Google today published the first edition of its AI & Economy ATLAS, a study of nearly 15 million de-identified interactions that maps generative AI usage against official labour statistics. The headline result cuts against the automation narrative: AI has spread across two thirds of occupations, but within any given job it covers roughly a fifth of the work.

The report, titled "Google's AI & Economy ATLAS v1.0: Mapping Gemini Usage in the Economy," lists seventeen authors across Google and Google DeepMind, with Zanna Iscenko and Scott Strand as corresponding authors. ATLAS stands for Activity, Task, Landscape, and Adoption Study.

The dataset covers 14,653,926 interactions sampled from the Gemini App, Google AI Mode, and the Gemini API between April 6 and April 19, 2026, mapped to more than 800 occupations, 4,000 work tasks, 300 household activities, 150 countries, and 140 languages. According to Google, 143 distinct languages clear the study's privacy thresholds, covering languages spoken by over 93% of first-language speakers among the 200 most spoken worldwide.

GoogleATLASv1
GoogleATLASv1.pdf
12 MB
download-circle

Broad reach, thin penetration

The central finding concerns the gap between how many jobs use AI and how much of each job it reaches.

Google observed usage above its minimum threshold in 68% of detailed occupations globally, representing 88.4% of employed civilian workers in the United States. But among occupations where any meaningful usage appears, the median share of constituent tasks performed with Gemini is 21%.

The distribution is skewed further than that median suggests. For 29% of detailed occupations, the study observed zero task saturation, meaning no single task drew at least 25 users. That group includes stockers and order fillers, food preparation workers, fast food cooks, refuse collectors, and police and sheriff's patrol officers. At the other end, only 3% of occupations showed AI usage across three quarters or more of their tasks: software quality assurance analysts and testers, human resources specialists, and document management specialists.

Google draws the comparison with prior industry research explicitly. Anthropic's Economic Index found that roughly 36% of occupations used AI for at least a quarter of their tasks when it launched in February 2025. Google's equivalent figure is around 30%, which the authors attribute partly to stricter privacy-preserving thresholds rather than to a genuine divergence in behaviour. Both point the same direction: diffusion is wide and shallow.

Automation remains the exception

The second finding addresses intent directly. Google built a classifier assigning conversation clusters to five categories: Task Automation, Partial Drafting and Generation, Review and Refinement, Ideation and Strategy, and Information Retrieval and Learning.

Non-routine cognitive analytic tasks, meaning work requiring judgment, strategic thinking, problem-solving, and creativity, account for 35% of all tasks in the O*NET taxonomy but 65% of Gemini work interactions. Within that heavily over-represented category, end-to-end automation represents the intent of fewer than 10% of conversations. Routine cognitive work behaves differently: more than a quarter of conversations in that subset target automation, a contrast the authors describe as potentially meaningful while cautioning that the classification is preliminary. Manual tasks concentrate overwhelmingly on information retrieval and learning.

That structure matters for anyone reading the labour market debate. Google states plainly that the data does not support claims that AI is about to cause massive automation and displacement of white-collar work, that AI is irrelevant to blue-collar work, or that the purpose of AI is strictly to automate tasks, while cautioning that these observations are early and could change as model capabilities advance.

Physical trades show up in the data

Nearly a third of heavily physical occupations register no observed AI usage. Yet the report identifies substantial activity in manual and technical trades where AI functions as a diagnostic collaborator.

Automotive service technicians and mechanics, whose tasks are 83% manual, generated more than ten thousand conversations covering vehicle component testing, system rewiring, and inspection of parts for wear, with multimodal conversation shares running more than twice the work average. Industrial machinery mechanics, at 44% manual, produced thousands of interactions analysing test results and machine error messages. Electrical and electronics repairers, computer user support specialists, and photographers also combine above-median manual task shares with notable usage.

The wage gradient

Google measures usage intensity as conversations classified into an occupation divided by that occupation's employment. On that basis, a 1% increase in an occupation's median earnings is associated with more than a 2.5% increase in usage intensity. Controlling for educational attainment, the coefficient falls to 1.86.

The weighting exercise is stark. Median annual earnings across US civilian workers in the study sample, weighted by employment, come to $62,252. Weighted instead by Gemini conversations, that figure rises to $82,919. Weighted by token usage, it reaches $86,157.

Financial and investment analysts, market research analysts, software developers, and network and computer systems administrators rank among the most over-represented occupations relative to US employment. Home health and personal care aides, fast food and counter workers, cashiers, and retail salespersons rank among the most under-represented.

There is a wrinkle in the expertise data. Using the Autor and Thompson methodology, which scores task expertise by the rarity and domain specificity of the vocabulary in each task statement, Gemini usage is most over-represented for the lowest expertise cognitive tasks, particularly the non-routine variety. Cited examples include rewriting news reports into specified languages and writing product specifications. High-earning workers adopt AI at the highest rates, but the tasks they bring to it skew toward the lower end of the expertise distribution.

