AI Marketing Attribution: How to Actually Measure What Works

I had a call last month with a business owner who was about to fire his entire marketing agency. His Google Ads dashboard said the campaigns were crushing it. His actual bank account said revenue was flat. His Facebook dashboard also said the campaigns were crushing it. His GA4 said something different from both. His Shopify said something different from all three. He asked me which number was right. The answer, which nobody wants to hear, is that none of them are right in 2026, and the old idea of clean attribution is gone. I'm Jake McCluskey, and after 25 years in digital marketing and working with 500+ businesses, I can tell you attribution is the single most broken part of marketing right now. But it's fixable if you stop expecting it to be perfect.

Why is marketing attribution broken right now?

Marketing attribution is broken right now because three things happened at the same time: browser cookies stopped working reliably, AI-driven traffic became opaque by design, and user behavior shifted from click-then-buy to research-in-AI and show-up-direct. The tools didn't keep up. Most dashboards are still reporting on a web that no longer exists.

Here's what changed. Safari's Intelligent Tracking Prevention has been eating cookies since 2020. Firefox followed. Chrome's slow-motion cookie deprecation is still dragging on, but in practice third-party tracking is already unreliable on more than half the web. Your pixel data is missing chunks and you don't know which chunks.

Then ChatGPT, Perplexity, Claude, and Google's AI Overviews started sending traffic without clean referrers. When somebody asks ChatGPT a question and then visits your site, GA4 often logs it as direct traffic, not as referral. You literally cannot see that channel in your reports without instrumentation you probably haven't set up.

And users now routinely research a purchase across six or seven sessions, half of which happen inside AI tools that never touch your pixel. By the time they convert, the attribution path looks like "direct" or "organic brand," which tells you almost nothing about which marketing actually worked.

If your attribution setup was designed before 2023, it's broken. You just might not know by how much yet.

What's the difference between last-click, multi-touch, and incrementality attribution?

Last-click attribution gives 100 percent of the credit to the final touchpoint before conversion. Multi-touch attribution spreads the credit across every touchpoint a user interacted with. Incrementality attribution ignores both of those and asks a different question entirely: would this sale have happened without the marketing? Each model tells you a different story, and they're all partially true.

Last-click is the default in most ad platforms. It's wrong, but it's at least consistent. The problem is it wildly overvalues bottom-funnel channels (branded search, retargeting) and undervalues everything that built the demand in the first place (content, brand, PR, word of mouth).

Multi-touch is better in theory and worse in practice. The models (linear, time-decay, position-based, data-driven) all depend on tracking a user across every touchpoint. Cookies are dying, so the tracking is broken, so the model has bad inputs. Garbage in, garbage out. A multi-touch report built on 40 percent of your actual touchpoints is not 40 percent right, it's confidently wrong.

Incrementality is what serious teams have moved to. Instead of trying to trace every click, you run controlled tests: geo holdout experiments, ghost bids, traffic-off periods. You compare what happened in the test market against what happened in the control market. The delta is the incremental lift. It's messier to set up, but it's the only method that survives the cookie apocalypse.

For most small and mid-sized businesses, I recommend a blend: use last-click as a rough directional read, skip multi-touch entirely, and run small incrementality tests on your largest channels quarterly. That's a practical stack that actually tells you something.

How do I build a measurement stack for a small or mid-sized business?

A small or mid-sized business can build a good-enough measurement stack with four tools and about 10 hours of setup: GA4 with custom dimensions, a server-side conversion feed for ad platforms, a self-reported attribution question on your contact form, and a simple spreadsheet that reconciles all three against revenue in your CRM. You don't need an enterprise stack. You need discipline with these four things.

Here's the stack I recommend and why each piece matters:

• GA4, properly configured. Set up custom events for every meaningful action (form fills, demo bookings, calls, newsletter signups). Add custom dimensions for marketing channel, landing page, and session source. Default GA4 reports are not good enough. Build your own in Looker Studio.
• Server-side tagging (Google Tag Manager server container, or Stape). Browser-side pixels are getting blocked. Server-side sends data directly from your server to Google and Meta, which recovers 15 to 30 percent of lost conversions.
• A CRM that stores UTM parameters. HubSpot, Pipedrive, and Salesforce all do this if you configure it. Every form submission should save the source, medium, campaign, content, and landing page of the lead's first session. Without this, you're flying blind.
• A self-reported attribution question on your contact form. One dropdown: "How did you hear about us?" It sounds dumb. It's the single most valuable data point you'll collect. More on this below.

The reconciliation spreadsheet is where it all comes together. Once a month, pull leads from your CRM, match them against ad platform conversions, check GA4 sessions, and compare to self-reported source. You'll see the gaps. The gaps are the story.

If building this stack yourself sounds like a full-time job, that's because it half is one. Our services include setting this up end-to-end for clients who want it done right the first time.

How do I track AI-driven traffic specifically?

You track AI-driven traffic by setting up referral exclusions, custom channel groupings, and a few regex patterns in GA4, then cross-checking against your self-reported attribution data. Most AI tools send traffic that GA4 misclassifies as direct or organic, so you have to build the rules manually. There's no out-of-the-box report for this yet.

