AI Provider Statement
This is a translation; the Turkish text is binding. AI Sağlayıcı Beyanı (Türkçe)
This statement explains what artificial intelligence does in Kararis, which data is sent to which provider and where this chain ends.
It is a separate document because it is the text that will change most often: the choice of provider and the model chain change much faster than the privacy policy. If it were embedded in the other texts, every provider change would raise the version of the privacy policy.
1. The boundary of artificial intelligence in the product
Kararis's decisions are not produced by artificial intelligence. Decisions, metrics and priorities are produced by a deterministic rule engine that calculates them from source data. The job of artificial intelligence is to turn the evidence produced by this engine into readable text.
Concretely, artificial intelligence:
- does not calculate a new metric and does not change existing figures;
- does not produce a new decision or recommendation; it explains the decision produced by the rule engine;
- does not rank priorities;
- makes no change to your advertising accounts.
The figure you see on the same screen does not change even if the AI text is regenerated.
2. Artificial intelligence is optional
The artificial intelligence feature can be turned on and off at the organisation level. When it is turned off the product keeps working: decisions, reports and audit rows continue to be produced; only the explanatory texts are not produced and no data is sent to any provider.
Text generation is also user-triggered: the monthly report assessment, the deep analysis text, the social content summary and the decision explanation are not generated unless requested on the relevant screen.
3. What is sent to the provider
The data sent varies with the type of text to be generated and is in every case limited to the fields needed for that text:
| Generated text | Data sent |
|---|---|
| Decision explanation | The structured evidence of the decision — only a predefined list of fields. No free-form data is sent. |
| Monthly report assessment | The total performance metrics for that month and the decisions of the same month |
| Deep analysis text | Periodic performance summaries and creative breakdowns |
| Social content summary | Organic content performance measurements and search term summaries |
| Creative attribute extraction | The text and name of the ad |
Not sent: API keys, tokens and passwords; user names, email addresses and passwords; the audit log; the names, email addresses, telephone numbers and addresses of the brand's end customers — these fields are not stored on the platform in any case (see the Data Retention and Deletion Policy, section 1); individual order records, order numbers and order IDs.
Shopify data. Order records received from a Shopify store are protected as personal data belonging to the store's customers even though they carry no name, email, telephone or address (Privacy Policy section 8). From this data only aggregate measurements go to a provider, and this content is sent only to a provider that commits not to use data sent through the API for model training; it is not sent to the router providers in section 5. If no such provider is available, no AI text is generated for that content. The same limit applies to data received from Google APIs (Privacy Policy section 7.d).
4. The provider chain
The active provider and model are selected from the administration panel. The providers the platform supports today:
| Provider | Location |
|---|---|
| Google Gemini | USA |
| Anthropic Claude | USA |
| OpenRouter (router) | USA |
| xAI (Grok) | USA |
| Groq | USA |
| DeepSeek | China |
| LlamaIndex | USA |
Which provider is active for an organisation is disclosed on request.
5. Router providers and why we cannot give a fixed list
Some of the providers above — in particular OpenRouter — are not model providers themselves but routers. When such a provider is selected, the request is handed over to a sub-provider according to the router's capacity, price and availability at that moment.
This has three consequences, and we state all three openly:
a. Routing is dynamic. The same feature, on the same day, may go to two different sub-providers in two different requests. For this reason it is not possible to give a fixed list of sub-providers; if one were given, it would be wrong.
b. The sub-provider may be in a different country. For example, a request made through OpenRouter has been observed to be routed to one of the DeepSeek models; DeepSeek is established in China.
c. In return, every request is recorded. For every AI run, the platform records which provider was called and which model responded. You can ask which model produced a given text; it is answered from the record.
If you do not want to use a router, a direct model provider can be selected for your organisation, or the artificial intelligence feature can be turned off entirely.
6. Model training
Kararis does not use your data to train a model on its own behalf and does not sell data to providers. We work with providers through the API, not through consumer interfaces.
