Three new cards, two new operators. Independent signals pointing toward the same structural shift: AI is rewriting where attribution goes, where cost goes, and which roles can do knowledge work at scale.
Theme 1, AI changes where traffic comes from before it changes search volume
Rand Fishkin and Similarweb published quantified evidence this week that AI brand recommendations produce a measurable direct-traffic lift. When AmEx or Capital One gets recommended by an AI tool, visitors arrive at the brand's domain 7 to 14 percent more often in the following seven days. Branded search volume falls slightly in the same window. The mechanism is simple: the AI satisfies the discovery need inside the conversation, so the user navigates directly rather than re-querying a search engine. AI brand recommendations produce a 7 to 14 percent direct visit lift within seven days documents the specific lift figures (7.2% for AmEx, 14.2% for Capital One) and the attribution implication: teams tracking only search share of voice will miss this entirely.
This connects to Fishkin's earlier analysis of zero-click behavior (Zero-click Google search has crossed 68% and will not reverse because Google's financial incentives now align with keeping users on-platform). The channel shift was already underway when AI recommendations were weak; now that AI tools actively name brands, the direct-traffic channel gains a new driver that is orthogonal to SEO.
- Rand Fishkin, AI brand recommendations produce a 7 to 14 percent direct visit lift within seven days. Quantified direct-visit lift from AI brand recommendations for large consumer brands, seven-day window.
Theme 2, AI-native companies run on an inverted cost structure
Tomasz Tunguz published a detailed breakdown of AI-native unit economics on June 29. The headline: Anthropic spends 2.3 times its payroll on compute, compared to 0.4 times at top traditional software firms. AI-native companies invert the traditional software cost structure, spending more on compute than on payroll maps the mechanism. When intelligence is the product, output scales with tokens consumed, not labor hours. Token price deflation running at roughly 10 times per year for three years has not reduced compute bills; it has expanded volume instead. Goldman Sachs projects 24 times consumption growth by 2030.
This card sits alongside Tunguz's earlier finding on model substitution (Model substitution savings are reinvested in more tokens, not returned to the budget): savings from switching to cheaper models get reinvested in more tokens rather than returned to the budget. Both cards together describe the same dynamic from different angles. Model substitution keeps the bill flat. The cost inversion keeps the ratio elevated. AI-native companies are not building toward traditional SaaS margins; they are building a different cost structure entirely.
- Tomasz Tunguz, AI-native companies invert the traditional software cost structure, spending more on compute than on payroll. Compute at 2.3x payroll at Anthropic vs. 0.4x at traditional software firms; mechanism of volume expansion under price deflation.
Theme 3, Agentic AI spreads beyond engineering when skills lower the adoption floor
Drew Johnston and David Holtz published a large-scale study of Codex adoption across OpenAI's own workforce on June 25. The finding that deserves attention: legal staff output grew 13 times, researcher output grew 50 times, and the factor that separated rapid adoption from plateau was skills, shareable workflow templates that let non-technical users run complex agentic workflows without writing prompts. Agentic AI spreads beyond engineering into legal and research roles when skills-sharing lowers the adoption floor documents the mechanism.
The skills layer is the key variable. Without it, agentic tools stay concentrated in the roles that built them. With it, diffusion follows the skills catalog rather than individual technical capacity. 26.6% of active Codex users used skills, and that cohort drove the majority of cross-functional growth. The practical implication: organizations deploying agentic tools should treat skills-authoring as a first-class priority alongside deployment, not an afterthought.
- Drew Johnston and David Holtz, Agentic AI spreads beyond engineering into legal and research roles when skills-sharing lowers the adoption floor. Codex deployment study at OpenAI: 5x user growth, 13-50x output growth in legal and research roles, driven by skills-sharing.