The practical brief
The finding
Consumer purchases accounted for 5.2% of all conversations in the global Gemini App and Google AI Mode sample; professional and personal care services accounted for 4.8%. [c1] [c5] [c8] [c15]
Why it matters
AI-assisted customer research deserves investigation, but conversation volume is not a forecast of enquiries or revenue.
| Topic | Global share of all conversations | US non-work conversation share / active-time share |
|---|---|---|
| Consumer purchases | 5.2% of all global Gemini App and Google AI Mode conversations | Nearly 3×: US non-work conversation share / US human non-work active-time share; work and sleep excluded [c7] [c33] [c31] [c32] [c37] |
| Professional and personal care services | 4.8% of all global Gemini App and Google AI Mode conversations | More than 7×: US non-work conversation share / US human non-work active-time share; work and sleep excluded [c7] [c33] [c31] [c32] [c37] |
What to try
BLURSOR’s practical interpretation
Our interpretation: audit answers to real customer questions in relevant languages before committing to a larger visibility programme.
Study boundary
This two-week observational study does not measure brand visibility, answer accuracy, conversions or the effectiveness of marketing changes.
Investigate customer research, not promised AI leads
If you sell products or professional services, the decision is whether AI-assisted research belongs in your customer-discovery monitoring. This study supports investigating that channel—not treating AI conversation volume as a forecast of leads or sales.
People already bring purchase and service questions to Google's conversational AI products. The missing link is what happens next: the research does not establish whether users discover a business, receive an accurate recommendation, contact a provider or buy.
What the researchers measured—and filtered out
Google's AI & Economy ATLAS report analyses 14,653,926 de-identified interactions from Gemini App, Google AI Mode and Gemini API between April 6 and April 19, 2026. An automated pipeline groups conversations and classifies them into work tasks and household activities.
The household analysis uses about 10 million logs from Gemini App and AI Mode, excluding API logs. Its categories describe conversation topics, not individual customers or completed transactions. Sampling excluded queries below 10 tokens—short text units—as well as internal Google accounts and prompts without model responses.
Privacy filtering removes clusters representing fewer than ten unique users, potentially hiding niche activities. Global conversations are mapped to US statistical categories, creating additional misclassification risk for activities without US equivalents.
Purchases and services attract disproportionate attention
Consumer purchases accounted for 5.2% of all conversations in the global conversational sample; professional and personal care services accounted for 4.8%. A separate US comparison found purchase topics represented nearly three times, and professional and personal care services more than seven times, their respective human time shares.
Those US ratios compare shares of non-work AI conversations with shares of human non-work active time. They exclude work and sleep and normalize across the American Time Use Survey's top-level activity categories. They do not compare AI attention with the full daily time budget, and their denominator differs from the global conversation shares.
This is an attention comparison, not evidence of answer quality, purchasing influence or time savings. Among productive household tasks, purchase research and household management led conversation shares; assistance concentrated on planning and organisation rather than physical execution.
Just below half of medical, legal, financial, government and civic-service conversations occurred outside weekday 9 AM–5 PM or on weekends. That indicates when questions arise, not whether an after-hours offering would attract paying customers.
A realistic response is a focused information audit
Our interpretation is to start with customer questions rather than a broad promise of “AI optimisation.” A modest audit can identify issues worth investigating, although this paper does not test whether correcting them changes AI answers.
For multilingual markets, avoid assuming English represents the whole audience. English accounted for just over one-third of conversations in the App and AI Mode language analysis. This supports checking relevant languages, not assuming equal answer quality across them.
- Select questions from actual customer enquiries, including purchase comparisons and service-process questions.
- Record inaccurate descriptions, omissions and unsupported claims separately from sales results.
- Check whether public information answers those questions clearly. Treat content changes as a test, not a proven ranking tactic.
What this snapshot cannot tell your business
The data covers current users of selected Google products, not non-adopters or the entire AI market. Paid API content and several Google experiences are excluded, potentially under-representing enterprise use. Automated classifications are uncertain, especially for detailed tasks.
US statistical categories may poorly fit activities elsewhere. The study cannot establish whether activities are disproportionately represented relative to another country's own time-use distribution.