To be honest, however: whether data sent through the API is used for training by the provider concerned is subject to that provider's own terms, and when a router is used, because the chain is dynamic, we cannot give an end-to-end guarantee on this point.
Content generated from data received from Google APIs and from Shopify is outside this uncertainty: that content goes only to a provider that gives a training commitment (section 3).
This uncertainty is the reason why the artificial intelligence feature is optional and can be turned off. If you do not accept this uncertainty for your sensitive data, turn the feature off: the product also works without artificial intelligence.
7. Stored AI records
| Record | Content | Period |
|---|---|---|
| Run telemetry | The provider called, the model that responded, duration, error status and the text of the request sent and the response received | 12 months |
| Generated texts | Monthly report assessment, deep analysis text, decision explanation, social content summary | For as long as the relationship lasts |
| Generation request ledger | By whom, when and for which subject the request was opened | 12 months |
It is deliberate that the telemetry record contains the request and response text: only in this way can it be answered why a text was generated as it was and which model produced it. This record is used only for fault diagnosis, quality measurement and accountability.
8. Transfer abroad
All of the providers above are established abroad. These transfers are made with appropriate safeguard mechanisms under Article 9 of the KVKK. If the artificial intelligence feature is turned off, this transfer does not take place at all.
9. Assistant connection (MCP)
The connection surface that lets you read Kararis's decisions, audit status and evidence through an AI assistant of your own choice (for example Claude or ChatGPT) is open. The conditions announced in advance in version 1.0 of this section remain in force unchanged; no condition has been relaxed below.
The connection can be established in two ways — from the panel, or from a consent screen started by the assistant client — and in both cases it can be revoked at any time by your organisation's administrator.
a. The connection is established with explicit consent and can be withdrawn. Unless a connection is established, no data goes to any assistant. There are two ways to establish the connection, and both rest on a moment of consent:
- When you establish it from the panel, a secret is generated and shown to you once; this secret is not stored in Kararis and cannot be shown again; you copy it and carry it to your assistant client yourself.
- When you start it from the assistant client, Kararis shows you a consent screen. That screen is the moment of explicit consent: there you see and approve the client's name, the address the data will go to, which organisation and brand scope can be read and the duration of the connection. When you approve, the secret is given directly to the client; it is not shown to you and, again, it is not stored in Kararis.
In both ways the duration of the connection is limited — an unlimited connection cannot be created — and access ends by itself when the duration expires. The duration of a connection established from the consent screen is not extended by the client renewing its secret. Revocation takes effect immediately.
b. You choose the model. This is a different class of transfer from the situation in section 4: there, Kararis selects the provider and records the chain. In the assistant connection you choose the model, the data flows into your own session, and how long the model you choose retains the data, how it uses it and whether it uses it for training cannot be known by Kararis. In that relationship the assistant provider acts on its own terms and under its own responsibility.
c. Read only. The connection is read-only: it makes no change to your advertising account, does not produce a new decision and does not open free queries on raw data. There is no write capability defined in the product.
d. The authorisation boundary applies unchanged. The assistant can read only data within the organisation and brand scope you are authorised for; an unauthorised target is not narrowed, it is refused.
e. Every call that reaches the surface is written to the ledger and the connection can be visibly revoked. The ledger is subject to the audit log class of the data retention and deletion document. Requests that hit the rate limit and never reach the surface are not recorded individually; these requests read no data.
