Anthropic's AI Exposure Index: What Real-World Usage Data Means for Your Website
On March 5, 2026, Anthropic published "Labor Market Impacts of AI: A New Measure and Early Evidence", a research paper by Maxim Massenkoff and Peter McCrory. The paper introduces "observed exposure," a metric that compares how AI is used in the workplace with how it could theoretically be used. The gap between the two is large. The most exposed occupations are knowledge work that happens largely on the web.
If you build or manage a website, the paper is useful as a frame for one question: what can AI do in theory, and what does it do in practice? The paper measures the labor market, not websites.
A new way to measure AI's impact
Earlier attempts to measure AI's impact on jobs relied on theoretical assessments, in which experts rated whether an LLM could speed up a given task. The Anthropic team combined three data sources instead:
- O*NET Database, task definitions across ~800 US occupations
- Anthropic Economic Index, real-world Claude usage patterns from millions of conversations
- Eloundou et al. (2023), theoretical LLM task feasibility ratings (β scale: 1 = fully feasible, 0.5 = needs tools, 0 = not feasible)
The result is "observed exposure," a metric that weights tasks by theoretical feasibility, actual usage frequency, work-context relevance, automation level, and task-share within occupations. Because it draws on Claude usage data, the metric shows where AI is used in practice and not only where it could be used.
The feasibility gap: theory vs reality
In the data, Claude usage concentrates on tasks that are rated as theoretically feasible, but it covers only part of them:
97% of what people use Claude for falls into categories rated as theoretically feasible. Coverage within those categories is far from complete. This is the feasibility gap: usage sits where AI is rated to work, and it reaches only a fraction of the tasks AI could theoretically handle.
Anthropic measures task exposure and observed AI use in labor data. It does not establish a predictable adoption path for websites, or show that any published web signal causes agent use. Website readiness needs its own evidence.
Which occupations are most exposed?
The most exposed occupations are mostly web-based knowledge work, where people use websites, SaaS tools and digital platforms every day:
Computer programmers lead with 74.5% task coverage. That means AI use covers about three quarters of their work tasks. As a front-end developer, I feel slightly attacked by this. Customer service representatives follow at 70.1% and data entry keyers at 67.1%. Roughly 30% of all occupations register zero AI coverage, among them cooks, mechanics, bartenders and lifeguards, whose work is physical and cannot be copy-pasted.
The adoption gap: theoretical vs observed exposure
A central finding is how far actual AI usage lags behind theoretical capability. The chart compares what AI could do in each occupation group with what it actually does:
Computer & Math occupations have 94% theoretical exposure but only 33% observed. That is a gap of 61 percentage points. It exists because of legal restrictions, software integration requirements, verification protocols and model limitations. Those constraints apply to the labor-market analysis. A website scan can separately identify missing structured data, crawler policies, and machine-readable interfaces, but it cannot infer an adoption trajectory from this gap.
Who are the exposed workers?
AI-exposed workers are, on average, well-educated and well-paid. The table compares them with unexposed workers:
| Demographic | Exposed Workers | Unexposed Workers | Difference |
|---|---|---|---|
| Female representation | Higher | Lower | +16 pp |
| Average earnings | Higher | Lower | +47% |
| Graduate degree holders | 17.4% | 4.5% | 3.5x |
| White representation | Higher | Lower | +11 pp |
| Asian representation | Higher | Lower | ~2x |
Exposed workers earn 47% more on average and are 3.5 times more likely to hold a graduate degree. Exposure is highest in well-paid, highly educated occupations, not in low-skill work. Much of their work runs on digital tools and web-based platforms.
Employment impact: the early signals
The headline finding on employment is that highly exposed workers show no systematic increase in unemployment since ChatGPT launched in late 2022. Younger workers show a different pattern.
BLS growth projections correlate with exposure
The Bureau of Labor Statistics' own employment projections through 2034 correlate with observed AI exposure:
For every 10-percentage-point increase in observed AI coverage, BLS growth projections drop by 0.6 percentage points. This correlation appears only with the observed exposure measure and not with theoretical exposure alone.
The young worker signal
The clearest sign of an effect concerns young workers entering exposed occupations. Workers aged 22-25 in AI-exposed occupations show reduced hiring:
Job-finding rates for young workers in exposed occupations have declined roughly 0.5 percentage points since late 2024, a 14% reduction in hiring rates. Workers over 25 show no equivalent decline. This matches findings from Brynjolfsson et al. The pattern suggests that companies hire fewer juniors while existing workers keep their jobs, presumably because AI tools handle tasks that would have gone to entry-level staff.
