Direct answer
AI-assisted discovery often breaks the ordinary referral trail. A user may see a brand in an answer, remember it, search for it later, visit through a direct URL, or contact sales without a referrer that identifies the AI interaction. A visible AI click is therefore one measurable event, not the full value of a citation or answer. Measure the chain in separate layers: exposure, source use, site visit, qualified action, and commercial outcome. Use controlled prompts, landing-page signals, self-reported discovery, and CRM evidence where lawful and practical. Do not convert missing attribution into an invented AI conversion rate.
The visitor who arrived without a referral
A procurement manager asks an assistant for European suppliers of a specialized component. The answer links to a manufacturer. The manager reads the answer on a phone, later searches the company's name from a work laptop, visits the homepage directly, and submits a request for a quote two weeks later.
The CRM records a direct visit. Analytics may record organic search or no usable campaign source. The link in the AI answer was real, but the final conversion is not visibly connected to it. If the team measures only click-throughs, it misses the path. If it labels every later branded visit as AI-driven, it overstates the evidence.
This is the broken referral trail: the user's decision path continues after the observable source link disappears. The solution is not a magical attribution model. It is a measurement design that shows which links are directly observed, which relationships are inferred, and which outcomes remain unknown.
Separate the events before joining them
Exposure is an answer or search feature displaying a page, brand, or claim. Source use is the page being cited or linked as support. Visit is a user reaching a site or landing page. Qualified action is a defined event such as a completed technical inquiry. Commercial outcome is a later opportunity, order, or contract under the organization's records.
These events can occur without one another. A page can be cited without a click. A user can click without staying. A direct visit can follow an unobserved answer. A form can be submitted without becoming a qualified opportunity.
Keep the event definitions separate in the data model. A column called “AI conversion” invites false certainty because it compresses a chain with missing links into one label. Use fields such as observed_ai_source, observed_click, self_reported_ai_discovery, qualified_inquiry, and attribution_confidence.
The Pew Research Center study observed 8 percent traditional-result link visits on pages with an AI summary versus 15 percent on pages without one, and 1 percent clicked a source link in the summary itself. It also found session endings of 26 percent versus 16 percent. This observational study describes search behavior in its sample; it does not identify a particular company's AI revenue or establish that every unclicked exposure influences a later purchase.
Official metrics have defined boundaries
Google's generative AI performance report reports impressions in supported Search AI features and groups data through specified property, page, country, date, and device dimensions. Its chart and page table use different aggregation rules. That report can show measured exposure under Google's definitions. It does not show all AI answers or prove that an exposure changed a buyer's decision.
Bing's AI Performance documentation defines Total Citations, Average Cited Pages, grounding queries, and page-level citation activity. It states that these measures do not indicate ranking, authority, or placement within an individual answer.
Treat those official measures as named layers in the chain. Do not add Google impressions to Bing citations, prompt-panel mentions, and CRM inquiries to create a single cross-platform numerator. The sources use different populations, units, and observation rules.
Ahrefs reports that its December 2025 study found a correlation between AI Overviews and a 58 percent lower average click-through rate for the top-ranking page, using 300,000 keywords split into groups. It is an external observational analysis, not a universal causal estimate or a count of all site clicks. Use it to understand why click-based reporting may change, not to assign a fixed loss to one brand.
Build an attribution ladder
Attribution should become weaker as the path contains more unobserved steps. Make the levels explicit.
| Level | Evidence | Appropriate statement |
|---|---|---|
| Direct link | Analytics or platform record connects a click | “A tracked click arrived from the observed source.” |
| Tagged landing visit | User reaches a campaign URL with a declared tag | “The tagged visit is associated with this campaign.” |
| Self-reported discovery | User says an AI answer influenced research | “The respondent reports AI-assisted discovery.” |
| CRM-assisted inference | Timing, brand search, and records suggest a path | “AI may have assisted; the evidence is indirect.” |
| Unobserved exposure | Brand appeared but no later path is known | “Exposure was observed; downstream value is unknown.” |
These labels prevent an untracked direct visit from being upgraded to a confirmed AI conversion. A weaker attribution is still useful if the limitation is visible.
Add evidence without contaminating the user journey
Use campaign landing pages or link parameters when the platform and site permit them and when the parameters do not make the URL misleading. Track the page, date, content version, market, and intended campaign. Do not assume every assistant preserves parameters or that a user will click the same URL later.
