The AI in your funnel might be the second opinion

When a visitor lands on your site from ChatGPT and buys something, your analytics dashboard tells a very comforting story. An AI referred a prospect, the prospect converted, and a neat little slice of attribution gets filed away.

The problem with that story is that it starts in the middle.

Long before that referral fired, the buyer may have already been on your site. Maybe they found you through an ordinary Google search, read your pitch, and paused. Instead of booking a demo or closing the tab, they took your offer to an AI assistant for a reality check: “Is this price actually reasonable?” “Will this work for a three-person team?” Or the question no marketer wants to imagine: “What’s the catch?”

If the assistant’s answer satisfies them and sends them back to your site, your reports log an "AI referral." A referral report alone won't tell you whether that visit was the buyer's first encounter with you. To the buyer, it may be a return trip—with a second opinion in between.

Once you take that pattern seriously, the standard playbook for "AI visibility" starts to look pretty thin. Businesses spend a lot of energy worrying about whether a model will volunteer their name when a stranger asks for a generic recommendation. They spend far less time thinking about what happens when someone who already knows their name asks the model to poke holes in their offer.

A study by Michael Iannelli and Alan Ai at Scrunch AI offers an intriguing glimpse into the timing of these interactions. Using February 2026 clickstream records from opted-in browsing panels in the US and Great Britain, the researchers mapped the sequence in which people used conventional search, visited web pages, and opened AI assistants.

Conventional search in this dataset followed a familiar order: people searched first, then visited content pages. With AI assistants, the flow frequently inverted. Content pageviews more often came before the first assistant event than after the last—placing the AI further down the chain of events.

Now, a good analyst should pump the brakes right here. The study defines a browsing "session" using a 30-minute window of inactivity. That tells us the sequence of clicks on a clock, not whether those clicks belonged to a single train of thought. A user could spend ten minutes pricing industrial suppliers and the next ten asking an assistant to draft a birthday speech. So the "second opinion" dynamic isn't something this dataset proves; it's a hypothesis the sequence invites us to test.

The paper also leaves a lot behind the curtain. Citing commercial sensitivity, the authors don’t share the overall panel size or the raw user counts behind specific findings. And because the tracking is limited to standalone assistant websites, it misses native apps, API integrations, and the AI summaries embedded directly inside search engines. It is a partial snapshot, and the patterns shift depending on the country and the AI platform you look at.

Even so, the underlying dynamic is hard to ignore: getting onto a buyer’s shortlist and surviving their follow-up questions are different problems.

Consider a small team weighing two software tools. They’ve read both websites. One option looks cheaper on paper, but the pricing page is vague about whether occasional collaborators require paid seats. So they ask an assistant to break down how the two plans actually compare.

At that stage, a brand mention tells you very little on its own—the buyer typed both brand names into the prompt themselves. What matters is the substance of the output. Did the assistant parse your seat limits accurately? Did it compare the actual tiers the buyer was looking at? Or did it hallucinate a licensing restriction that quietly killed the deal?

That’s a hypothetical scenario, but you don't have to guess at it. You can test how models interpret your pricing and documentation directly. You can also just ask new customers whether they bounced your offer off an AI before buying, and what they asked. Their memory won't give you a perfect click-by-click transcript, but it will surface real objections you would never uncover by endlessly prompting models for "the best software for small teams."

None of this will fix your attribution math. A referral count alone won't tell you how many times someone visited before, and an AI session that ends without a click won't tell you whether the buyer got the answer they needed, got talked out of buying, or went to a competitor. There is no clever multiplier you can apply to your referral column to make those invisible customers show up on a spreadsheet.

What you can change is the set of questions you optimize for. When someone asks an AI if your price makes sense, you don't have a discovery problem anymore. You have a sales call happening in another tab.