punchline!

Blog 路 November 4, 2026 路 Punchline Analytics

Your attribution model has a new blind spot

Buyers are sending AI agents to do the research your attribution model was built to track. Here's where the gap is, why your tools can't see it, and four ways Ops can measure it this quarter.

If we had a dollar for every pipeline review where someone asked "so where did this deal actually come from?", we'd have retired to a beach a long time ago. (And we'd still be checking the attribution report from our phones. Old habits.)

Every Ops pro we know has a love-hate relationship with attribution. We build the model, we defend the model in front of Finance, and we quietly know where the bodies are buried.

"Direct" is usually where they're buried.

Here's the thing - there's a new set of bodies, and they're piling up fast.

The research step moved

Think about how a B2B buyer used to build a shortlist. They'd search, click through to a handful of sites, read your pricing page, maybe download a whitepaper (or, let's be honest, close the tab the second they saw the form). Every one of those steps left a trail your tools were built to follow: a session, a referrer, a UTM, a form fill.

A growing share of that research now happens somewhere else. You know this - your process has probably changed as well. Your buyer opens ChatGPT or Claude or Perplexity and types "shortlist analytics tools for a 200-person software company." The assistant goes and reads your site for them. Your pricing page, your security page, your comparison page. It summarizes, compares, and hands your buyer a shortlist.

Your buyer never visited your site. Their agent did.

A few days later, someone from that company books a demo. Your attribution model looks at the session, finds no referrer, and does what it always does: "Direct." 馃し

It's a bit like scouting an athlete. The scout watched three of your games from the stands, wrote the report, and the GM made the call. Your box score only shows the day the player signed.

Why your tools can't see it

This isn't a configuration problem you can fix with a stricter UTM convention. (We know. We wanted it to be, too.) It's structural, for three reasons.

  • Most of these agents don't run JavaScript. Your web analytics tag, your marketing automation tracking script, your ad pixels - they all run in a browser. An assistant fetching your page to answer a question usually requests the HTML and leaves. The tag never fires. As far as your tech stack is concerned, it never happened.
  • The ones your tools do notice get thrown out. Google Analytics 4 excludes known bot traffic automatically, and you can't see how much it removed. Most privacy-friendly tools filter even harder. That made total sense when "bot traffic" meant scrapers and uptime monitors. It makes a lot less sense when it means your buyer's research assistant.
  • The click, when it comes, is the last step. GA4 now has an AI Assistant channel, which is genuinely helpful. But it counts people who clicked a link in an AI answer. That's the punchline of the journey. The setup - the research that got you on the shortlist - happened before anyone clicked anything.

So your model is crediting the end of the story and filing the rest under noise.

"But," you say, "if the buyer still shows up and books a demo, who cares how they did their homework?"

Fair question. Here's why it matters.

What this does to the numbers you report

When a meaningful part of the research step goes dark, a few things start to happen:

  • Direct and organic look better (or worse) than they are, depending on how your model handles sessions with no source. Either way, the trend line is telling you a story that isn't quite true.
  • Your best content looks like your worst content. Your comparison page might be doing a ton of work getting you onto shortlists. If agents are the ones reading it, it shows up as low traffic and zero conversions, and it's first on the chopping block at the next content review.
  • Sales hears about interest late. By the time a person fills out your form, their agent may have read your pricing page three times. You had no way to know.

None of this throws an error. Nothing breaks. That's what makes it a blind spot instead of a bug - and blind spots are exactly the things that come up in a QBR at the worst possible moment.

What Ops can do this quarter

You don't need to rebuild your attribution model. (Please don't. Finance is still recovering from the last one.) You need to start measuring the part you can't see, so you can make a decision with real numbers instead of vibes.

Here are four approaches, roughly in order of effort.

Approach 1: Size it from your logs (a few days)

Your server or CDN logs already contain every request, including the ones your tags miss. Look for the user agents AI companies publish: ChatGPT-User, Claude-User and Perplexity-User for live fetches on someone's behalf, and GPTBot, ClaudeBot and friends for crawling. (There's a directory here with how to verify each one, because some of those user agents are fakes.)

Pros: free, and you get a number this week. Cons: someone has to own the query, the list of agents changes constantly, and agents that don't say who they are won't show up at all. This is a snapshot, not a system.

Approach 2: Separate "reading for someone" from "crawling for training" (same project, an extra afternoon)

These are two different businesses. A training crawler reading your blog is a policy question for Legal and Marketing. A live fetch of your pricing page means someone asked an assistant about you, moments ago. That's a buying signal, and it belongs in a different report.

Pros: turns a scary "bot traffic" number into something you can act on. Cons: still a snapshot unless you automate it.

Approach 3: Add one question to your forms (an hour, plus a sync to your CRM)

"Did an AI assistant help you find or choose us?" Optional, one click. It sizes the influence that never turns into a referral. Podcast advertisers have been calibrating with "how did you hear about us?" for years - same idea.

Pros: cheap, and it's data straight from the buyer. Cons: self-reported, so treat it as directional. And only the people who fill out a form answer it.

Approach 4: Give agents something to hand back (a sprint, with a developer)

When an agent builds something on your site - a quote, a demo booking, a configuration - give it a link to pass to its person, with a token in it. When that person opens the link days later, the visit connects to the agent's work instead of landing in "Direct." That's what handoff links are for, and in our tests agents passed them on every time they finished the task.

Pros: the only approach of the four that connects agent research to an actual conversion, exactly, without guessing. Cons: needs engineering time, and it only works where an agent can actually do something on your site.

Our recommendation

Do 1 and 3 this month. They're cheap, and they get you a number and a buyer-reported signal you can put side by side. If the number is big enough to matter (and for most B2B sites we'd bet it is), make 2 a standing report and put 4 on the roadmap for the flow that matters most - usually the demo booking.

Then, whatever you find, tell your stakeholders before they ask. This one's for anyone who's ever been blindsided in a QBR. If agent research is changing your numbers, you want to be the one explaining it, not the one being asked about it.

The bottom line

Your attribution model isn't broken. It's measuring a world where every visitor was a person in a browser. That world is going away, one shortlist at a time.

The good news: this is exactly the kind of problem Ops was built for. It's messy, nobody owns it yet, and the people who figure it out first are going to look very, very smart in next quarter's pipeline review.

Punchline was built to see that research step: every agent, why it came, and what it led to. See it on your own site.

Start free