The report is observational, so it cannot establish causal purchasing changes or productivity gains. Household-benefit calculations assume time savings rather than measure them. Most importantly, it offers no tested recipe for earning citations, recommendations or customers.
The source and evidence
Google's AI & Economy ATLAS v1.0: Mapping Gemini Usage in the Economy
Study limitations and disclosures
- All listed authors are affiliated with Google or Google DeepMind. The report studies Google's own products using internally governed data and Gemini-based processing, creating a material commercial interest in the subject.
- Sampling excluded queries below 10 tokens, internal Google accounts and prompts without model responses. Reported conversation shares therefore do not represent all queries, particularly very short ones.
- The main classified dataset excludes paid Gemini API content and several Google experiences, including Workspace and AI Overviews. Enterprise use may be under-represented, and non-adopters are not represented.
- Global conversations are mapped to US statistical taxonomies, increasing misclassification risk for activities without US equivalents. US time-use comparisons cannot establish over-representation relative to other countries' local time-use distributions.
- The two-week observational sample measures interactions, not successful task completion, answer accuracy, brand visibility, sales or causal productivity gains. Household benefit valuations rely on assumed rather than measured time savings.
- Automated classifications are uncertain, especially for detailed tasks. Synthetic validation can introduce bias; human approval of a plausible label is not equivalent to exact ground-truth accuracy.
- Privacy thresholds exclude clusters with fewer than ten unique users, potentially hiding niche activities and languages. The principal language analysis excludes Gemini API.
Evidence behind this briefing
[c1] ATLAS analyzes 14,653,926 de-identified interactions across the Gemini App, Google AI Mode, and Gemini API during April 6–19, 2026, using a fully automated privacy-preserving pipeline.
Section 2, Data and Methods Overview · Read the study
[c2] The executive summary reports disproportionate non-work usage for administrative and professional-service topics: government services and civic obligations were represented at almost twenty times their share of actual time spent. Nearly half of medical, legal, financial, and government AI consultations occurred outside 9-to-5 business hours. These are usage patterns, not evidence of successful consultations or business discovery.
Section 1, Executive Summary, finding 8 · Read the study
[c3] In the US comparison of Gemini interaction shares with employment shares, financial/market analysts, software developers, and systems administrators were among the more over-represented detailed occupations. Observed high-use tasks included data collection and analysis and information monitoring; this is relative usage representation, not a worker adoption rate.
Section 3.3, Occupational AI Usage, discussion of Table 1 · Read the study
[c4] The preliminary intent classification assigned automation intent to less than 10% of non-routine cognitive work conversations, versus more than a quarter of routine cognitive work conversations. Non-routine use centered on drafting, refinement, ideation, and information retrieval. The authors explicitly warn that these findings are not definitive.
Section 3.6, Task Intent · Read the study
[c5] The dataset excludes paid Gemini API usage, including enterprise usage through Google Cloud, and several other Google products. It measures interactions rather than productivity or economic outcomes, excludes non-adopters, and uses probabilistic classifications whose granular results are more uncertain than broad trends.
Section 2.3, Limitations · Read the study
[c6] Privacy filtering discards clusters with fewer than ten unique users. Processing is automated, original conversations and individual summaries are absent from the final dataset, and final-data access is restricted under internal privacy oversight. Gemini models also perform summarization, classification, taxonomy enrichment, and synthetic validation-data creation.
Section 2.2, Privacy and Data Governance · Read the study
[c7] The household analysis uses about 10 million logs from Gemini App and Google AI Mode, excludes API logs, and maps summarized conversation clusters to the ATUS activity taxonomy. Its reported coverage is 74% of classified ATUS categories, representing about 98% of Americans’ average non-sleep daily time—not 98% of people or AI products.
Section 4.1, The Broad Picture · Read the study
[c8] In the global conversational sample, 86.5% of conversations are classified as non-work. Consumer purchases account for 5.2% of all conversations, and professional and personal care services account for 4.8%. These are conversation shares, not customer conversion, answer accuracy, brand visibility, or purchasing outcomes.
Figure 10, Notes · Read the study
[c9] For US non-work usage compared with ATUS human time allocation, consumer purchases over-index by nearly threefold and professional and personal care services by more than sevenfold. This measures relative conversation share versus time share, not demonstrated usefulness or causal purchasing influence.
