The practical brief
The finding
Content pageviews more often followed conventional search but preceded standalone assistant use; assistant sessions were also more often contained within the observed assistant surface. [c1] [c2] [c3] [c7] [c9] [c10] [c11]
Why it matters
AI referral counts cannot capture browsing that happened before assistant use, although the study does not establish that earlier pages influenced an answer or purchase.
| Observed pattern | Standalone assistant sessions | Non-assistant search sessions |
|---|---|---|
| Content before first anchor event | 45.5% of February assistant sessions, user-weighted (95% interval: 44.9–46.1%) | 46.7% of February qualifying search sessions, user-weighted (95% interval: 46.4–47.1%) [c2] |
| Content after last anchor event | 38.9% of February assistant sessions, user-weighted (95% interval: 38.4–39.5%) | 60.2% of February qualifying search sessions, user-weighted (95% interval: 59.9–60.6%) [c2] |
| Contained within the constructed session | 34.1% of February assistant sessions, user-weighted: no external search or content step (95% interval: 33.5–34.7%) | 19.5% of February qualifying search sessions, user-weighted: no content pageview (95% interval: 19.2–19.8%) [c3] [c4] |
What to try
BLURSOR’s practical interpretation
Our interpretation: retain AI referrals as a traffic measure, not a complete measure of AI-related discovery, and do not equate absent onward visits with satisfaction.
Study boundary
The study measures event order, not brand visibility or conversions, and cannot separate information seeking from producing work inside an assistant. Exact rates vary by market, provider, and measurement rules.
Do not make referrals your whole AI scorecard
If you are deciding whether AI helps customers discover your business, avoid treating assistant referral traffic as the whole story. Browsing can precede assistant use rather than follow it. This study supports that measurement caution—not a claim that your pages influenced an answer.
Michael Iannelli and Alan Ai studied linked assistant, search, and browsing activity from an AI-forward, opt-in panel in the United States and Great Britain. Their findings concern where standalone assistants sit in observed activity, not which businesses win visibility.
The study linked events, not validated customer tasks
The researchers combined the same users’ prompts, responses, searches, and content pageviews into time-ordered sessions, separated by 30 minutes of inactivity. These “temporal sessions” are measurement windows: nearby events need not belong to one shopping decision or information need.
The main analysis covered February 2026 and repeated the construction in March. It measured standalone assistant websites, not native mobile apps, API use, or search-embedded answers such as Google AI Overviews and AI Mode.
The privacy-safe extract excluded conversation text and detailed page content. Consequently, the analysis cannot distinguish information seeking from workbench production—using an assistant to produce work rather than seek information. Calling every observed session a discovery or customer journey would overstate the evidence.
The comparator was these adopters’ search sessions without assistant events, not their pre-AI behavior. User-weighted shares give each contributing person equal weight rather than letting frequent users dominate.
Browsing often came before the assistant
Around search, content appeared after the last search in 60.2% of sessions, versus 46.7% before the first. Around assistants, content appeared before the first assistant event in 45.5%, versus 38.9% after the last. These are user-weighted shares; a session can contain both before and after activity.
Among people contributing both session types, the paired difference between those directional patterns was 20.6 percentage points, with a 95% uncertainty interval of 19.9–21.3 points. The result repeated in March, but replication in the same panel is not independent-population validation.
No single assistant pattern dominated: user-weighted, 34.1% of sessions had no observed external search or content step, 37.1% mixed web activity around or between assistant events, 18.3% had web activity only before, and 10.5% only after.
These pooled rates are not typical rates for individual markets or assistants. Containment varied substantially across market and provider splits. The directional ordering persisted, but its magnitude varied too; Gemini’s assistant-side asymmetry was small.
What this means—and what it cannot establish
Our interpretation is to distinguish traffic arriving from assistants from possible AI-related activity elsewhere in a journey. An assistant-first funnel is an incomplete default. Earlier browsing does not prove that an assistant read, synthesized, or discussed those pages.
“Contained” does not mean satisfied. Assistant containment meant no observed external search or content step; search containment meant no content pageview. The paired within-user difference was 13.0 percentage points, not simply the subtraction of the headline averages.
Containment also depends on measurement choices. Session-weighted assistant containment fell from 46.9% with a 15-minute inactivity gap to 29.2% with a 60-minute gap. Neither definition establishes task completion.
The study cannot tell you whether AI displaced search, improved credibility, named your brand, or generated sales. It tested no marketing intervention.
- Label AI referrals as observed arrivals, not total AI influence.
- Treat missing onward visits as missing observed activity, not proof of resolution.
- Do not reallocate channel budgets solely from these descriptive patterns.