Ask a short, optional discovery question on an inquiry form: “How did you first hear about us?” Include AI assistant as one choice and allow a free-text answer. This is self-reported evidence, not a perfect truth source. Avoid forcing users to attribute a journey they cannot remember.
Use CRM fields for first-touch and influence separately. A salesperson may learn that an AI answer shaped the shortlist even when the last-touch source is direct. Record the evidence and confidence rather than changing the last-touch channel.
For privacy and compliance, collect only what the organization has a lawful and transparent reason to use. Do not fingerprint users to reconstruct a hidden assistant journey. An attribution system that damages trust can cost more than the measurement it produces.
A measurement workflow for a broken trail
- Define the business event that matters, such as a qualified RFQ rather than a pageview.
- List the directly observable AI and site signals available to the organization.
- Add a self-report or sales-record field for assisted discovery where appropriate.
- Freeze campaign and content versions so the same link can be interpreted later.
- Join events only when time, entity, market, and privacy rules support the join.
- Assign an attribution level and retain the evidence behind it.
- Report direct, assisted, and unknown outcomes separately.
- Compare changes over time with controls or alternative explanations.
The workflow is a proposed measurement design. It does not create missing data. If the site has no reliable referral record, publish that limitation and improve future capture rather than backfilling historical AI influence from assumptions.
Use controlled tests for influence, not just clicks
A prompt panel can test whether a page is cited or whether an answer contains a supported recommendation. It cannot prove that a human changed a buying decision. A user study, controlled landing-page experiment, or CRM comparison may add evidence, but each has its own confounders.
The foundational GEO paper measures visibility in a controlled benchmark and reports domain-specific gains. It does not establish organic traffic or revenue. The critical survey finds that reviewed techniques do not establish a stable longitudinal cross-platform causal effect on downstream behavior. That is a reason to separate the business question from the citation question.
Design a test around a decision. For example, compare qualified inquiries from matched markets after a content intervention, while logging exposure and other marketing activity. If a randomized design is impossible, use a documented comparison and state residual confounding. A later inquiry rise may be consistent with AI influence without proving it.
Make direct traffic less ambiguous going forward
A brand can improve the odds of learning how a buyer arrived by making the next step explicit. A landing page can explain the service, offer a market-specific form, and ask the discovery question. A sales team can record the sources mentioned in a buyer's shortlist. A product page can use a stable URL that remains identifiable after a redirect or migration.
Do not add “AI” to every page title or URL merely for attribution. The page should serve the buyer first. Use a content version ID or campaign register behind the scenes, where it does not confuse readers.
Keep an “unknown” bucket operationally useful. It should record why the path is unknown: direct visit with no self-report, missing analytics consent, a copied URL, an offline introduction, or a CRM record that lacks the discovery field. Different unknowns suggest different improvements. A copied URL may be addressed by a memorable landing page; a missing consented analytics event may require a privacy-safe measurement review. Treating all unknowns as one failure prevents the team from learning where the trail breaks.
Review attribution rules with sales and finance before using them for budget decisions. A marketing report may count assisted discovery, while a finance report counts only booked revenue. Both can be correct for their purpose. The handoff between them should show the event definition and confidence rather than force the commercial system to accept an unverified AI label.
For Xindar's US, UK, and European work, retain the market and language of the first observed exposure where available. A UK buyer who later contacts a global sales team should not be silently counted as a US outcome simply because the CRM owner is in the United States.
Frequently asked questions
Does a citation without a click have no value?
- It can contribute to awareness, verification, or a later unobserved journey. The evidence only shows that a click was not observed. Report the downstream value as unknown or assisted when supported by other evidence.
Can direct traffic be counted as AI traffic?
Not from the channel label alone. Use self-report, a tracked link, or a defensible assisted-attribution rule, and state its limits.
Should every AI answer link use tracking parameters?
Only when the platform and user experience support it and the organization has a legitimate measurement purpose. Parameters can be stripped, copied, or ignored. They cannot repair every broken trail.
What should be shown to a client?
Show direct source events, clicks, self-reported influence, qualified actions, commercial outcomes, and unknowns separately. Include the join rules and evidence level behind any assisted claim.
Source and method note
Official reporting documentation and external studies were retrieved on September 21, 2026. The Pew and Ahrefs findings retain their samples and observational boundaries. Scenarios, attribution ladder, field names, and workflow are editorial proposals. No customer conversion, revenue, or hidden referral path was measured, and no named human review is claimed.