Attribution when an AI agent sits between the buyer and you

The IAB said at the end of August that it is building a framework for crediting AI’s role in conversions. The reason it needs one is in the brief: when an agent reads your content, compares you against two competitors and either recommends you or completes the purchase, the resulting visit carries no referrer and no UTM. The two fields most B2B attribution is built on are empty, and the platforms doing the mediating are the only ones who can see what happened.

The same month, Brainlabs data across 54 advertisers put AI referral traffic up 163% while organic sessions fell 10.5%, with AI-referred visitors completing key events at roughly 1.5 times the rate of organic. The traffic that matters most is the traffic you can see least. Waiting for the IAB to publish is a choice to run a year of reporting that under-credits your best channel and over-credits direct.

Why the current model breaks

Three mechanics, each of which removes a signal you relied on.

The click disappears. A buyer who asks ChatGPT or Gemini to shortlist vendors reads the answer and types your URL in, or searches your brand name. Your analytics logs that as direct or branded search. The AI session that created the intent is invisible.

The identifier does not travel. When a link does come through from an assistant, it often arrives stripped. Where a referrer does survive, your channel grouping probably files it under referral alongside a partner blog from 2019.

The timing stretches. Google Ads data-driven attribution ignores offline conversions uploaded more than seven days after the click, and ChatGPT’s Conversions API has its own seven-day window. A B2B deal that takes three weeks to become an SQL never informs bidding on either platform, so your dashboards look complete while the optimisation model learns from a fraction of your conversions.

Log what you control

You cannot make the platforms transparent. You can build an evidence base that does not depend on them.

1. Capture conversions server-side. Move form submissions, demo requests and signups to a server-side event that writes to your warehouse before anything goes to an ad platform. Include timestamp, landing page, user agent and whatever identifiers the visitor supplied. This is the record you will reconcile platform claims against later.

2. Ask. A self-reported attribution field on every high-intent form, free text with a short placeholder, catches the AI-mediated journey that no pixel can. “Asked ChatGPT for options” is a sentence buyers already write unprompted when the field exists. Store it raw, categorise it monthly, and report it alongside the tracked channels rather than beneath them.

3. Separate the AI channel. Create a channel grouping for traffic from ChatGPT, Gemini, Perplexity, Copilot and Claude domains, and weight it by key events rather than sessions. A report that ranks channels by traffic volume scores AI referrals as a failure while they outconvert organic.

4. Read your server logs for AI crawlers. GPTBot, ClaudeBot, PerplexityBot and Google-Extended hitting a page is the earliest signal that the page is being read into an answer. A weekly count of crawler hits by URL tells you which content is in the consideration set, weeks before any referral shows up. The mechanics of getting that content cited at all are in citation infrastructure.

Fix the timing

The seven-day windows are a plumbing problem with a plumbing fix. Replace the weekly or fortnightly offline conversion batch with an event-driven job: a HubSpot or Salesforce stage-change webhook into n8n that sends the GCLID or the hashed identifier and the value to the platform within minutes. Where the stage you want to optimise towards lands beyond seven days from click, optimise to an earlier stage that fits the window and keep the later stage for reporting. The ChatGPT Ads version of this setup is in ChatGPT Ads as a B2B pipeline channel.

Check what arrives on its own

Not every attribution change is one you opted into. ChatGPT Ads turned on automatic advanced matching for existing web pixels on 17 August, which hashes customer data from your site forms in the browser and attaches it to conversion events. If your legal team scoped those forms for a different purpose, that transfer needs a review. Audit the pixel settings on every ad platform quarterly for anything that started sending data without a decision from you.

What this gets you

When the platforms start reporting agentic conversions, and they will, with no referrer and no way to verify, you will have four independent records to compare them against: server-side conversions, self-reported source, AI channel key events, and crawler hits by page. Those four records let you audit a platform’s number instead of accepting it. The gap this leaves in the zero-click world was already there before agents arrived; the background is in the attribution gap nobody’s measuring.