Claim
The most consequential AI search behavior changes are not announced in changelogs; practitioners who rely on release notes alone will miss shifts that directly control which domains get cited, making behavioral tracking tools the primary signal for AEO work.
Mechanism
When OpenAI published a brief August 6, 2026 note about making "GPT-5.6 Sol more reliable with facts," there was no mention of search sourcing changes. Promptwatch tracking revealed that the site: operator's share in ChatGPT search queries jumped from 0.3-0.5% to 16-17% on August 8, coinciding with the update. The underlying search tool appears to accept a domain parameter rather than exposing the site: operator directly to users, meaning OpenAI can adjust domain-weighting silently through the model's tool-call behavior. The same update appeared to reduce Reddit's prevalence in results. Willison frames this as the emergence of a "GEO" (Generative Engine Optimization) discipline: reading AI system behavior through instrumented query analysis rather than through product documentation.
Conditions
Holds when: the AI search provider makes undisclosed changes to search sourcing, tool behavior, or domain weighting.
Fails when: the provider publishes explicit documentation of sourcing changes, making changelog monitoring sufficient; or when the tracked query sample is too narrow to detect real changes.
Evidence
Promptwatch data showed the site: operator share in ChatGPT search hovering between 0.3% and 0.5% for weeks before jumping to 16-17% on August 8, 2026. Willison notes the jump coincided with the GPT-5.6 reliability update. He also noted a simultaneous apparent reduction in Reddit-sourced results, though he could not confirm whether this reflected explicit system prompt changes or the domain weighting shift.
Signals
- Sudden drops in ChatGPT-attributed referral traffic to a domain that track with undocumented model updates
- AEO ranking tools showing volatility on dates when AI providers publish vague release notes
- Changes in which competitor domains appear in AI-generated answers without any corresponding SEO shift
Counter-evidence
Promptwatch sampling may not represent total ChatGPT search traffic; the temporal correlation with GPT-5.6 does not prove causation. The behavior may vary by topic category, making the 16-17% figure non-generalizable to all query types. Kevin Indig's data shows AI platforms deliver under 1% of organic traffic referrals for most sites, meaning citation tracking may matter more than visit tracking for most AEO practitioners.
Cross-references
- Telling a frontier model to self-route tasks to cheaper subagent models outperforms pre-written routing rules on both cost and quality: a related pattern where models make undisclosed routing decisions that affect output quality
- Single-task visual benchmarks have decoupled from frontier model rankings; agentic tool-calling reliability now differentiates model performance: how behavioral signals diverge from published capability claims