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

When an AI assistant recommends a brand to a user who hasn't interacted with it recently, that user's subsequent behavior changes predictably. They are more likely to search for that brand, visit its website, and navigate to retailer pages. However, most digital marketing tools are blind to this transition.

Standard web tracking relies on "last-click" attribution (a method that credits only the final link clicked before a conversion). If a user asks ChatGPT for a running watch recommendation and then later types "Garmin" into Google, the search engine gets all the credit. The AI's role as the original catalyst remains unlogged and invisible.

A new study from Scrunch AI addresses this visibility gap. The researchers report that for users with no recent observed engagement, an AI recommendation increases same-name Google searches by +4.3 percentage points (pp), brand own-site visits by +2.4 pp, and brand-specific retailer-page visits by +1.0 pp .

This work provides a methodology to recover these "unlogged" exposures by joining conversational logs with clickstream data (a record of a user's web navigation).

The blind spot in digital attribution

Current web measurement assumes a traceable path of clicks. In the era of generative AI, this assumption fails. Conversational assistants act as "upper-funnel" touches (initial interactions that spark interest before the main shopping phase). As seen in, the journey typically follows a sequence. A prompt leads to a brand mention, which triggers a search, which eventually leads to a brand site or a retailer.

The problem is that the first step is a "black box" to standard analytics. Because the assistant does not typically pass a referrer header (metadata identifying the source of a visitor) to the browser, the connection is lost. Furthermore, the authors find that simply counting every brand mention in an AI chat produces biased results. Many mentions are merely incidental. For example, an AI might say, "Check your Netflix download." This captures existing customer behavior rather than new discovery. This confusion makes it difficult to determine if the AI is actually driving new interest or just riding the wave of existing habits.

Isolating the acquisition effect

To solve this, the authors implement a multi-layered identification strategy. This design strips away the noise of existing customers and incidental mentions. Their approach relies on four key technical pillars:

1. Non-customer conditioning: The researchers define a "non-customer" as an observably-unengaged user. This is someone with no recent recorded search, site visit, or retail interaction with the brand. This focuses the measurement on "acquisition" (moving new users into the brand ecosystem).
2. Stance classification: Using a small language model, the authors classify every brand mention as a "recommendation," a "neutral" name-drop, or a "caution." This distinguishes between an AI actively suggesting a product and an AI simply referencing a brand in passing.
3. Pre-trend event study: To ensure they aren't just seeing a coincidence, the authors look at the days leading up to the AI response. They found that for "all-mention" data, brand activity was often already rising before the AI spoke .

By filtering for non-customers, they remove this "pre-trend" and isolate the true lift. 4. Within-response controls: The authors compare the named brand against other brands mentioned in the same response. If the AI says, "Consider Garmin, Coros, or Polar," they compare the behavior following "Garmin" to the behavior following "Coros." Both brands shared the same user session and response. Therefore, any difference in behavior can be more confidently attributed to the specific recommendation.

Quantifying the recommendation dose

The results show that the "dose" of the recommendation matters significantly. The authors report that a genuine recommendation to a non-customer produces a much stronger behavioral response than a neutral name-drop .

Specifically, while a neutral mention might move search behavior by +1.8 pp, a recommendation moves it by +4.3 pp.

The study breaks the impact down into three distinct stages of the consumer journey: * Recall: Measured as a same-name Google search, which saw a +4.3 pp lift. This represents the user remembering and actively seeking the brand. * Discovery: Measured as a visit to the brand's own website, which saw a +2.4 pp lift. This indicates the user moved from search to the brand's direct digital presence. * Retail: Measured as a visit to a brand-specific product page on a third-party site, which saw a +1.0 pp lift. This is a "purchase-adjacent" signal of high intent.

Interestingly, the authors note that the "retail" measurement requires precision. If they only looked at the website host (e.g., amazon.com), the effect would appear near zero. Instead, they used "path-aware" matching. They looked for the brand name within the specific URL path (e.g., amazon.com/Garmin-watch) to accurately capture where the user landed .

Limits of the observational lens

While the evidence for a lift is robust, the authors define the boundaries of their findings carefully. Because the study is observational rather than a randomized controlled trial, they cannot definitively prove causality in every instance. They note they cannot fully rule out a "brand-specific within-session intent shock." This is a scenario where a user's sudden, unobserved desire for a brand happens to coincide with an AI mention.

Furthermore, the study measures "purchase-adjacent" behavior. While a visit to a retailer's product page is a strong signal, the researchers do not observe actual completed transactions. They also acknowledge that their definition of a "non-customer" is limited to what is captured in their clickstream panel. It cannot account for users who interact with brands through offline channels or closed mobile apps.

Finally, the findings regarding different AI assistants are nuanced. While ChatGPT users appear to show higher off-platform activity, the authors argue this is a matter of "composition" rather than the AI's performance. Their analysis suggests that ChatGPT users are simply more web-active individuals overall .

This means the higher lift is a trait of the audience rather than a superior recommendation engine.

The verdict: A new requirement for attribution

The verdict is clear: if you rely on traditional last-click attribution, you are likely undercounting the commercial impact of conversational AI. The study demonstrates that these assistants serve as a powerful, unlogged entry point. They drive new users into the web navigation funnel, particularly during the acquisition phase.

For practitioners, the takeaway is twofold. First, measurement frameworks must evolve to include "mention-based" tracking. This helps capture upstream exposure. Second, when evaluating retail impact, companies must move beyond simple host-matching. They should adopt path-aware URL analysis to avoid missing the subtle signals of AI-driven discovery. The "prompt-to-purchase" path is no longer a straight line. It is a search-mediated loop that begins in a chat window and ends on the open web.