Section 4.2, Comparisons with Human Time Usage and Heterogeneity; Figure 11 · Read the study
[c10] Among productive household tasks, researching purchases and household management lead Gemini conversation shares. Cluster inspection suggests assistance concentrates on planning and organization rather than physical execution; it does not establish measured productivity gains.
Section 4.3, Non-Market Welfare: Valuing Household Time Savings; Figure 14, Panel A · Read the study
[c11] Household economic benefits are scenario calculations, not measured causal time savings. The authors assume average savings of 0.5–5%, acknowledge that cited tasks may not represent ATLAS activities, and warn that time savings omit improvements in quality and breadth.
Section 4.3, household time-savings scenario assumptions · Read the study
[c12] The API geography analysis includes enterprise, educational, and paid requests, uses originating server IP rather than end-user location, and is not adjusted for Gemini API penetration. Its country quintiles therefore should not be treated as comparable measurements of end-user AI adoption.
Figure 23, Notes · Read the study
[c13] API work-use estimates exclude European countries and assume all paid usage is work-related because logging does not permit the usual classification. Among 106 countries/regions with data, 16% have less than 90% work conversations and fewer than 2% have less than 80%; these are conditional descriptive estimates.
§5.2, Gemini API work/non-work analysis preceding Figure 24 · Read the study
[c14] Country rankings differ when work use is measured as a share of total conversations rather than per-capita volume: high-income regions fall into lower share quintiles, while developing regions, especially Africa, rise. The authors offer competing explanations, including constrained casual adoption; this is not evidence of a causal productivity benefit.
§5.2, paragraph introducing Figure 25 · Read the study
[c15] The language analysis of Gemini App and Google AI Mode, excluding Gemini API, identifies 143 privacy-cleared languages. English accounts for just over one-third of conversations, Spanish 12%, Arabic just under 7%, and Portuguese just under 6%. Minimum cluster thresholds make observed diversity an underestimate; coverage of over 93% refers to first-language speakers among the 200 most spoken languages, not all global speakers.
§5.3, language distribution paragraph preceding Figure 27 · Read the study
[c16] Non-primary-language conversation shares are similar across work and non-work: 26% versus just below 24%. For countries where English is non-primary, English shares are 14% versus 13%. Figure 28 includes countries with over 1,000 conversations and defines primary language by the most spoken language for AI queries within a country; the comparison challenges systematic professional code-switching but does not measure response quality.
§5.3, paragraph preceding Figure 28; Figure 28 notes for inclusion threshold and primary-language definition · Read the study
[c17] With granular ATUS activity and language fixed effects, non-primary English conversations are associated with 9–12% higher turn counts and 18–20% more total tokens, plus slightly greater technical complexity. These ranges are reported estimates, not identified confidence intervals. The authors explicitly allow within-task sorting of more complex queries into English as an alternative to a language-related interaction cost.
§5.3, final paragraph · Read the study
[c18] ATLAS measures de-identified interactions rather than realized productivity, lacks enterprise-contracted logs for corporate integration analysis, and is observational. It cannot establish macroeconomic productivity gains or causal effects on labor substitution, wages, or skills.
§6, open questions paragraph · Read the study
[c19] ATLAS samples approximately five million interactions per surface from Gemini App, Google AI Mode and Gemini API over April 6–19, 2026, then reweights by source interaction shares. It excludes queries below 10 tokens, internal Google accounts and prompts without model responses.
Appendix A.1.2, Sampling · Read the study
[c20] Paid Gemini API request counts inform geographic dispersion, but paid-request content is unavailable and therefore excluded from more granular analyses.
Footnote 2 · Read the study
[c21] On synthetic-data validation, exact O*NET task classification accuracy is 22.6%, whereas agreement after aggregation into Autor–Thompson task types is 70.4% of clusters. These are classifier validation results, not the accuracy of Gemini answers to users.
Appendix B.2.1, Result Overview; Table 5 · Read the study
[c22] The Gemini-generated synthetic benchmark can bias measured accuracy in either direction: extra context or model-specific linguistic cues can make classification easier, while ambiguous prompts can make a defensible classification disagree with the seeded label. The first bias cannot be fully measured under the study's privacy constraints.
