-1.png?width=1132&height=292&name=Asset%201@2x%20(12)-1.png "Odins.ai Logo")
All your marketing data in one structure, and a model you can take apart.
Spend sits in twenty platforms and an agency's spreadsheets, sales sit in your warehouse, and no version of the truth is agreed on. Odins collects and structures every source, offline included, into one dataset in your hierarchy, delivers it to your warehouse, and builds a Bayesian model on top whose priors, curves and forecasts are yours to inspect, in the platform or from your own tools.
Data
Every channel, online and offline, plus sales and revenue. Structured, fully managed.
Model
Bayesian MMM, built on what your team already knows and verified against actuals.
Decisions
Budget recommendations, scenarios and reports. Reviewed by our team before delivery.
Results feed the next model, and the loop repeats.
Twenty sources, no single version.
The marketing team wants an answer that cuts across channels. The pipelines were built one platform at a time, and the question falls between them.
Every source in its own shape
Each platform's API has its own schema, currency and time zone. The TV plan is a spreadsheet from the agency. Someone maps it all into one table, and maps it again when a schema changes.
Offline is a black hole
The largest budget lines have no API. They arrive as monthly totals, if they arrive at all, so nothing downstream can see which weeks the spots ran or what a flight did to search.
The model nobody maintains
A model built once on Robyn or Meridian describes last year. Retraining it every month takes someone who knows Bayesian statistics and has the time, and the knowledge leaves with them.
What you get
Two data layers and an open model.
Everything in this table is part of the service. The two data layers are a product on their own, and some customers use Odins for them alone.
Delivery to your own warehouse is included. The model is open in the platform and through the MCP.
Built by our data scientists, open to yours.
Our team builds and retrains the model. Nothing about it is hidden from the team that has to trust it. The question to ask any vendor is how they set their priors; here is where you can check ours.
- Every prior is documented: where it came from (historical spend, the platforms' own conversion data, structured interviews, your experiments) and what the posterior became
- Saturation curves, carryover and the weekly decomposition into baseline, control variables and channel contributions, per channel, with ranges
- The control variables the model uses, and why there are deliberately few of them
- A verification view with forecasts stored before the actuals arrived, week by week, and every earlier model kept so you can check how it held up
- What the model does not do, stated plainly: no cross-channel interaction effects, no ad-level claims the data cannot carry, and long-term brand lift as a beta rather than a standard output
From access to a delivered dataset.
What happens on our side, and the little that happens on yours.
Connect
You grant access through an onboarding link, OAuth or an API key per source. We load two to three years of history where the platform allows it, and read directly from BigQuery or similar where you already hold structured data.
See data integrationStructure
Managed pipelines sync every source daily. Everything lands in your hierarchy at a daily grain in one currency, offline next to digital, sales next to spend. Our team monitors every sync and fixes what breaks when an API changes.
See the connectorsModel
Priors are set with you in two or three meetings. The first model is typically live four to six weeks after kickoff, data collection and structuring included, and its forecasts are checked against actuals before anyone plans on it.
How the model worksDeliver
The structured dataset and the raw tables flow to your warehouse. The model is open in the platform and through the MCP, which your own agents can use as the API. Data flows in daily; the model is retrained monthly.
Your data: ownership and exportWho does what
What stays with your team, and what moves to ours.
The boundary most data teams want: you keep the warehouse, the BI and the questions; we take the plumbing and the modeling.
Stays with you
- Your warehouse and BI tools, with the Odins layers as tables in them
- Your own analyses on the raw tables, ad level included
- The experiments you run; we take the results into the model as priors
- The decisions: what goes into action each month
Moves to Odins
- Setting up, maintaining and monitoring every connector, API changes included
- Offline pipelines from your media agency, broken down to the spot
- Building the model, setting the priors with you, retraining it monthly and verifying it
- Reviewing every recommendation before it reaches the marketing team
Both ways
Delivery to your warehouse is included, not an add-on. Some customers use Odins only for the structured data and add the model later on the same foundation.
The MCP is the API.
Odins is an MCP server. Connect Claude, ChatGPT, Microsoft Copilot or your own agent and work with the data and the model without the interface. The large majority of platform use across our customers goes this way.
- Forecast, create scenarios, pull spend and sales, and query the raw data at ad level, from a conversation or from an agent
- The computation runs on our side. The assistant gets the same numbers the platform computes, with the meaning of each value encoded so an agent can interpret the results
- The model's setup is exposed through the same door: the channels, the priors, the control variables
- Point an agent at the MCP to feed your own dashboards. There is no separate read-only API, because the MCP can read and write
Frequently asked questions.
Can we get the raw data, not only the structured dataset?
Yes. Both layers can be delivered: the structured dataset in your hierarchy, and the raw platform tables as detailed as each API offers, with the original structures preserved for your own queries.
Which warehouses can Odins deliver to?
BigQuery, Snowflake, Redshift or Databricks, with Power BI, Looker, Tableau or whatever your team already uses on top. Delivery is included in the service.
Can we see how the priors were set?
Yes. Our data scientists set the priors with you, from historical spend, the platforms' own conversion data and two or three structured interviews, and every prior and its posterior can be inspected afterwards. Customers do not set priors in the platform themselves. The technical view on priors.
Does the model estimate interactions between channels?
No. The model estimates how each channel affects sales on its own, and the optimizer treats channels as independent saturable resources. When heavy TV lifts search sales in the following weeks, that lift is credited to TV. Natural variation in the history is what separates channels that often run together. We say this plainly rather than claim otherwise.
How is the model validated?
Forecast first, compare afterwards. Each model stores its forecast before the actuals arrive and the verification view shows the two week by week. Every earlier model stays available, so you can check how the model from six months ago predicted the months since. The platform also alerts when predictions diverge from actuals. How we stay accurate.
We already run lift tests and use an attribution tool. How does that fit?
Both go into the model. Experiment results become priors on that channel's return, and platform data informs the priors after debiasing. The attribution tool stays where it is for optimizing inside each platform; the model takes the questions across platforms. Triangulation, in our words.
We built a Meridian model ourselves. Why not keep going?
Keep what it taught you; we learn from the earlier work, especially its assumptions about each channel. What usually stalls an in-house model is not the maths but the upkeep: the data collection, offline spend, the monthly retrain and turning output into a plan. That is the part Odins takes over. Google Meridian: build vs buy.
How granular can the model go?
As granular as the spend supports. One national weekly series typically carries around ten activity types reliably; a heavily used platform can be split into campaign types, a lightly used one is grouped with similar sources. Ad-group and ad-level claims from an MMM are more than the data can carry, and we do not make them.
What do you need from us to get started?
Access to the ad platforms through an onboarding link, an automated sales feed (daily is best), and an introduction to your media agency for TV, radio and other offline spend. Two or three clarification meetings while the model is built. The first model is typically live four to six weeks after kickoff.
Where is the data stored, and who owns it?
In Google Cloud's EU regions, with storage locations configured for EU residency. Everything you provide, and everything we collect and structure for you, remains your property throughout, and if you leave you get a final export in a standard format. The modeling data is aggregated and contains no personal data. Your data and security.
Is there an API?
The MCP is the API. It is not a read-only layer: an agent connected to it can forecast, create scenarios, and write and pull data, and the interpretation of each value is encoded in the server. Odins MCP.
Can we use Odins only for the data?
Yes. Some customers use Odins purely as managed marketing data infrastructure: every channel, offline included, structured into their own warehouse. The model can be added later on the same foundation.
See the data layers and the model on your own sources.
A 30-minute technical walkthrough: your sources, the hierarchy, the delivery to your warehouse, and what the model exposes.
Book a demo