Vertex AI Search for retail pricing
Prices are listed in US Dollars (USD). If you pay in a currency other than USD, the prices listed in your currency on Cloud Platform SKUs apply.
Search charges
Search enables you to provide high quality product results that are customizable for your business needs. You can leverage Google's query and contextual understanding to improve product discovery across your website and mobile applications.
The only search operations that incur charges are requesting search or browse results by calling the Search method. There's no charge for importing or managing user events or catalog information. There is also no charge for using the pretrained Recommendations LLM.
Search and browse queries are charged at $2.50 per 1000 requests.
Example
This example illustrates how search queries are charged.
In this example, a customer's application made 15 million keyword search queries and 10 million browse queries in one month. Here is how we would calculate the total cost to the customer:
- Search queries = 15 million
- Browse queries = 10 million
- Total queries for the month = 15 million + 10 million = 25 million
- Vertex AI Search for retail search pricing = $2.50/1000 queries
Total cost to the customer = 25 million queries x $2.50/1000 queries = $62,500
Recommendations charges
Free trial: You can try recommendations with $600 free credits. Free credits are automatically granted when you sign up, and expire six months after signing up. These credits are granted to your billing account, and are not affected by the number of projects connected to your billing account. For example, if you have three projects linked to your billing account, your billing account still receives $600 of free credits. These credits are typically sufficient to train a model and test its performance in production through a two-week A/B test. See Implement Vertex AI Search for retail.
There's no charge for importing or managing user events or catalog information. The only recommendations operations that incur charges are training, tuning, or requesting predictions by calling the predict method.
Training (per node per hour) costs are charged on a daily basis if your model is actively training or if you have submitted a request to resume training. Once you pause or delete the model, you will no longer be charged. See the documentation for managing training.
Tuning (per node per hour) costs for active models are charged after the tune completes successfully. You will only be charged for an incomplete tune if you pause or delete a model during an ongoing tune. In this case, you will then be charged for the node hours that were consumed before the model tuning stopped. See the documentation for managing tuning.
| Prediction requests per month | Price per 1000 predictions |
|---|---|
| Up to 20,000,000 | $0.27 |
| Next 280,000,000 | $0.18 |
| After 300,000,000 | $0.10 |
| Feature | Price |
|---|---|
| Training and tuning | $2.50 per node per hour |
Examples
Example A
This example illustrates how each tier of pricing for monthly prediction requests is applied.
In this example, a large retailer's application made 1,000,000,000 prediction requests in this particular month. It trains three models which, by default, automatically retrain once per day. This amounts to about 500 node hours of model training a month. By default, recommendations models are tuned quarterly; in this example, the model tuning accrued about 300 node hours per tuning, which on a monthly basis would be 100 node hours.
To calculate the cost for this month, we'll first find the cost of prediction requests. Pricing is calculated in 1000-request blocks, and cost is tiered by number of monthly prediction requests.
- First 20,000,000 predictions = 20,000,000 predictions / 1000 * $0.27 = $5,400
- Next 280,000,000 predictions = 280,000,000 predictions / 1000 * $0.18 = $50,400
- Next 700,000,000 predictions = 700,000,000 predictions / 1000 * $0.10 = $70,000
Next, let's calculate the cost of training and tuning.
- Training charge = 500 node hours * $2.50 = $1,250
- Tuning charge = 100 node hours * $2.50 = $250
The total cost of predictions, training, and tuning in the given month is $127,300.
Example B
This example illustrates a lower volume use case.
In this example, a retailer makes 10,000,000 prediction requests a month and trains a single model per day, which, by default, automatically retrains once per day. This amounts to about 150 node hours of model training per month. The model's quarterly tuning accrued about 90 node hours per tuning; to find the cost per month, we'll use the monthly average, 30 node hours.
Let's calculate the price for one month of usage. Because this retailer's number of prediction requests this month doesn't exceed 20,000,000, requests will all be charged at the first tier of pricing, $0.27 per 1000 requests.
- 10,000,000 predictions = 10,000,000 predictions / 1000 * $0.27 = $2,700
To calculate the cost of training and tuning:
- Training charge = 150 node hours * $2.50 = $375
- Tuning charge = 30 node hours * $2.50 = $75
The total cost of predictions, training, and tuning in the given month is $3,150.
Google Cloud Observability charges
Vertex AI Search for retail logs an error to Google Cloud Observability for each API request that results in an error, such as a user event request that contains malformed JSON, or a catalog item import request with a negative price. Vertex AI Search for retail also logs an error for every prediction request with a catalog item that is not in the imported catalog.
Google Cloud Observability charges by the GiB of logs stored. (Logs are retained for one month.) The first 50 GiB of logs per month per project is free. After that, Google Cloud Observability charges $0.50 per GiB of logs. The size of the logging data depends on the size of your JSON payload, but a GiB would be approximately 200,000 recommendations errors.
For more information, see the Google Cloud Observability pricing page.
What's next
- Read the Vertex AI Search for retail documentation.
- Try the Pricing calculator.
- Learn about Vertex AI Search for retail solutions and use cases.