Most usage happens at home

Over 86% of conversational AI interactions in the sample occur outside formal work. Restricting to the Gemini App and AI Mode and excluding API traffic, work and work-related conversations account for 13.5% of the total.

Mapped to the American Time Use Survey taxonomy, the largest categories are Socializing, Relaxing, and Leisure at 26.4% and Education at 20.7%. Household Activities take 11%, Personal Care 9%, Consumer Purchases 5.2%, and Professional and Personal Care Services 4.8%. The data covers 74% of classified ATUS categories, representing about 98% of Americans' average daily non-sleep time.

Comparing conversation share against actual time spent produces sharper contrasts. Education over-indexes by a factor of roughly 5.8, Professional and Personal Care Services more than sevenfold, Consumer Purchases nearly threefold. The largest outlier is Government Services and Civic Obligations, over-represented by almost twenty times relative to the hours Americans report spending on it. Eating and Drinking runs about eighteen times under-represented, with Traveling, Sports and Recreation, and Caring for Household Members also showing negative gaps.

Roughly half of medical, legal, financial, and government consultations occur outside standard business hours, removing a scheduling cost that previously required substituting away from paid work.

Putting a number on unpaid gains

Google attempts a valuation, and hedges it heavily.

According to ATUS data for 2024, the average American spends about 17.8 hours per week on productive household tasks under the third-person criterion, meaning activities that could theoretically be delegated to a paid third party. Valuing that time at the Bureau of Economic Analysis replacement wage of $12.02 per hour, a 0.5% average time saving of roughly 5.4 minutes per week yields $56 annually per person and $14.9 billion across the US adult population. A 2% saving produces just under $60 billion. A 5% saving, described as the optimistic upper bound, produces about $149 billion.

The authors chose a conservative range of 0.5% to 5% against published estimates running considerably higher, one of which puts time savings for shopping browsing at 43% to 64%. The report notes the $12.02 wage was not adjusted for inflation, acknowledges that assigning time savings is a flawed metric since improvements may show up in quality rather than reduced hours, and observes that any savings would distribute unevenly given that men spend as much as 30% less time on productive household activities than women. These gains fall entirely outside GDP.

The geography of adoption

Per-capita conversational usage tracks national wealth closely. A 1% increase in GDP per capita is associated with roughly a 0.9% increase in AI conversations, using figures adjusted for platform penetration through StatCounter web referral data. Countries in the lowest usage quintile hold about 17% of world population and generate 2% of conversations. The top quintile holds 11% of population and drives 30%, 2.8 times its proportional share.

Several middle-income countries break the pattern. Türkiye, the United Arab Emirates, Qatar, and Israel sit in the highest tiers, and Chile, Peru, Brazil, Argentina, and Colombia reach the high and very high quintiles. India and Russia fall into the low tier despite large populations and substantial STEM graduate output.

Search interest inverts the picture. Google Trends data from January through April 2026 places India, Bangladesh, Nepal, the Philippines, Indonesia, Thailand, Vietnam, Kenya, and Ethiopia in the highest quintile for relative AI search interest, while Western Europe and Japan sit in the low and very low quintiles. The authors call this an interest-adoption gap: AI has become background utility in mature markets while curiosity elsewhere has not yet converted to usage. Adjusting for internet access nearly triples the adoption metric for Sub-Saharan Africa without closing the gap.

A further inversion appears when work usage is measured as a share of each country's total rather than per capita. Under that lens, the United States and European Union drop into lower quintiles while African countries surge into the top. The report offers competing readings: professionals in developing economies may be using AI more intensively as a productivity multiplier, lower baseline casual adoption may be concentrating usage within work contexts, or the exclusion of Gemini Apps enterprise subscription data may be suppressing professional usage in wealthier markets.

Language patterns resist the obvious hypothesis

English accounts for just above a third of conversations. Spanish follows at 12%, Arabic just below 7%, Portuguese just below 6%, Turkish just above 4%, Korean at 4%, and Vietnamese just below 2%.

The study tested whether users switch to English for economically important work on the assumption that model quality is higher there. The data does not support it. Non-primary language usage averages 26% for work activities and just below 24% for non-work, clustering tightly along the parity line across countries. Where code-switching does concentrate is in Volunteer Activities at 21.9%, Religious and Spiritual Activities at 20.5%, and Government Services and Civic Obligations at 18.5%, which the authors read as sociolinguistic rather than performance-driven. Non-primary English conversations do carry measurable overhead: 9% to 12% higher turn counts and 18% to 20% more total tokens after controlling for activity type and language.

Method and stated limits

The pipeline runs through Observation Clustering and Taxonomy Organisation, or OCTO, a clustering and hierarchical taxonomy tool built by Google DeepMind. Conversations pass through Data Loss Prevention filters stripping personally identifiable information, then receive replacement identifiers unlinked from user logs. Text is summarised individually with full text discarded, then aggregated and re-summarised at cluster level. Any cluster representing fewer than 10 unique users is discarded. Gemini 3.1 Flash Lite performed the intent and task classification, validated against synthetic ground-truth data, inter-rater agreement testing, and human approval of labels.

