Key findings
- A 2026 Scrunch study found higher one-week rates of brand search, site visits and retailer views after recommendations to new users. Its consumer results cannot be applied directly to B2B fintech.
- Mentions, recommendations, position and citations are different signals. A brand can be cited without being named or recommended without receiving a direct click.
- Geolix.ai data shows why referrals alone are incomplete: high monitored visibility on one engine coincided with only three direct sessions. Different collection windows prevent an engine ranking.
- Fintech measurement should combine AI answers, brand demand, first-party attribution and commercial outcomes without crediting the same conversion twice.
What the Scrunch study measured
Scrunch published its prompt-to-purchase study on June 17, 2026. It analyzed millions of events from a privacy-safe opt-in panel, linking anonymized AI conversations with the same users’ anonymized searches, brand-site visits and retailer product-page views from February through May 2026. The analysis compared the week after a recommendation with matched earlier weeks for the same person and excluded people who had recently engaged with the brand.
| Observed behavior among new users | Matched baseline | Week after recommendation | Reported relative lift |
|---|---|---|---|
| Google search for the brand | 3.28% | 9.24% | Approximately 182% |
| Visit to the brand website | 3.43% | 7.45% | Approximately 117% |
| Retailer product-page view | 0.94% | 2.68% | Approximately 185% |
Sources: Scrunch prompt-to-purchase study
The study separated passing mentions from active recommendations. It reported increases of 3.3 percentage points in brand search and 1.9 points in site visits after a mention, compared with 6 and 4 points after a recommendation. This supports a distinction between being present and being put forward, not a universal conversion rate. The panel was unusually AI-engaged, and the replicated categories were beauty, apparel and audio rather than fintech.
Why direct attribution misses part of the journey
Last-click systems observe sessions, not the full decision path. If a buyer sees a recommendation in ChatGPT and later searches the company name, analytics may credit organic search. If the buyer types the domain, the visit may appear as direct. If the buyer talks to sales two weeks later, the AI interaction can disappear unless the CRM or a self-reported question captures it. Peec AI makes this same argument in its public KPI guide and recommends asking customers how they discovered the company, then following the resulting accounts through revenue.
A first-party view from Geolix.ai
Geolix.ai compared AI monitoring, GA4 and Google Search Console for its own site. GA4 recorded about 403 sessions during the 12 days with usable data from July 25 to August 5, 2026. Twenty-five sessions, or 6.2%, came from identifiable AI assistant referrals. ChatGPT accounted for 18, Gemini for four and Perplexity for three.
During the post-July 28 monitoring window, Perplexity mentioned Geolix.ai in roughly 77% to 86% of tracked responses and placed it first in about 72% to 82%, yet it generated only three direct referral sessions. ChatGPT mention rates were much lower, generally 5% to 8%, but ChatGPT produced 18 referrals. Monitoring coverage began at different times and answer visibility is not comparable to GA4 sessions one for one. The evidence shows that visibility and referral clicks measure different behaviors; it does not prove which engine produced more business value.
Other signals remain non-attributable. Direct traffic represented 70.5% of sessions. Search Console recorded 44 Geolix.ai brand-query clicks, 45% of all 97 clicks in the export, while July sitewide clicks rose from 19 to 56. Multiple releases and channels occurred concurrently. Some zero-click impressions also came from Geolix.ai monitoring fan-out and must be removed before demand analysis.
Mention, recommendation and citation are not interchangeable
Commercial interpretation starts with the answer state. A citation means a page supplied information. The brand may never appear in the prose. A mention means the entity was named, but the answer may be neutral or negative. A recommendation means the assistant actively presents the brand as suitable for the user’s need. Position and framing then influence prominence, while factual accuracy determines whether the exposure is safe.
This distinction appeared in a separate Geolix.ai English-language fintech case covering nine purchase-intent questions in Singapore. Across 15,495 valid answers and 223,196 citation records collected from July 20 to 31, ChatGPT cited the target domain in 22.9% of answers, mentioned the brand in 6.9%, and ranked it first in 0.8%. A citation-only score would have overstated how often the brand entered the recommendation set.
A four-layer measurement model for fintech
| Layer | What to record | Decision it supports |
|---|---|---|
| 1 Answer exposure | Relevant prompt coverage, mention, recommendation, position, sentiment, cited URLs and answer evidence. | Determines whether the brand is present, accurately represented and actively considered. |
| 2 Assisted demand | Brand-query trends, direct visits, returning users, sales-site engagement and regional changes. | Finds demand patterns that may follow AI exposure without claiming attribution. |
| 3 First-party attribution | How-did-you-hear-about-us responses, demo notes, signup fields, CRM source detail and the exact question used. | Connects a person or account to AI-assisted discovery with declared evidence. |
| 4 Commercial outcome | Qualified lead, activated account, pipeline, revenue, retention and disqualification reason. | Tests whether AI-influenced discovery reaches a business result. |
Use two labels: AI-referred for sessions with a detectable assistant referrer, and AI-influenced for journeys supported by self-report, sales notes or another approved first-party signal. A conversion can have both labels but must be counted once. Preserve the window, market, engine, prompt group and attribution confidence.
What changes in APAC fintech
Fintech buying cycles extend beyond the consumer journeys in the Scrunch study. A recommendation may affect a shortlist before compliance, procurement or technical review. The later signal may be a documentation visit, partner introduction or enterprise demo, so attribution should be account-aware and connected to CRM stages.
APAC also requires separate measurement by language and engine. A ChatGPT click pattern cannot be assumed for DeepSeek, Doubao or Qwen, and Chinese recommendations may lead to local search, social channels or offline referrals.
Where Geolix.ai fits
Geolix.ai can establish the answer-side baseline across relevant engines, languages, markets and purchase-intent questions, then connect those observations to the client’s analytics and CRM evidence. The measurement design should be agreed before optimization begins: which answer states matter, which commercial events are configured, how self-reported discovery is collected, and who approves the final attribution view.
The company’s own data also shows the current limit. GA4 had no configured key events for consultation, registration or form submission during the analyzed window. Visibility and traffic could be compared, but conversion and revenue could not. Improving the measurement system is part of the GEO operating work, not a calculation that can be reconstructed reliably after the fact.
Method and limitations
Scrunch figures are vendor-published research, not an independent fintech benchmark. Geolix.ai figures come from its own monitoring, GA4 and Search Console exports. The tools cover different dates and units; GA4 starts on July 25, Search Console includes monitoring-generated queries, and there was no control group. The data identifies measurement gaps, not incremental revenue caused by AI recommendations.
Frequently asked questions
Can AI search ROI be attributed precisely?
Not with one channel report. Direct referrals can be observed precisely, while assisted influence requires a combination of self-report, CRM evidence, brand-demand trends and controlled experiments where practical. The final report should distinguish observed referral, supported influence and modeled incrementality.
Which leading indicators are useful?
Use relevant-prompt recommendation rate, position, accurate product representation, citation quality, brand-query movement and first-party discovery responses. Report them by market, engine and buyer-intent group rather than as one global visibility score.
How can teams avoid double counting?
Assign a stable lead or account identifier, preserve the first and latest known touchpoints, and define a lookback window. Count each commercial outcome once, then attach multiple supporting touches. Do not add direct, organic and self-reported AI conversions as if they were separate customers.