How AI search is changing customer discovery

Marketing has a mental model problem. Most teams still picture discovery as a funnel they can observe: impressions they can count, clicks they can attribute, landing pages they can test. Every stage leaves a trace in an analytics dashboard.

AI search breaks the model not by shortening the funnel but by relocating it. The comparison table your prospect used to build in a spreadsheet now gets built by Gemini in four seconds. The "best X for Y" listicle they used to read gets composed on demand, personalized to the exact constraint they typed. All of it happens before your analytics ever see a visitor.

The invisible middle

Call it the invisible middle of the journey: everything between "I have a problem" and "I am on your pricing page." That middle used to be fought over in public, on search results pages and review sites you could monitor. Now a growing share of it happens inside a model's context window, and the only observable artifact is the visitor who arrives already decided.

10+questions a buyer asks an assistant before their first site visit

2 to 3brands that survive into a typical AI-generated shortlist

0impressions your dashboard records while it happens

You have probably already seen the downstream symptom without naming it: direct traffic that converts unusually well, from visitors who seem to know your product before their first session. They were not psychic. They were briefed.

What the machine actually does

To influence the invisible middle you have to understand what happens inside it. When an assistant answers a commercial question, it blends two things: what the model already believes about your category from training, and what it retrieves live from the web to ground the answer. Both are shaped by the same underlying material.

Parse intentConstraints extracted: budget, team size, use case

→

Recall entitiesWhich brands the model associates with the category

→

Retrieve + groundLive sources fetched: reviews, comparisons, docs

→

ComposeTwo or three names, with reasons attached

Notice what carries weight in that pipeline. Entity recall depends on how often and how consistently your brand appears next to your category across the public web. Grounding depends on whether credible third-party sources describe you accurately. Composition favors content that hands the model a clean, quotable claim: who you are for, what you cost, why you win.

The audit most businesses have never run

Here is the exercise we run in the first week of every engagement. Take your ten highest-intent commercial queries. Ask them, phrased naturally, across ChatGPT, Claude, Gemini and Perplexity. Record three things for each answer: were you named, how were you described, and which sources got cited.

The results are usually humbling in a useful way. Companies discover their category description is five years stale because the model learned it from an old crunchbase page. They discover a competitor owns the phrase "best for small teams" because of two well-structured comparison posts. They discover review sites they stopped monitoring in 2022 are the primary grounding source for their entire category.

Discovery you can shape, not track

The honest conclusion is that a share of customer discovery is no longer directly measurable, and pretending otherwise leads to bad strategy. What you can do is shape the inputs: publish the comparisons buyers ask for before they ask, keep every public description of your business consistent down to the phrasing, earn presence in the sources models retrieve, and structure your site so a machine can lift answers from it cleanly.

Then measure what is measurable: your share of answer across assistants, tracked monthly like you once tracked rankings. It is a young metric with rough edges. It is also the closest thing to ground truth about how the invisible middle sees you.