The source and evidence
The New Shape of Search: How Conversational AI Recomposes Information Seeking
Study limitations and disclosures
- The AI-forward, opt-in US and Great Britain panel covers two adjacent months, not a representative sample of all customers. March replication uses the same panel. Pooled containment levels and directional magnitudes vary substantially by market and provider, even where the ordering persists.
- Conversation content was excluded, so the analysis cannot distinguish information seeking from workbench production. Observed assistant sessions are not validated discovery or customer journeys.
- Inactivity-defined sessions do not establish shared intent, satisfaction, completion, verification, or causal search displacement. Containment changes with session boundaries and observation density; different coverage thresholds also select behaviorally different users.
- Native mobile apps and API use were unobserved. Search-embedded AI was excluded from assistant activity but could appear in comparator searches. Coarse domains, capture timing, unrecognized assistant hosts, and excluded captured events can affect classification. Onward visits were timing-linked, not observed clicks, and do not establish reading depth.
- Both authors, Michael Iannelli and Alan Ai, list Scrunch AI affiliations. Aggregate panel size, analysis-user counts, and cell counts are withheld as commercially sensitive; raw conversations, searches, and individual URLs are not released. These restrictions limit independent assessment and reproducibility. The manuscript provides no separate funding disclosure.
- The manuscript is under review and not yet peer reviewed. Panelists consented under the provider’s governance process; the secondary analysis was not submitted for separate institutional ethics review.
Evidence behind this briefing
[c1] The researchers combined same-user prompts, responses, searches, and content pageviews into sessions separated by 30 minutes of inactivity. These are temporal proxies, not validated tasks; the privacy-safe extract excluded conversation text and detailed page content.
§3, Session reconstruction with assistant events as steps · Read the study
[c2] Content more often followed conventional search but preceded assistant use. Among users contributing both session types, the paired difference in before-minus-after content shares was +20.6 percentage points (95% user-clustered bootstrap interval 19.9–21.3). This is temporal association, not evidence that assistants synthesized prior pages.
§6, The mirror · Read the study
[c3] Assistant sessions were more often contained than the same adopters’ non-assistant search sessions: user-weighted shares were 34.1% versus 19.5%, with a paired within-user difference of +13.0 percentage points (95% interval 12.5–13.6). Assistant containment means no observed external search or content step; search containment means no content pageview.
§6, Within the same user · Read the study
[c4] No single assistant journey dominated: user-weighted shares were 34.1% contained, 37.1% bridge/interleaved, 18.3% web-only-before, and 10.5% web-only-after, with 95% user-clustered intervals. Marketers should not treat all standalone assistant use as an initial discovery step followed by a click.
§4, opening paragraph; Table 2 · Read the study
[c5] The main contrasts repeated in March in the same panel: the paired containment difference was +13.6 percentage points (95% interval 13.0–14.1), and the paired directional difference was +20.3 points (19.6–20.9). This is adjacent-month replication, not independent-population validation.
§6, Replication in a second month · Read the study
[c6] Containment is sensitive to session construction and observation density, and does not establish satisfaction or task completion. Session-weighted containment fell from 46.9% at a 15-minute gap to 29.2% at 60 minutes; more active-day requirements also selected behaviorally different populations.
§10, observation density and session boundaries; Tables 3–4 · Read the study
[c7] The design cannot establish shared task identity, successful resolution, verification, causal displacement, or comparative assistant credibility. Search and assistant sessions select different needs; coarse destination-domain robustness is not semantic task matching.
§8, Boundaries of the claims · Read the study
[c8] Sample sizes and raw data are withheld for commercial/privacy reasons, limiting independent assessment of the denominators. The manuscript is not peer reviewed; both authors are affiliated with Scrunch AI, and the secondary analysis was not submitted for separate institutional ethics review.
§3, Disclosure and reproducibility; front matter; Ethics and Human-Subjects Statement · Read the study
[c9] The analysis cannot distinguish information seeking from workbench production because prompt-content classification was intentionally excluded. Observed assistant sessions therefore should not all be described as discovery or customer journeys.
§8, Boundaries of the claims · Read the study
[c10] Within the February 2026 panel, containment levels varied substantially by market and provider despite preserving the assistant-versus-search containment ordering. User-weighted assistant containment was 50.9% in the United States versus 20.7% in Great Britain, and 48.8% for Gemini versus 27.3% for ChatGPT. Provider splits used single-assistant sessions; Perplexity was too thin to split, and search sessions were attributed to each user's most-frequent provider. These pooled levels should not be treated as typical market- or provider-specific rates.
§6, Stability within the window · Read the study
[c11] The directional ordering persisted across the observed market/provider splits, but its magnitude was contextual. Gemini's user-weighted assistant-side before-minus-after content difference was only +1.2 percentage points, with a 95% user-clustered interval of +0.2 to +2.2: positive, but small and close to zero at its lower bound.
§6, Stability within the window · Read the study