Here are the concrete steps I use for clients:

1. Create a custom channel group in GA4 called "AI Search." Include referrals from chat.openai.com, chatgpt.com, perplexity.ai, claude.ai, copilot.microsoft.com, gemini.google.com, and you.com. Add new domains as they appear. I update this list quarterly.
2. Add a regex filter for AI-related utm_source values. When you tag links inside AI tool plugins or in Perplexity Pages, use consistent UTMs like utm_source=perplexity or utm_source=chatgpt so your own marketing gets attributed properly.
3. Track "direct with no prior session" as a proxy for dark AI traffic. A new user with no prior touch who lands on a deep page (not your homepage) is almost always coming from an AI answer or a word-of-mouth source. Build a report that isolates this segment. Watch it grow month over month.
4. Add a UTM discipline policy. Every paid link, email link, and partner link gets a tagged UTM. No exceptions. If you don't enforce this internally, your attribution data will be polluted by your own team's lazy links.

The dark AI traffic problem is real and getting bigger. In my own data across client sites, direct traffic as a percentage of total has jumped 40 to 90 percent in the last 18 months. That's not "people typing your URL." That's AI referrals with stripped referrers.

You can't fix the tracking at the source. You can fix it at the destination by asking people how they found you.

Why are self-reported attribution surveys actually useful?

Self-reported attribution surveys are useful because they capture the signal that tracking misses entirely: the human memory of how someone discovered you. When a user types "ChatGPT recommended you" or "heard you on the Marketing Over Coffee podcast" into a form field, that's data no cookie, pixel, or model will ever give you. For a small business, this is often the most reliable signal you have.

The implementation is simple. Add one required dropdown to your contact or demo request form with the question "How did you hear about us?" and options like:

• Google search
• ChatGPT or another AI tool
• Podcast (specify which one in the next field)
• LinkedIn
• Referral from a friend or colleague
• YouTube
• Event or conference
• Other

Add an open-text "tell us more" field. You'll be amazed what people write. I've had clients discover entire traffic sources they didn't know existed because 20 leads in a row wrote in a specific podcast they'd never heard of.

The data is self-reported so it's noisy. People forget. People lie to themselves. But in aggregate, over hundreds of leads, the directional signal is remarkably stable. Combine it with your platform data and your CRM data, and you'll have a three-legged stool that's more reliable than any single source.

I now consider a contact form without this question a basic hygiene failure. If yours doesn't have it yet, that's the single fastest attribution fix you can make this week.

What does "good enough" attribution look like if you can't afford a data team?

Good enough attribution for a business without a data team is a monthly review where you can answer three questions with reasonable confidence: which channels are producing revenue, which channels are producing leads that don't convert, and what percentage of new customers came from something your marketing touched. You don't need second-by-second dashboards. You need a monthly story that matches the money.

Here's the minimum viable attribution workflow I teach clients:

1. Once a month, pull a revenue report from your CRM or accounting system. Group new customers acquired that month.
2. For each new customer, note the self-reported source from their form submission. This takes 20 minutes for most small businesses.
3. Cross-check against ad platform reports. Are the platforms claiming customers that the self-reported data says came from somewhere else? That's where you're paying for credit you didn't earn.
4. Kill or shrink the channels that can't prove they worked. Being unable to prove it is itself a signal.
5. Document the monthly conclusion in one paragraph. "This month, 62 percent of new revenue came from Google organic and direct. Meta claimed 18 conversions but only 4 showed up in self-reported data. We're cutting Meta spend 50 percent next month."

That one-paragraph monthly summary, done consistently for 12 months, is worth more than any $50,000 attribution platform that you don't have time to interpret.

The goal isn't perfect attribution. Perfect attribution doesn't exist anymore. The goal is being less wrong than your competitors. Most small businesses aren't measuring anything real right now, which means even a mediocre attribution setup puts you ahead.

If you don't know where to start, a free audit will show you exactly what's instrumented and what's leaking today.

What attribution mistakes do most businesses make?

Most businesses make attribution mistakes that fall into three buckets: trusting a single platform's dashboard, ignoring what customers actually say, and changing models every time a new report looks bad. Any one of these will wreck your decision-making. All three together guarantee you'll spend money on things that aren't working and cut things that are.

The single-platform trap is the most common. Google Ads will always claim credit for Google Ads. Meta will always claim credit for Meta. They're not lying, exactly. They're optimizing reports to look good inside their own walled garden. If you run a campaign in both, and the total claimed conversions exceed your actual total, you've got double-counting. It's almost always happening.

The ignoring-customers trap is more subtle. I've had clients look at platform data and ignore 40 self-reported "podcast" responses on their contact form. The platform said podcasts drove 2 percent of traffic. The customers said podcasts drove 30 percent of intent. Guess which one was right when we ran the test.

The model-hopping trap is the worst. A business sees a report they don't like, changes the attribution model, gets a different number, likes it better, and moves on. This isn't measurement. This is justification. Pick a model, commit to it for at least two quarters, and compare period over period. Otherwise you're just grading your own homework.

The point of attribution isn't to feel good. It's to make better decisions about where the next dollar goes. That requires consistency, honesty, and a willingness to hear bad news.

Marketing attribution will never be perfect again, and anyone who promises you a single clean dashboard is selling you comfort, not accuracy. The businesses making good decisions right now are the ones running a layered setup: real tracking where they can get it, self-reported data where they can't, incrementality tests on their biggest channels, and a monthly reconciliation against actual revenue. None of it is pretty. All of it is useful. If you want help building an attribution stack that survives the next platform change and actually tells you what's working, book a discovery call and I'll show you exactly what your setup is missing. I'll also tell you which of your current tools you can stop paying for.