The limitations section is unusually direct for a corporate research publication. ATLAS v1.0 excludes paid Gemini API usage, which covers enterprise traffic through Google Cloud, meaning enterprise use cases may be under-represented. It also excludes Google Workspace, AI Overviews, Google Translate, Google Maps, and Gemini Notebook, and API content is unavailable for the European Economic Area, the United Kingdom, and Switzerland. The study measures behavioural interactions rather than productivity outcomes, captures only current usage rather than the full set of tasks AI could address, and its classifications are probabilistic.

Google also lists claims ATLAS cannot address, including whether AI is deepening the global digital divide and whether it is dampening entry-level hiring. The report carries a guest commentary on household production and the measurement boundary from Diane Coyle, Bennett Professor of Public Policy at the University of Cambridge, who is credited alongside David Autor of MIT for contributions and review.

Why this matters for marketers

The occupational data lands directly on marketing functions. Market Research Analysts and Marketing Specialists appear as one of the five most task-saturated occupations in the entire dataset, and also rank among the most over-represented detailed occupations relative to US employment. Whatever the aggregate picture of shallow diffusion, marketing research sits at the deep end of it.

That aligns with earlier findings. Research analysing four million Claude conversations found that 5% of occupational tasks account for 59% of all AI interactions, clustering in tasks scoring high on creativity and cognitive complexity rather than routine execution, and Microsoft research covering 2024 found AI applicability concentrated across marketing communications occupations. Google's non-routine cognitive figure of 65% against a 35% baseline is the same pattern on a different dataset.

The self-report gap is where the picture gets uncomfortable. A WARC study conducted with TikTok in May 2026 found 88% of marketers reporting higher creative volume since adopting generative AI but only 45% reporting significant quality improvement, with 87% rating their organisation's use effective. MiQ's survey of 3,169 marketers across 16 countries found 72% planning expanded AI use while only 45% felt confident applying it. Google's behavioural measurement of 21% median task coverage sits closer to the low self-reported figures than the high ones.

For the search and publishing side, the 86% non-work share carries a different weight. Google measured these conversations at a moment when AI Mode passed one billion monthly active users, confirmed at Google I/O on May 19, 2026, and the Gemini App reached 900 million monthly users. The categories over-indexing most heavily against time use, meaning education, comparison shopping, financial services, and government navigation, are precisely the informational queries that historically generated publisher referral traffic and search inventory.

A randomised field experiment with 1,065 Chrome users found AI Overviews cut outbound organic clicks by 39.8% and raised zero-click searches by 34.5%, and Ahrefs measured a 58% click-through rate reduction for top-ranking pages by February 2026. ATLAS measures no referral traffic and excludes AI Overviews entirely. What it supplies instead is a taxonomy of what those absorbed queries consist of, mapped to standardised categories rather than inferred from keyword sets.

The comparison shopping finding carries weight for retail and performance teams. Consumer Purchases over-index nearly threefold against time spent, and researching purchases ranks among the two largest productive household categories by conversation volume. That activity is occurring inside a surface where Google has been building connected app integrations and commerce pathways.

One structural caution applies throughout. This is Google measuring Google's own products. The study is transparent about that, adjusts cross-country figures using third-party StatCounter data specifically to avoid presenting Google's regional footprint as global diffusion, and enumerates its exclusions at length. But the finding that AI is complementary rather than substitutive is also the finding most favourable to the company publishing it. OpenAI's September 2025 analysis of 700 million ChatGPT users reported roughly 70% non-work usage, a directionally similar result from another interested party. Convergence across vendor datasets strengthens the observation without removing the incentive shaping what gets measured.

Timeline

Summary

Who: Google and Google DeepMind, through seventeen listed authors including corresponding authors Zanna Iscenko and Scott Strand, with contributions and review credited to Diane Coyle of the University of Cambridge and David Autor of MIT.

What: ATLAS v1.0, an economic research report built on 14,653,926 de-identified interactions mapped to more than 800 occupations, 4,000 tasks, 300 household activities, 150 countries, and 140 languages. Principal findings are that AI usage appears in 68% of detailed occupations covering 88.4% of US employment while reaching a median of only 21% of tasks within them, that end-to-end automation is the intent of fewer than 10% of non-routine cognitive conversations, and that over 86% of conversational usage occurs outside formal work.

When: Published July 23, 2026, drawing on interactions sampled between April 6 and April 19, 2026.

Where: Global, covering more than 150 countries and territories, with occupational analysis benchmarked against US Bureau of Labor Statistics and O*NET taxonomies and household analysis against the American Time Use Survey.

Why: The report supplies behavioural evidence for a labour market debate that has run largely on projections and self-reported surveys. For marketing specifically, Market Research Analysts and Marketing Specialists rank among the five most task-saturated occupations in the dataset, while the informational and comparison shopping categories that over-index most sharply against human time use are the same query types driving documented publisher referral decline.