Compare
How Siper compares with the alternatives.
First the systems we tested on the same Turkish documents. Then the kinds of product you may own already, and the alternatives to buying a tool.
Results
Siper has the highest recall among the 16 systems we evaluated.
Measured 2 November 2026, "Turkish and English" option. All systems ran on the same 403 human-verified Turkish documents. Siper finds 85.7% of the personal data in this set. The table lists 9 of the 15 alternatives.
| System | Recall | False alarm | Median latency | Can run on site |
|---|---|---|---|---|
| Gurubase Siper | 85.7% | 14.3% | 176 msOne 8-vCPU server, no GPU | Yes |
| GPT-5-mini | 69.3% | 12.1% | 18,115 msCloud API | No |
| Gemini 2.5 Flash | 52.2% | 1.9% | 1,790 msCloud API | No |
| Limina (Private AI) | 51.5% | 48.6% | 561 msCloud API | No |
| Qwen2.5-32B | 48.3% | Not measured | 19,242 msLocal GPU | Yes |
| Azure AI Language | 26.7% | 3.4% | 221 msCloud API | No |
| Google Cloud DLP | 24.5% | 40.2% | 220 msCloud API | No |
| Tonic Textual | 22.2% | 42.7% | 491 msCloud API | No |
| Microsoft Presidio | 19.2% | 98.1% | 7 msLocal CPU | Yes |
| OpenAI Privacy Filter | 8.9% | 18.4% | 4,600 msOne 8-vCPU server, no GPU | Yes |
On-prem systems: median per document on one 8-vCPU server without GPU, unless the row states another setup. Cloud services: API response time including network. Systems ran on different infrastructure, so this is not a direct hardware comparison. The benchmark gives the method, the confidence intervals and the results per category.
Product types
Three kinds of product you may compare with Siper
This table describes each kind of product from its vendors' public descriptions. It names no product.
| Kind of product | What it focuses on | What to check |
|---|---|---|
| AI gateways | Model access and routing. | Whether it detects Turkish special-category data. The public descriptions we read do not offer this. |
| PII detection and masking | Generic detection across many languages. | How much it finds in Turkish sentences. A cloud PII API also raises the question of transfer abroad. |
| Network and DLP layer | Traffic and data rules across the company. | Whether it finds KVKK Article 6 special-category data and reads Turkish in context. |
Firewall and DLP
If you already own a firewall or DLP
Keep your firewall and DLP. Test how much Turkish data they find on your own text. Finding an 11-digit ID is a number pattern. Finding "epilepsi tanısı" ("epilepsy diagnosis") inside a sentence is a different job.
Alternatives
Alternatives to buying a tool
Each alternative has a real reason behind it. The table gives that reason and our reply.
| Alternative | Why teams choose it | Our reply |
|---|---|---|
| A suite you already pay for, which blocks rather than masks | It needs no new vendor. | A suite that blocks stops the request. Siper masks the personal data it detects, so the work continues. |
| An enterprise AI contract | The contract sets how the provider may use your data. | Real names still reach the provider, even with a good contract. A contract does not remove the transfer question. |
| Your own LLM on your own servers | The prompts stay in your building. | Siper can route to your own model server, and it masks detected personal data before the request reaches the model. |
| A ban | It costs nothing. | KVKK advises against full bans. |
| Policy only | Your legal team is satisfied. | A policy states a rule. It does not change what is in the prompt. |
Next step
Compare on your own text.
Only a test on the same documents is valid.