A buyer asks an AI system which vendors fit their problem, gets a useful answer, and starts forming a shortlist. Nothing appears in your analytics. No click. No session. No attribution fight to lose.

That is not an argument that the website no longer matters. It is an argument that the buying journey has acquired a quiet, imperfectly observable surface. Executive teams now need to measure it without inventing precision they do not have.

There is evidence that click behavior changes when answers arrive in the results page. In a Pew study published July 22, 2025, traditional result links were clicked in 8% of visits with Google AI summaries, versus 15% without. The study observed 900 US adults in March 2025. It was observational—not B2B-wide, and not proof that the summary caused the difference. Still, it describes a measurement problem worth taking seriously.

Last touch is not dead. It is incomplete.

Last-touch attribution remains useful for answering a narrow question: what did a known visitor do immediately before converting? Keep it. But it cannot describe an influence path that happens inside a model interface, in a private evaluation, or before the visitor identifies themselves.

Nor is “AI visibility” pipeline. A mention can be wrong, stale, irrelevant to your ICP, or disconnected from a buying motion. Forrester’s February 12, 2026 perspective makes the larger point: buyer AI evaluation also happens internally, beyond public discovery. The work is to assemble evidence, not crown a dashboard metric.

Five signals, each with a boundary

1. Sampled AI mentions and citations

Create a stable set of prompts by job, problem, and category. Run the same set on the same platforms on scheduled prompt dates. Record whether your brand is mentioned, cited, and positioned accurately. Your share is brand-mentioned runs ÷ eligible sampled runs. It is a share of your sample, not market share. Preserve the raw responses; models and rankings change.

2. Self-reported AI discovery

Recommend a fixed, optional source question in prospect intake: “Search engine,” “AI assistant,” “Peer,” “Event,” “Partner,” “Other,” or “Prefer not to say.” Do not force an answer or let salespeople recode it later. This is a recommended instrument, not a claim that Sophizo has implemented tracking for every client.

3. Verified AI crawler access

Ask the infrastructure team to verify known AI crawler access in server logs, using documented verification methods. Report it separately from human sessions and conversions. A crawl is evidence that a system could access a resource—not proof it used, understood, or recommended it, and never proof of buyer intent.

4. Branded and direct traffic

Watch these as supporting signals alongside campaign activity, PR, product releases, and seasonality. A rise may be consistent with discovery; it is not causal evidence. Direct traffic is especially noisy: it includes bookmarks, privacy-preserving referrers, and mislabeled channels.

5. Matched pipeline cohorts

Compare prospects exposed to the same motion, segment, geography, and sales capacity where feasible. Log differences in pricing, campaigns, rep coverage, product maturity, and time. This is stronger than a simple before/after chart, but still has confounding limits. Do not call it causal incrementality without a design capable of supporting that claim.

A practical 30-day protocol

Days 1–5: define the ledger. Select 20–30 decision prompts that represent real executive questions, identify the platforms to sample, assign prompt dates, and define eligibility before the first run. Establish a one-page coding guide: mention, citation, accuracy, competitor set, and response type.

Days 6–14: capture the baseline. Run and save the stable sample. Add the optional intake question prospectively; do not backfill guesses. Set a separate log view for verified crawler activity. Take a timestamped baseline of branded/direct traffic and open pipeline by cohort.

Days 15–23: inspect the substance. Read the output, not just the count. Are you named for the right problem? Are the cited sources current? Which competitors appear in eligible runs? Fix factual gaps in the content that humans and systems can actually access. Google says AI Overviews and AI Mode traffic are included in Search Console’s Web reporting; it also says no special AI schema or file is required and appearance is not guaranteed.

Days 24–30: publish a restraint-first readout. Show the five signals side by side, annotate what changed, and list alternative explanations. Decide only what the evidence supports: continue sampling, improve a source asset, test an intake wording, or hold. The point is not to prove that AI created revenue in thirty days. It is to establish a defensible operating cadence before someone asks for a bigger claim.

What to say in the boardroom

“We are measuring AI discovery as an influence layer. We can see sampled presence, declared discovery, verified access, supporting demand movement, and cohort outcomes. We cannot yet assign all movement to AI, and we will not pretend otherwise.”

References

Build the measurement layer into the work: modernize pipeline, assemble a signal stack, and redesign the operating model around evidence rather than anecdotes.

Your next buyer may never visit your website.