f. Data classes the assistant can read. The connection reads the following classes already produced in the panel and cannot go beyond them:
| Data class | Content |
|---|---|
| Brand summary | Connected sources and their freshness, period, number of open decisions, currency |
| Decision list | Produced decisions; finding, class, priority, monetary scale and confidence |
| Evidence reference | The source, period, method, sample and freshness of a figure |
| Audit rows | The five states of the structural audit and the canonical text of the row |
| Report archive index | The ID, period and delivery status of stored reports — the report body is not returned |
| Connection health | The status of platform accounts and the freshness of data sources |
| Monthly summary measurement | The canonical monthly report of a calendar month: spend and platform breakdown, source reconciliation, funnel, commercial result, product breakdown, the calculated results of the profit waterfall, booking result and freshness |
| Campaign breakdown measurement | The campaign rows and performance metrics of the selected period |
| Product profitability measurement | Profitability per product and variant; this class also contains the cost information entered by the brand itself (unit cost, cost source, channel commission rate and cost of goods sold rate) |
| Traffic source measurement | The channel and source/medium breakdown of web analytics: the distribution of sessions, engaged sessions, key events, purchases and analytics revenue |
| Search term measurement | The search terms of the advertising account and the organic search queries of your site — together with the term and query text — and the number of terms seen in both |
| Ad breakdown measurement | The distribution of ad spend and results by placement, demographics, geography, device and hour |
| Attribution window measurement | Sales count and revenue according to the advertising platform's own attribution windows (view and click), per campaign and ad |
| Weekly summary measurement | The canonical summary of a week: spend, the brand's primary result and efficiency, the web analytics funnel, the previous week's counterpart and percentage change for each item, the number of decisions opened and closed that week and freshness. The explanatory text shown in the panel is NOT INCLUDED in this class and is not returned |
| Decision explanation (AI) | The explanation text produced by artificial intelligence for a decision; returned only within the decision list class, only while the evidence it rests on is still current and clearly marked as produced by AI |
| On-site search measurement | The terms visitors typed into your site's own search box — together with the term text — and the sessions, purchases and analytics revenue of these terms |
| Keyword measurement | The keywords purchased in Google Ads — together with the keyword text, match type and ad group name — spend, clicks, conversions and search network impression share |
| Landing page measurement | The address the ad click actually went to and the spend, clicks and conversions per address |
| Product status measurement | The point-in-time status of the products in Google Merchant Center: the number of products per serving status and stock declaration, the number of products per issue code and the rows of products not fully serving — together with the product title |
| Deviation band measurement | The days on which five metrics (spend, primary result, efficiency ratio, sessions, purchases) fall outside the band built from the last 28 days: the observed value, the edges of the band and the reason when the band could not be built. This class carries no judgement — it contains no severity label, recommendation or forecast |
| Brand list | The list of brands this connection can see: brand name, business model, currency and which source types are connected. This class is not a measurement — it contains no figures, freshness or connection health |
| Order attribution measurement | The aggregate distribution of store orders read from the order's own visit sessions: order count and net revenue per traffic channel, UTM tags, device and operating system, landing path and first/last click journey. This class does not contain individual orders: no order number, order ID, visit/visitor ID, IP address, referring address or any field belonging to the customer is returned — these fields are not stored in Kararis in any case. Rows carrying three or fewer orders are combined into a single “other” row |
| Order payment and customer measurement | The aggregate payment method mix of store orders, the number of pending payments and refunds, discount/coupon/shipping measurements, tax rate totals and the shares of first orders and guest checkouts; lossless category totals in product profitability and separate aggregate tax evidence. No individual order or customer row, customer ID, coupon code, payment provider name or unverified delivery time is returned. The category name is the brand's own text at the provider and is marked as data |
| Abandoned cart and store setting measurement | The store's periodic abandoned checkout count and total value, recovered count, recovery email send share, checkout-to-order conversion rate; the free shipping threshold and a fixed-shape aggregate summary of shipping/tax settings. No individual checkout, customer, email, address or telephone field is returned. Store texts in the shipping/tax settings are not passed to the assistant |
