“You’re giving me 40% of my traffic from ChatGPT”: Why that metric is misleading (and how to measure the real impact)

Someone told me proudly while we were talking about Mi Seguro: “You’re sending me almost 40% of my traffic from ChatGPT.” It’s tempting to believe it and go share it everywhere. But if you’re building products, that line is where the problem begins, not where it ends.

Why “40% comes from ChatGPT” is an incomplete truth

No one can accurately measure how much of your traffic “comes from” an LLM. ChatGPT, Perplexity, and others rarely appear as a clean referrer in your analytics. What actually happens, more often these days:

• Traffic shows up as “direct” or “unknown” (dark traffic).
• The journey is multi-touch: someone asks an LLM, then Googles your brand, and finally lands directly.
• Attribution breaks easily if it isn’t properly instrumented.

So that “40%” could be just as misleading as it is impressive. It’s not a lie, it’s a poorly defined metric.

The right question isn’t “who sent me the click”

When you’re building an intent-driven product, a comparator, a SaaS tool, a programmatic SEO matrix, what matters isn’t the last click. What matters is:

• How you’re measuring attribution: GA4, server logs, real referrers, Search Console.
• How much of your “direct” traffic is actually assisted: people who already knew you from a previous recommendation.
• Which queries and pages are acting as bridges to conversion.

Shifting from “who sent me the click” to “what path did the user take before converting” is what separates perception from data.

How we measured it in Mi Seguro

Mi Seguro is a case where attribution breaks fast: users research, compare, leave, and come back. Only looking at the last click is self-deception. What actually works:

• Split landing pages by intent (car, motorcycle, city-specific, model-specific) and measure conversion per page, not in bulk.
• Track real origin by combining referrer + UTMs + Search Console, don’t trust GA4’s “direct” label.
• Measure assisted conversions using short windows (e.g., 7 days): how many conversions had prior touchpoints from Google, an LLM, or social media.

With this, you stop arguing “ChatGPT vs Google” and start seeing what really moves the business.

The minimum useful setup to avoid insanity

You don’t need a data warehouse to do this. Start with:

1. Landing pages split by intent, each page has a measurable goal.
2. Real-origin conversion tracking, combine referrer, UTMs, and Search Console, not just “direct”.
3. Assisted conversion windows, even 7 days, to capture prior touches before conversion.

The conclusion is boring (and more valuable)

When you measure properly, you almost never find “one magical channel giving you 40%.” Instead, you find something less sexy but more real: it’s two or three channels combined, and they shift with intent. Someone searching “motorcycle insurance in Córdoba” doesn’t arrive the same way as someone asking ChatGPT, “What’s the best car insurance?”

That’s the difference between believing a pretty metric and building on real data. For an organic acquisition product, measuring assisted attribution isn’t a luxury, it’s what tells you where to invest next month.

Want to see how Mi Seguro grew? The full case study is here. And if you’re building something where attribution keeps breaking, let’s talk.

---

By Esteban Aleart, Founder & Lead Engineer at PairProgramming.