Measuring AI search in five layers, because attribution alone will not do it

The decision is now often made before anyone reaches your website. A B2B buyer asks ChatGPT which vendors to consider, uses Gemini to weigh up implementation approaches, checks a technical claim against Perplexity, and only then opens a tab for the two or three companies that survived. Everything that mattered happened somewhere your analytics cannot see.

The scale is not marginal. Green Hat's 2025 research into B2B buying journeys found as much as 94% of the people involved in a purchase using language models during selection — to validate, to summarise, to confirm a decision already forming. Wynter's 2026 figures put around 84% of CMOs using AI tools while discovering vendors in the first place.

That leaves SEO reporting with an awkward gap. The metrics readily available describe what happens after the visit, and the visit is now the last step rather than the first.

The temptation is to look for one number that closes the gap. There isn't one, and the reason is structural rather than technical. A buying journey now runs across search engines, AI assistants, communities, analyst coverage, peer recommendation and direct contact, and no single touchpoint can honestly be named as the cause of a sale. A framework set out in Search Engine Land proposes measuring influence in five layers instead, each answering a different question.

The first layer is access, and it comes before everything else because nothing can be recommended that has not been read. If models cannot reliably reach your pages, or reach them and find nothing worth retrieving, the remaining layers have nothing to measure.

The practical work is monitoring verified AI crawler activity — how often it arrives, how deep it goes, how much of the site it covers. Rising bot traffic guarantees no visibility whatsoever, but it does indicate that systems are finding your content and coming back to it.

One caution matters more here than it does with conventional search crawlers. User-agent strings are trivially forged, so a log full of confident-looking AI bot names proves nothing by itself. Verify against published IP ranges, reverse DNS, or whatever verified-bot service your CDN provides. The objective is not a large request count; it is confidence that the systems doing the requesting are the ones you think they are.

The second layer asks whether that access turns into use — mentions, citations, appearances in generated answers. This is where measurement most often goes wrong, because the natural instinct is to run a handful of prompts after publishing something and see what comes back. That produces anecdotes rather than a trend.

The fix is a standing prompt library built around the questions buyers actually ask across the journey, and then leaving it alone. Stability is the entire point: a fixed set of prompts run repeatedly shows movement, while a set that keeps changing shows noise.

Several metrics each contribute a partial view — mention rate, citation rate, Search Console impressions from AI Overviews, AI Mode, and grounding queries in Bing Webmaster Tools. None of them answers the question alone. Together they get at something usable: when buyers ask AI about this market, does our name come up?

The third layer is the first one traditional analytics can touch at all. GA4 will record referral traffic from recognisable assistants when somebody actually clicks through, and that behaves much like organic or paid traffic in a report.

It is essential to be precise about what this excludes. AI Mode and AI Overviews are not counted as AI traffic. They generally fold into Google organic, and depending on how the click is passed they can land in Direct. So the surfaces most likely to shape a buyer's shortlist are among the least visible in the report.

Beyond that, a great deal of AI influence never produces a click at all. Someone reads about you in a generated answer, does nothing, searches your brand name four days later, and arrives by a route carrying no trace of where the idea came from. Revenue credited to AI referrals is therefore one input to the picture, not a verdict on what AI contributed.

The fourth layer is where that missing influence becomes observable indirectly. A buyer who first met you through an AI surface tends to return through branded search, a direct visit or an email, and converts under a label that says nothing about the original encounter.

Watching branded clicks in Search Console alongside branded organic conversions in GA4 gives you a way to see the shape of it. No individual branded conversion proves AI created the opportunity — but when visibility gains and branded demand rise together, repeatedly, the case stops being speculative.

The fifth layer is the one that was always there. Pipeline, closed-won business and revenue decide whether any of this was worth doing, and AI changes none of that.

Which is the useful conclusion. AI has not altered which outcomes matter; it has altered the evidence available for showing progress towards them. Presented as separate numbers, the four layers underneath read as activity metrics. Presented in sequence they read as an argument: systems reached the content, visibility inside AI answers rose, branded demand followed, attributable AI revenue grew, and pipeline moved.

No link in that chain proves causation, and it is worth saying so plainly in front of an executive rather than being caught later having claimed more than the data supports. What the chain does offer is a body of evidence about channels conventional attribution cannot see into — which, for the moment, is the honest maximum.