look. None is proof that software paid.
  • A zero is not evidence. If you have no way to detect AI-Executed purchases, a count of zero means you did not measure it, not that none happened.

Our post on what a purchase by an AI agent leaves in your revenue numbers lists what the two revenue routes would and would not show if an agent bought from you.

Answer fetches: a read, not a person

There is one more AI signal a server can see, and it belongs to the content side of the story, not the conversion side. AI crawler tracking is a separate server side path, because crawlers do not run JavaScript. The @usepeeko/bot-track middleware is ESM only, needs Node 20 or newer, has zero runtime dependencies and weighs under 3 KB gzipped. Peeko matches a crawler user agent against a curated list of 55 agents from 21 companies, and files each recognized hit under AI answers, Search indexing, AI training or Other AI bots.

An answer fetch says a page was read while an assistant was replying to somebody. The question itself never reaches your server, so nobody can report it. No address and no visitor hash exist on the crawler path at all, so a fetch can never be tied to a person, a journey or a sale. It tells you which pages assistants lean on. It cannot move a conversion into a bucket.

It is also a small slice of what AI crawlers do. Cloudflare reported that over the 12 months to its August 2025 post, 80% of AI crawling was for training, 18% for search and 2% for user actions. Cloudflare's categories do not match Peeko's, so treat user actions as a rough neighbor of answer fetches, not the same count.

80%18%2%TrainingSearchUser actions

Source: Cloudflare, The crawl-to-click gap: Cloudflare data on AI bots, training, and referrals, 2025-08-29. Retrieved Nov 12, 2026.

What to write when someone asks for the three shares

You will be asked for the split, by a client, a vendor or yourself. Give them this, with your own counts in the brackets:

Conversions in period: [total]
AI-Assisted, counted floor: [n] (arrived from an AI assistant referrer, later reached the goal)
Direct, unattributed: [n] (an unknown share is AI-influenced)
AI-Executed: not measured
Human-Owned: not measured, only what remains by exclusion

The Human-Owned line is the one people skip. Without a way to see AI influence, you cannot confirm that a conversion had none, so the remainder is not a count of human-owned conversions. It is the conversions your data could not place.

The IAB asks AI platforms to label AI-originated traffic consistently. Until they do, the floor stays a floor. If your numbers move, you will know because you wrote down what each line could and could not see.