The web readiness parallel
The report's gap is specific to occupational tasks. The comparison below is a website-audit analogy, not a result from Anthropic's dataset:
THE LABOR MARKET GAP
94% of Computer & Math tasks are theoretically automatable, but only 33% show actual AI usage. Diffusion constraints (legal, software, verification) slow adoption.
THE WEB READINESS GAP
This research does not measure websites. Missing structured data, blocked crawlers, and absent machine-readable interfaces are observable implementation gaps, not evidence of adoption or task success.
The website column below is an analogy, not a finding from Anthropic's research:
| Labor Market Pattern | Website audit boundary |
|---|---|
| Theoretical exposure (94%) vs observed (33%) | Presence and validity of published machine-readable signals |
| Legal/software barriers slow adoption | Observable access policies and structured-data coverage |
| Most exposed = high-skill knowledge workers | A static scan does not identify which website types will be affected |
| Young worker hiring slows in exposed fields | A static scan does not measure task replacement |
| BLS projects lower growth for exposed occupations | A static scan does not predict visibility |
| Adoption follows feasibility (β=1 tasks first) | A static scan does not predict adoption order |
What the research does not establish for websites
Anthropic's gap between theoretical exposure and observed use applies to occupational tasks. The reported percentages cannot be projected onto websites, and the research does not say that the gap must close.
The Yale Budget Lab's parallel research reports that economy-wide disruption has not materialized while occupational composition shifts. As Fortune reported , that finding concerns labor markets. It is context for product planning, not proof that a particular website change will improve adoption.
What this means for your website
For website owners, the report can motivate audit questions without predicting outcomes:
OPEN THE DOOR TO AI CRAWLERS
Review crawler access policies and publish machine-readable content where it serves a documented audience. These checks measure configuration, not future traffic.
MAKE YOUR DATA MACHINE-READABLE
JSON-LD and Schema.org expose structured facts in initial HTML. Their presence can be validated without claiming that they cause agent use.
EXPOSE AGENT PROTOCOLS
MCP discovery and OpenAPI are measurable interface signals. A2A is emerging bonus evidence. WebMCP remains experimental and non-scored.
BUILD TRUST SIGNALS
HTTPS and security headers are measurable web controls. The scanner reports their configuration without predicting trust or task success.
The scanner groups related published signals into five categories, while keeping score roles explicit:
- AI Content Discovery (30%): crawler access, robots.txt, llms.txt, sitemaps. See how ChatGPT selects sources .
- AI Search Signals (20%): JSON-LD, Schema.org, entity linking, FAQPage schema. See the State of AEO .
- Content & Semantics (20%): SSR, heading hierarchy, semantic HTML, ARIA. See how agents see your website .
- Agent Protocols (15%) includes MCP discovery, OpenAPI, and agents.json. An A2A Agent Card is emerging bonus evidence and does not affect the headline score. WebMCP is experimental, non-scored evidence. See what is WebMCP .
- Security & Trust (15%): HTTPS, HSTS, CSP, security headers.
An early warning system for jobs, not a website forecast
Anthropic frames this research as an "early warning system" for labor-market impacts. The figures below remain useful in that scope, but they do not forecast website traffic, adoption, or task success:
- AI use covers 74.5% of programming tasks in the observed exposure measure
- Hiring of young workers is lower in exposed occupations
- BLS projections show lower growth for occupations with higher observed exposure
The research supports monitoring how AI is used in work. Website teams can verify their published technical signals and collect runtime evidence separately, and should not read labor-market data as proof of future web outcomes.
Sources
- Labor Market Impacts of AI: A New Measure and Early Evidence (Anthropic Research) . Massenkoff & McCrory (March 2026). Primary source for all statistics cited.
- Appendix to "Labor Market Impacts of AI" (Anthropic Research) . Methodology details and supplementary data.
- Anthropic Economic Index: January 2026 Report (Anthropic Research) . The real-world usage data behind the observed exposure metric.
- Evaluating the Impact of AI on the Labor Market (Yale Budget Lab) . Independent CPS analysis showing stability at economy-wide level.
- Incorporating AI Impacts in BLS Employment Projections (Bureau of Labor Statistics) . How BLS accounts for AI in occupation-level growth projections.
- IsAgentReady: The State of AEO: Key Insights from Vercel's 2026 Report
- IsAgentReady: How AI Agents See Your Website: The Accessibility Tree Explained