| Incremental attribution measurement | The estimate of the advertising platform's own incremental model — the sales count and revenue estimated to have been caused by the ad — and its ratio to the result reported on the advertising platform, per campaign and ad. This is not an experiment result. It comes from the same pull as the attribution window measurement; it contains no new data type and carries no text other than campaign/ad names |
| Daily order series measurement | The order count and net revenue (in the local currency and in the presentation currency) for each closed day of the store, at most 92 days. It contains no order rows, customers or products; the same figures already appear in the monthly summary measurement, and it contains no new data type |
| Google device breakdown measurement | The distribution of the spend, clicks, impressions, conversions and conversion value of Google Ads campaigns in the selected period by device type (mobile, desktop, tablet, connected TV, other and unrecognised) and the ratios derived from them. It is the device breakdown of the same ad figures in the campaign breakdown measurement and contains no new data type |
| Google hour breakdown measurement | The distribution of the spend, clicks, impressions, conversions and conversion value of Google Ads campaigns in the selected period by hour of the day (0–23, in the advertising account's own time zone) and the ratios derived from them. It is the hour breakdown of the same ad figures in the campaign breakdown measurement and contains no new data type |
| Google province breakdown measurement | The distribution of the spend, clicks, impressions, conversions and conversion value of Google Ads campaigns in the selected period by the province where the ad was shown (81 provinces, abroad as a single bucket, and unmatched location) and the ratios derived from them. It is the province breakdown of the same ad figures in the campaign breakdown measurement; it contains no sub-province or person-level location and is not joined with the order province |
| Blog post measurement | The title, web address (URL), category and tag names and first publication date of the posts published on your site's Wix Blog; the number of views, likes and comments gained in the selected period and the click and impression totals of the same page in Search Console. It does not contain the post body, the author or comment texts; the title, address, category and tag are your site's own text and go to the assistant marked as untrusted text |
The abandoned cart and store setting measurement entered the list with
version 1.6. The new class carries only aggregate and fixed-shape
measurements; it does not open an individual checkout or customer grain. Free
texts belonging to the store in the shipping and tax settings are not sent to
the assistant. The actual delivery time is not included in this class; because
of delivery_timestamp_unverified the delivery audit remains a declaration.
Because no new provider chain or new connection purpose arises, and the class stays within the same read-only connection established with the existing explicit consent and within the fixed projection produced in the panel, no new consent round is needed. The existing consent can always be withdrawn; the old version 1.5 remains accessible.
The brand list entered the list with this version and is NOT a new data class. The assistant previously did not know which brands it could read and found them by trying brand IDs one by one. This class makes that trial unnecessary: only brands this connection can already read appear in the list, and the brand's name was already returned within the brand summary class. The list is not a measurement — it contains no spend, revenue, decision count or freshness — and an empty list means not "no brands" but that there is no brand within this connection's scope. No new provider chain arises and no new consent round is needed.
The measurement classes are fixed-shape readings of ready-made views that produce the same calculation in the panel: the assistant cannot choose metrics, dimensions, filters or sorting; no free queries on raw data are opened. An answer that is not in the panel is not in this connection either.
The decision explanation entered the list with this version. Up to 1.4, no text written by artificial intelligence was returned from this connection. In this version the explanation has become readable only within the decision list class, and the gate is narrow in three places at once: no measurement class returns an explanation, the text is given only while the evidence it rests on is still current (when the evidence changes the text is not shown, section 10), and the text is given to the assistant with an explicit statement that it was produced by artificial intelligence — it is asked not to follow the instructions in it, and the text does not replace the evidence.
Product status entered the list with this version. The class reads whether the products in the brand's Google Merchant Center account are considered publishable in ads. The text itself states three limits:
- It is a point-in-time observation, not a period total. The answer belongs to a single polling day; for a day that was not polled it cannot be said that a product "was serving".
- The stock declaration is your feed's, not your store's inventory. This class does not measure the stock in your warehouse; it reads what Merchant Center declares.
- The product title is your text and is given to the assistant marked as data — it is explicitly asked not to follow the instructions in it. The free English text of Google's issue descriptions is not stored at all and is not returned in this connection either; only the closed code name of the issue is returned.
