From Prompt to Purchase: How AI Brand Recommendations Move Consumers on the Open Web

Technical Summary: From Prompt to Purchase: How AI Brand Recommendations Move Consumers on the Open Web

Problem Statement
Conversational AI assistants (specifically text-based LLMs like ChatGPT, Claude, and Gemini) have become a routine entry point for information seeking, yet their role in the commercial journey remains invisible to standard web analytics. When an assistant recommends a brand, the user often leaves the chat to search for that brand or visit its site via organic channels. Last-click attribution models credit these downstream actions to organic search or direct traffic, rendering the assistant's "mention" an unlogged, unattributed exposure. Furthermore, existing measurement approaches often conflate genuine recommendations with incidental references to brands a user already engages with (e.g., "your Netflix download"), leading to confounded estimates of the assistant's true incremental impact.

Methodology
The authors utilize an observational panel dataset that joins opt-in pageview clickstream data with the same users' conversations with ChatGPT, Claude, and Gemini. The study employs a rigorous causal inference framework to isolate the effect of AI-generated brand mentions:

• Unit of Analysis: A (user, response, brand) row for a curated lexicon of high-recognition consumer brands.
• Outcome Variables: Three downstream behaviors measured over a 7-day post-response window:
  1. Recall: A Google search query containing the brand name.
  2. Discovery: A visit to the brand's own website.
  3. Retail: A visit to a retailer product page where the brand is identified in the URL path (not just the host).
• Identification Strategy: The study uses a matched backward-placebo design, comparing post-response behavior against the same user's behavior in windows 14, 21, and 28 days prior.
• De-confounding Controls: To separate the true "acquisition" effect from noise, the authors apply several filters:
  • Non-Customer Conditioning: Restricting analysis to users with no recent observed engagement (search, own-site, or retail) with the brand in the prior weeks.
  • Stance Classification: Using a small language model to distinguish between "recommendations" and "incidental name-drops" (neutral/cautionary mentions).
  • Within-Response Controls: Comparing the named/recommended brand against same-category brands that the response did NOT name (unnamed same-category brands), which share the same category and session context but do not receive the brand exposure.
  • Pre-Trend Event Studies: Verifying that behavior does not rise prior to the response (ruling out existing customer episodes).
• Retail Measurement: The authors emphasize "path-aware" matching (identifying the brand in the URL path, e.g., /Garmin-.../dp/) rather than host-only matching, which they found to be a near-null estimator for product-page visits.

Key Results
The study isolates a clean, acquisition-like effect where AI recommendations move observably-unengaged users into open-web brand navigation:

• Incremental Lift: For a user with no recent observed engagement, a brand recommendation leads to:
  • +4.3 percentage points (pp) increase in same-name Google searches.
  • +2.4 pp increase in visits to the brand's own site.
  • +1.0 pp increase in visits to brand-specific retailer pages.
• Stance Sensitivity: The effect is dose-dependent. Incidental name-drops (neutral mentions) to non-customers move behavior significantly less (+1.8/+1.1/+0.3 pp), suggesting that the intent of the mention matters.
• Specificity: The lift is specific to the named brand. In the same response, unnamed same-category brands show negligible movement (+0.6/+0.3/+0.3 pp), ruling out general category intent or session-wide activity bursts as the primary driver.
• The "Pooled" Fallacy: Naive pooling of all mentions (including existing customers and incidental drops) yields a misleadingly attractive but confounded funnel (+2.08/+2.37/+0.52 pp). Event studies reveal that for these pooled groups, brand activity often rises before the response, indicating the user was already in an active engagement episode.
• Navigation Path: The downstream path is primarily search-mediated. Users do not typically click links within the chat; instead, they read the brand name, search for it, and then navigate to own sites or retailers. The sequence is not strictly linear; own-site and retailer visits often occur as parallel destinations following a search.
• Audience Composition: While ChatGPT shows higher raw lift than Gemini, within-user analysis suggests this is due to audience composition (ChatGPT users are generally more web-active) rather than a causal difference in the assistant's interface.

Significance and Claims
The paper claims to provide the first observational measurement of how conversational AI moves consumers from prompt to purchase-adjacent actions on the open web. Its primary contributions are:

1. Measurement of the Unlogged: It demonstrates that assistants create brand exposures that standard web logs cannot attribute, effectively creating a "blind spot" in current attribution models where the assistant acts as an upstream touchpoint.
2. Methodological Decomposition: It offers a framework for separating genuine acquisition effects from confounding factors like existing customer behavior and incidental mentions, using stance classification and non-customer conditioning.
3. Acquisition-Like Reach: The measurable effect is concentrated among users with no recent observed engagement. For existing customers, the mention coincides with an active episode but does not clearly accelerate behavior beyond the existing trend.
4. Practical Implications: The authors argue that last-click attribution misallocates credit to organic search and direct visits. They posit that a "purchase-adjacent" retail visit (product page view) is the observable proxy for purchase in this context, and that measurement systems must account for the "prompt-to-search" bridge to accurately value AI-mediated journeys.

Limitations
The authors explicitly note that the study is observational and does not observe transactions; "retail" is defined as a product-page visit, not a sale. They acknowledge that a brand-specific "intent shock" within a session (where a user's internal intent drives both the prompt and the subsequent action) remains unidentifiable without a randomized design. Additionally, the "non-customer" definition is behavioral (no recent observed clicks) rather than commercial (never owned the product), meaning the effect is a lower bound on true acquisition.