Three new cards from two operators. Kevin Indig's Search Signals Index surfaces a structural gap between aggregators' Google dominance and their ChatGPT citation share. Armin Ronacher contributes two pieces that deflate common assumptions in AI development: one on how LLMs change language selection, one on what model reasoning actually is under the hood.
Aggregators are winning Google and losing the ChatGPT citation pool
Kevin Indig's Growth Intelligence Brief #23 (August 20) includes a measurement that has direct implications for any business whose visibility strategy was built around aggregation. His Search Signals Index, tracking approximately 2,600 companies across 26 verticals, found that aggregator-style domains hold 76.0% of Google's SEO visibility between the two company types, but only 38.3% of ChatGPT's citation pool.
"Aggregators hold 76.0% of the SEO visibility split between the 2 company types, and only 38.3% of ChatGPT's citation pool."
The 37.7-point gap is not a product of one vertical or one query set. It reflects the structural difference between how Google and LLMs weight authority. Google's algorithm rewards breadth: more pages, more inbound links, higher domain authority from content volume. LLMs weight directness: a primary source that answers the query specifically, without an intermediary layer. A domain that built Google authority by aggregating content from other sources now competes against those same sources in AI search. The implication is not that aggregators need to optimize their aggregation. It is that aggregation as an authority-building strategy requires a parallel track for AI search that looks fundamentally different. Aggregators hold 76% of Google's SEO visibility but only 38% of ChatGPT's citation pool, requiring separate authority-building strategies for each surface.
- Kevin Indig, Aggregators hold 76% of Google's SEO visibility but only 38% of ChatGPT's citation pool, requiring separate authority-building strategies for each surface. Search Signals Index data shows aggregators lose 37.7 points of citation share between Google (76%) and ChatGPT (38.3%), signaling that Google authority does not transfer to AI search.
Two assumptions Ronacher dissolves in AI development
Armin Ronacher published two essays in the past week that address the same underlying question from different angles: what actually changes about software development when LLMs become capable? The answers he reaches in each case are more specific and more useful than the standard "AI will change everything" framing.
The first, "Fast and Hard Code" (August 22), makes the claim that the language-familiarity tax on development choices is gone.
"LLMs make language choice much less consequential than it used to be"
Months of ramp-up time in a new language was the historical cost of choosing Zig for performance or Rust for safety. That cost disappears when an agent can read and write unfamiliar syntax fluently and a developer can review the output without having spent years writing it themselves. Ronacher names Cloudflare's Zig-based Git engine and Vercel's fx agent as current examples. The constraint that survives is ecosystem fit and tooling maturity, not familiarity. Language selection becomes a deliberate engineering preference rather than a team skill inventory. LLMs have dissolved the language-familiarity tax, turning programming language selection from a skill constraint into a deliberate preference.
The second, "What Is Reasoning" (August 19), addresses the question of what frontier model reasoning actually is at the implementation level. The answer Ronacher arrives at is deflating: it is text output routed through special tokens by trained convention.
"they really are just text: the model is trained to emit its thinking into a scratchpad as part of its response"
"in some sense the only 'special' behavior for some models is not to think"
When reasoning is enabled, the model emits tokens like
- Armin Ronacher, LLMs have dissolved the language-familiarity tax, turning programming language selection from a skill constraint into a deliberate preference. LLMs remove the familiarity tax from language selection, making ecosystem fit and performance profile the deciding factors rather than team expertise.
- Armin Ronacher, Reasoning traces in frontier models are text routed through special tokens by convention, not a separate cognitive process. Reasoning traces are text routed through trained token conventions, not a separate cognitive process, with concrete implications for token budget and prompt design.