No new provider chain arises and no new consent round is needed: the class is a fixed-shape reading of a view already produced in the panel.
Search term text entered the list with version 1.4. Up to 1.3, raw search terms were also outside the list; in this version they have become readable, only within the search term measurement class, as the text users typed into the search box. The text is given to the assistant marked as data — it is explicitly asked not to follow the instructions in it — and is not written to Kararis's access ledger.
The deviation band entered the list with this version and carries no judgement. The band is a comparison of already readable measurements with their own recent history: no new data class is collected, no new provider chain arises and no figure that is not visible in the panel is produced. A day outside the band does not mean "bad"; this class has no severity label, recommendation or forecast, and when the band cannot be built, the reason is returned, not zero.
On-site search text entered the list with this version. In 1.4 only the advertising account's search terms and organic search queries could be read; in this version the text visitors typed into your site's own search box has also become readable. As with search terms, the text is given to the assistant marked as data — it is explicitly asked not to follow the instructions in it — and is not written to Kararis's access ledger. The on-site search measurement is populated only if enhanced measurement is turned on in your web analytics account; if it is off this class returns empty, and the emptiness does not mean "nobody searched".
The web and app split was added with this version. The traffic source measurement and the monthly summary measurement also carry the sessions, purchases and analytics revenue of the same period split into web / Android / iOS. It is not a new data class: it is a second breakdown of the same measurement and contains no person-level information.
The keyword and landing page measurements entered the list with this version. Both come from your advertising account's own configuration: the keyword is the target the brand CHOSE to buy — it is not the text users typed into the search box, which is a separate class (search term measurement) — and the landing address is the address of the brand's own site. Both are given to the assistant marked as data; it is explicitly asked not to follow the instructions in them. Ad texts and free description fields remain OUTSIDE the list in this version as well: although the text of the purchased target has been opened, the text of the published ad has not.
Impression share is an ESTIMATE, and Google does not return it at all for below-threshold impressions; therefore an empty share does not mean "your share is zero" but "it was not measured that day". The landing page measurement, on the other hand, does NOT COUNT that the page loaded: it says which address the click went to, not whether that address was reached.
Among the items outside this list are the following, and they are not opened to the connection: ad texts, free text fields other than the decision explanation, order and customer details, internal alert and threshold records, the audit log.
10. Human oversight
Text produced by artificial intelligence is advice, not a decision. The text is presented together with the numerical evidence beneath it; when the evidence changes, the old text is not shown. The final decision is made by the user of the platform.
11. Changes to this statement
When the provider chain changes, a new version of this document is published. Every version carries a number and an effective date; old versions remain accessible.
Version note (1.7). The operator and data controller of the platform has been updated to Zehra Eslem Karaca (04.09 Digital), and the contact address has been written in full as the business address in the tax registration. The provider chain, the data classes and the other provisions have not changed.
Version note (1.8). The statement in section 3 that "personal data belonging to the brand's end customers is not stored on the platform" has been corrected so that it does not conflict with Shopify order records being protected as personal data belonging to the store's customers: the fields not stored are name, email, telephone and address. Section 3 now states that only aggregate measurements are sent from Shopify data and that this content goes only to a provider that gives a training commitment, and section 6 now states that this limit is outside the model training uncertainty. The provider chain, the data classes and the other provisions have not changed.
Version note (1.9). Blog post measurement has been added to the table in section 9.f: the title, address, category and tag names of the published posts on Wix Blog, the number of views, likes and comments in the period and the Search Console clicks and impressions of the same page. The post body, the author and comment text are not sent. The provider chain and the other provisions have not changed.
12. Contact
To ask which provider is active, which model produced a given text, or to have the artificial intelligence feature turned off:
Zehra Eslem Karaca (04.09 Digital) Kayabaşı Mah. 75. Yıl Cad. 1. Bölge -Akik- C8 Blok No: 4 B İç Kapı No: 25, Başakşehir/İstanbul Email: kvkk@kararis.com