Four new cards, one new operator profile. Two themes that pulled in different directions from the scan. The first is about AI search behavior splitting citation from recommendation, making the relationship between "getting found" and "getting recommended" structurally different from what it was in organic search. The second is about judgment as the part of agentic workflows that should not be pre-encoded into rules.
Theme 1, AI search now rewards different behaviors for citation and recommendation
Lily Ray published a 90-day study on July 7 using 100 B2B "best software" queries tracked across three checkpoints. The finding in A brand's self-promotional "best of" listicle gets cited by AI Overviews 74% of the time but excluded from the recommendation 69% of the time is precise: a brand's own self-promotional "best of" listicle appears as a cited source in AI Overviews 74% of the time, while that same brand is absent from the recommendation 69% of the time. The mechanism is not a penalty. It is a decoupling. Google appears to be treating the list of competitors named in a self-promotional article as editorial votes for those competitors, not for the brand that wrote the article.
"Across the 80 prompts that triggered an AI Overview (notably, about 1 in 5 'best software' queries didn't surface an AI Overview at all), when a brand's own self-promotional listicle got cited as a source, that brand was left out of the actual recommendation 69% of the time (224 of the 323 self-promotional listicles cited across those prompts) on AI Overviews."
"Google appears to have decoupled what it cites from who it recommends."
Kevin Indig published the structural complement in Primary research earns 3.3x more AI citations than ordinary pages but only when structured as a comparison benchmark, not as narrative analysis on July 6. Primary research pages earn 3.3x more AI citations than ordinary pages, but only when structured as comparison benchmarks. Only 8 of 301 cited pages in his dataset qualified as primary research, yet those 8 earned 90 of 1,075 total citations. The Fivetran warehouse benchmark, which ranked named platforms on speed and cost in a table, captured 44 citations alone. The mechanism is structural: AI retrieval for buying comparisons reaches for content shaped as a liftable comparison artifact, not for narrative analysis that contains the same data.
"The win is not 'we published original data.' The win is 'we published a benchmark that answers a buying comparison,' and almost nobody builds one."
Together, these two cards describe a single structural shift in how AI search rewards content. Citation and recommendation are now separate signals. Self-promotional framing earns citation while losing recommendation. Benchmark-shaped content earns citation by being structurally liftable, not by being authoritative in the traditional sense.
- Lily Ray, A brand's self-promotional "best of" listicle gets cited by AI Overviews 74% of the time but excluded from the recommendation 69% of the time. Self-promotional "best of" listicles are cited 74% of the time in AI Overviews but excluded from the recommendation 69% of the time, with named competitors benefiting instead.
- Kevin Indig, Primary research earns 3.3x more AI citations than ordinary pages but only when structured as a comparison benchmark, not as narrative analysis. Primary research earns 3.3x more AI citations than ordinary pages, but only when structured as a comparison benchmark with results first, methodology boxed, and a stable canonical URL.
Theme 2, Judgment in agentic workflows works better when self-assigned than pre-encoded
Simon Willison documented on July 3 what Cat Wu and Thariq Shihipar from the Claude Code team shared at the AIE 2026 conference. Telling a frontier model to self-route tasks to cheaper subagent models outperforms pre-written routing rules on both cost and quality captures the mechanism. Telling a frontier model to decide for itself which subagent tier to use per task produces better cost-to-quality outcomes than pre-writing routing rules. The model's in-context read of task complexity is more granular than any static heuristic. Judgment-heavy work stays in the main loop; implementation routes to cheaper subagents. The model's training has implicitly learned to distinguish task types that a rule set must approximate in advance.
"One of the most interesting tips I got from the Fireside Chat I hosted with Cat Wu and Thariq Shihipar from the Claude Code team at AIE on Wednesday was to let Fable (and to a certain extent Opus) use their own judgement rather than dictating how they should work."
This pairs with the broader Execution is becoming free; judgement is the part that doesn't compress pattern in the corpus. Self-routing is the practical instantiation of that pattern: judgment stays in the expensive main model; implementation routes to the cheap subagent. The decision about who does what is itself a judgment call, so pre-encoding it defeats the point.
Colin Matthews, writing in Lenny's Newsletter on June 30, describes the PM parallel in PM output compounds through three AI adoption ladders, each rung delivering an order of magnitude more than the one below. Three AI adoption ladders (personal output, product shipping, repeatable systems) each with ascending rungs where each rung shifts work from manual execution to autonomous delegation. The mechanism is the same as Willison's at the product level: the PM's judgment about what to specify stays human; the execution of that specification routes to agents.
"As you ascend each ladder rung, you get an order of magnitude more leverage."
- Simon Willison, Telling a frontier model to self-route tasks to cheaper subagent models outperforms pre-written routing rules on both cost and quality. Frontier models self-routing to cheaper subagent tiers based on their own judgment outperforms pre-written routing rules on both cost and quality.
- Colin Matthews, PM output compounds through three AI adoption ladders, each rung delivering an order of magnitude more than the one below. PM output compounds through three AI adoption ladders, with each rung shifting execution from the PM to agents while keeping task specification with the human.