Google Agent Search now lets developers disable semantic search, KPI personalization and generative answers for each API request. The control is generally available in the v1 API, and it matters because applicable add-ons are billed and enabled by default under configurable pricing unless a request explicitly turns them off.
Google announced general availability for searchAddonSpec on September 14, 2026. It gives programmatic search clients three Boolean switches. A team can keep expensive features on for complex queries and send simpler requests through a lower-cost path.
The default is the first thing to understand
If searchAddonSpec is omitted, Google treats applicable add-ons as enabled. The same applies when a field is left unset. Under configurable pricing, every applied request add-on is billed through its corresponding SKU.
That turns an absent field into a spending decision. A migration that enables configurable pricing but leaves existing search calls unchanged can activate billable features whenever their prerequisites are satisfied.
Three flags control three different capabilities
| Request field | What true disables | What the add-on normally provides |
|---|---|---|
disableSemanticAddOn | Semantic query add-on | Embedding-based handling for longer and more complex queries |
disableKpiPersonalizationAddOn | KPI and personalization add-on | Event-based re-ranking and personalization for business KPIs |
disableGenerativeAnswerAddOn | Generative answer add-on | AI overview features including summaries and follow-up questions |
The REST reference uses negative Boolean names. Setting a field to true turns the capability off. That is easy to reverse mentally, so request builders should use explicit configuration objects and tests rather than scattered Boolean literals.
A minimal request can make the cost intent visible
{
"query": "BigQuery data platform",
"searchAddonSpec": {
"disableSemanticAddOn": false,
"disableKpiPersonalizationAddOn": true,
"disableGenerativeAnswerAddOn": true
}
}This example keeps semantic query enabled while turning off personalization and the AI overview. The exact choice should come from query intent, not from one universal cheapest setting.
The add-ons have dependencies and eligibility rules
Semantic query depends on semantic embeddings being available. Google says the semantic add-on is applied automatically when that prerequisite is met unless the request disables it. KPI and personalization charges apply when data quality reaches the required tier and the training pipeline is ready. The AI overview depends on semantic query.
A flag therefore controls application, not necessarily capability readiness. Observability should record both the requested state and what Google actually applied. Without that distinction, a team can misread a zero charge as successful cost routing when the add-on was merely ineligible.
A query policy can map cost to user intent
| Query class | Likely configuration | Reason to test |
|---|---|---|
| Exact identifier or known document | Disable all three add-ons | Keyword filtering may already be sufficient |
| Long natural-language discovery | Keep semantic; consider disabling the other two | Meaning may matter more than personalization or summaries |
| Personalized recommendation | Keep semantic and KPI personalization | Ranking quality may justify the extra feature |
| Answer experience with citations | Keep semantic and generative answer | The user expects synthesis rather than a result list |
| Background batch retrieval | Start with all add-ons disabled | Machine workflows may not need an AI overview |
These are test hypotheses, not universal defaults. Build a labeled query sample, compare relevance and task completion under each combination, and attach the current SKU cost. The cheapest request that fails the user’s job is not an optimization.
The routing rule should also be explainable. Log the query class, the three requested flags, the features Google applied, latency, result engagement and attributed cost under one trace identifier. That record lets product and finance teams answer a practical question later: which add-on improved an outcome enough to justify its charge? A feature toggle without outcome telemetry only moves uncertainty from the invoice into the application.
Configurable pricing still has subscription mechanics
Google describes configurable pricing as a mixed model with storage and search-query subscriptions plus optional add-ons. Storage is measured in GiB and search capacity in queries per minute. Overage can fall back to the general consumption model when usage exceeds a threshold.
Billing begins when configurable pricing is enabled, and apps and data stores must be enrolled individually. Google also warns that a Gemini Enterprise data store should use the general pricing model. Per-query switches solve feature selection; they do not replace subscription sizing, enrollment checks or overage monitoring.
Four tests prevent surprise charges
- Send a request with the specification omitted and confirm the applied add-ons.
- Send all three flags as true and verify that no request add-on is billed.
- Turn on one feature at a time and measure relevance, latency and cost.
- Replay the same test when embeddings, personalization training or answer configuration changes.
Our Google Agent Platform sandbox guide covers a different control plane for agent execution and persistent state. The Gemini 3.8 Live migration guide shows why status and pricing defaults should be tested at the request level.
The practical verdict
searchAddonSpec is a small API field with a large operational consequence. Google Agent Search customers can now align paid features with individual query value, but the safe posture is explicit configuration. Under configurable pricing, silence can mean “turn applicable add-ons on.”
Primary sources
- Google Agent Search release notes
- Google SearchAddonSpec REST reference
- Google configurable pricing guide
Checked September 15, 2026. API behavior and billing defaults come from Google Cloud documentation. Query policies and test recommendations are MustHave.ai analysis.
