A prospect spends twenty minutes with an AI assistant comparing four suppliers, forms a clear preference, then searches your company name and fills in a contact form. Your analytics records a direct or branded organic conversion. The twenty minutes that actually decided it are invisible, and so is whatever content the assistant drew on.
This is not a tracking configuration problem. It is a structural mismatch between how measurement works and how research now happens.
The five assumptions that no longer hold
- Journeys are linear. Models assume a sequence of touches leading to a conversion. AI-mediated research is exploratory and often collapses several stages into one session.
- Influence produces a click. An assistant can describe your positioning accurately without the buyer ever visiting your site. Influence occurred; no session exists.
- Channels are identifiable. Referral data from AI interfaces is inconsistent, frequently absent, and often resolves to direct traffic.
- Content value equals content traffic. A page that an assistant reads and synthesises may generate no pageviews while shaping every answer given about your category.
- Consideration is observable. Much of it now happens in a private conversation with a system you do not control and cannot instrument.
How each metric is affected
| Metric | What happens | Direction of error |
|---|---|---|
| Sessions and pageviews | Research completed without a site visit | Understates reach |
| Last-click attribution | Credits the branded search that followed the decision | Misattributes cause |
| Multi-touch attribution | Cannot include touches it never observed | Systematically incomplete |
| Channel reports | AI referrals resolve to direct or are absent | Inflates direct |
| Content performance by traffic | High-influence, low-traffic pages look like failures | Penalises the most useful content |
| Keyword rankings | Position on a results page is less relevant when no page is viewed | Measures the wrong surface |
| Cost per lead by channel | Denominators are wrong because sources are misassigned | Distorts budget allocation |
| Time to conversion | Compresses because early research is unobserved | Understates true cycle length |
What still works
The pessimism should be bounded. Several measures remain sound because they sit at points you still control.
- Enquiry volume and quality. Whatever happens upstream, the enquiry lands in your systems.
- Self-reported source on forms. Crude, free-text and inconsistent — but it is the only place a buyer can tell you they used an AI assistant, and enough of them do to be informative.
- Conversion rate from enquiry onward. Everything after first contact remains fully observable.
- Cohort analysis. Comparing groups of accounts created in different periods survives attribution failure entirely.
- Sales conversation evidence. What prospects already believe when they arrive is direct evidence of upstream influence.
Proxy signals worth tracking
| Signal | How to capture it | What it indicates |
|---|---|---|
| Branded search volume | Search Console impressions for your brand terms | Category research resolving into name recall |
| Direct traffic to deep pages | Analytics: direct sessions landing on non-homepage URLs | Someone was told specifically where to look |
| Self-reported "AI assistant" mentions | Free-text source field on enquiry forms | Direct, if partial, evidence of the channel |
| Prospect knowledge at first call | A single field the rep completes: how well-informed was this buyer? | Upstream education quality |
| Assistant answer audits | Ask several assistants your category questions monthly; log whether you appear and how you are described | Whether you are referenced, and whether accurately |
| Unlinked brand mentions | Monitoring across relevant public sources | Corroboration signals assistants can draw on |
| Enquiry specificity | Whether the enquiry names a specific service versus a general request | Depth of prior research |
A measurement approach that survives the gap
- Accept that upstream attribution is partially unrecoverable and stop optimising against a number you know is wrong.
- Move the primary measurement point to enquiry and beyond, where observation is reliable.
- Track the proxy set above monthly, as a trend rather than an absolute.
- Judge content on whether it answers a real buyer question completely, not on its pageviews.
- Report cohorts, not calendar periods, so that mix changes do not masquerade as performance changes.
Limitations of this approach
Proxies are correlational. Branded search can rise for reasons unrelated to AI. Self-reported source data is sparse and biased toward buyers who remember. Assistant audits are non-deterministic — the same question can produce different answers on the same day, which means a single audit tells you little and only the trend across months is meaningful. None of this reconstructs the lost attribution; it establishes whether the direction of travel is right.
Key takeaways
- Attribution assumes clicks and linear journeys; AI-mediated research provides neither.
- Last-click reporting credits the branded search that followed the decision, not what caused it.
- Measurement from enquiry onward remains fully reliable — move your primary metrics there.
- Track branded search, deep direct traffic, self-reported source, prospect knowledge and assistant audits as a proxy set.
- Read proxies as trends over months; they are correlational and individually noisy.