Appendix B.1.2, Classifier Validation Under Information Constraints · Read the study
[c23] Human approval measures whether a proposed label is economically reasonable, not exact ground-truth accuracy. Three raters reviewed 120 work and 120 non-work clusters across the three surfaces; reported approval is an unweighted average across raters. Table 7 reports 85.8% approval for work task statements and 92.2% for detailed non-work labels.
Appendix B.3.2, Human Approval of AI Labels; Table 7 and notes · Read the study
[c24] The supplementary figure notes report that just below half of conversations across medical, legal, financial, government and civic service domains occur outside weekday 9 AM–5 PM or on weekends. This is conversation timing, not evidence of conversions, unmet service demand or causal effects.
Appendix D, Figure 38 and notes · Read the study
[c25] The main ATLAS API pipeline samples free Gemini API requests, not paid request content. Initial samples exceed 5 million per surface, but privacy filtering removes clusters with fewer than 10 unique users, leaving smaller and unequal final surface samples.
Footnote 65; related disclosures in Footnotes 2 and 67 · Read the study
[c26] OECD occupational comparisons likely understate coding applications because European Gemini data availability is more limited and API data is unavailable for classification.
Footnote 8 · Read the study
[c27] Adjusted global usage estimates use StatCounter web referrals as a proxy, not conversation counts. They exclude AI Mode and assume its country-level popularity falls proportionally with below-average Gemini Apps penetration; relative-diffusion interpretation also assumes referrals positively correlate with combined usage.
Footnote 49 · Read the study
[c28] Observed end-to-end task automation must not be interpreted as job automation: coordination costs, complementary tasks, and organizational frictions remain relevant.
Footnote 24 · Read the study
[c29] The granular language breakdown offers weak evidence for task sorting to bypass model-quality penalties. The authors instead interpret non-primary-language use in volunteer, religious, and civic activities as suggesting sociolinguistic context; this is an interpretation, not a causal test or accuracy measurement.
Footnote 61 · Read the study
[c30] Because the synthetic validation dataset is work-skewed, reported classification accuracy averages 100 bootstrap simulations, each using a balanced sample of 2,750 work and 2,750 non-work conversations. This footnote supplies no numerical accuracy result.
Footnote 73 · Read the study
[c31] The US comparison uses shares of non-work AI conversations and human non-work active time across ATUS Tier 1 categories, excluding work and sleep. It is not a comparison against the full daily time budget.
Section 4.2, Figure 11 notes · Read the study
[c32] Within the US non-work comparison defined by Figure 11, Professional & Personal Care Services over-indexes by more than seven-fold and Consumer Purchases by nearly three-fold. These ratios concern conversation-share versus human-time-share representation, not measured time savings or total daily time.
Section 4.2, paragraph beginning “Figure 11 compares” · Read the study
[c33] The global distribution has a different denominator: all AI conversations. Non-work interactions account for 86.5% of all conversations; Consumer Purchases and Professional & Personal Care Services account for 5.2% and 4.8% of total usage, respectively. These global shares should not be conflated with the US non-work representation ratios.
Section 4.1, Figure 10 notes · Read the study
[c34] Mapping global AI interactions to US statistical taxonomies creates additional misclassification risk where occupations, tasks or activities have no US equivalents and sufficient usage to pass the study's privacy thresholds.
Appendix A.2.2, Primary Statistical Taxonomies · Read the study
[c35] The approach cannot establish whether an occupation or activity is over-represented in a non-US country's AI use relative to that country's employment shares or population time use.
Appendix A.2.2, Primary Statistical Taxonomies · Read the study
[c36] The appendix's granular time-use comparison explicitly compares US AI conversation shares with ATUS human time shares and excludes work and sleep; it ranks detailed tasks by absolute percentage-point difference.
Appendix D, Figure 34 notes · Read the study
[c37] The time-share comparison excludes sleeping and work activities and normalizes across ATUS Tier 1 categories; it is not a comparison against the full daily time budget.
Footnote 37 · Read the study
[c38] The time-share comparison controls for baseline human attention; the authors explicitly do not interpret information retrieval as physical task execution or the comparison as evidence of task substitution.
Footnote 36 · Read the study