Three new cards from two operators. One from Kevin Indig on diagnosing AI category-level decline. Two from Hamel Husain on sequencing and cost-cutting in AI product engineering.
Theme 1, Category repricing vs. competitive re-ranking
Kevin Indig's Growth Intelligence Brief #22 covered 30 days of AI search visibility data ending July 15, 2026. Most of that brief was captured on August 9. One distinct claim was not: the synchronized decline pattern across an entire vertical.
When one brand loses AI visibility and another gains, that is competitive re-ranking. When every brand in a vertical declines simultaneously, with no single competitor absorbing the redistributed share, the AI surface is doing something different. It found a substitute for the category's core function and stopped routing users to any external site. For stock photos, all 9 named brands declined between 8.7% and 44.7% in the same 30-day window. Indig's read: "A decline this broad and synchronized reads as category repricing rather than nine separate site problems." The correct response is a pivot out of the category, not better positioning within it. When every brand in a content category declines together in AI search, the platform is repricing the category, not re-ranking competitors.
- Kevin Indig, When every brand in a content category declines together in AI search, the platform is repricing the category, not re-ranking competitors. Synchronized AI visibility decline across all brands in a vertical signals the platform found an in-place substitute. The category is being repriced, not re-ranked.
Theme 2, Sequence and cost in AI product engineering
Hamel Husain published his AI Product Engineering Notes index on August 12, covering the 13-session curriculum he runs as part of the AI Evals for Engineers course. Two distinct claims are extractable.
The first is a sequencing rule. Teams improving AI products tend to jump to post-training when retrieval improvements would close most of their error cases at a fraction of the cost. Husain's hierarchy: retrieval and context first, systems and harness second, post-training only after exhausting both. Five of 13 sessions address context improvements. Post-training requires labeled data, a compute budget, and a retraining loop measured in days to weeks. Retrieval improvements use the same model more effectively at no model cost. The hierarchy reflects cost and reversibility, not opinion. Optimize retrieval and context before touching systems, and post-train only after exhausting both.
The second is a routing technique that belongs in the systems-and-harness stage of that sequence. Most classification inputs cluster at the high-confidence end of a small model's probability distribution. For those inputs, the small model's prediction matches the large model's output, and the expensive call adds nothing. A confidence threshold routes high-confidence inputs through and escalates uncertain ones to the large model. Average inference cost falls because the escalation fraction is typically small, and large-model accuracy is preserved on the full task. Route classification tasks to a small model when confident and a large model when uncertain to preserve accuracy while cutting inference cost.
- Hamel Husain, Optimize retrieval and context before touching systems, and post-train only after exhausting both. Retrieval and context improvements should come before systems work and before post-training; each stage is cheaper and faster than the next.
- Hamel Husain, Route classification tasks to a small model when confident and a large model when uncertain to preserve accuracy while cutting inference cost. Route classification tasks to a small model when confidence is high and a large model when it isn't; average inference cost falls while full-task accuracy holds at the large-model baseline.