You know your shipments. You do not know who bought, whether they switched to you or back, or whether the media you funded changed anything — at every chain outside the two or three big enough to give you household data. We are looking for a small number of brands to build that measurement with us, at one retailer each, starting from a relationship you already have.
The deal, plainly
No one on your side has to understand the technology, and no one has to build anything.
How that is even possible
Keys are released only to code the retailer approved. Different code, even by one byte, and the data stays encrypted. This is not a promise about who will look — there is no mode in which anyone looks.
What you would learn
Did the media actually sell anything?
Households that saw your campaign against a matched control that did not: incremental sales with a confidence interval, and which channel earned its money.
See it live →Where did my growth come from?
New category buyers, switching from named competitors, more trips, bigger baskets — and who took volume back from you, by name and by dollar.
See it live →Who should I be reaching?
Competitor loyalists, lapsed buyers, category heavies who have never bought you — sized before you spend, and activatable in the retailer's own media.
See it live →Where is demand I am not serving?
Stores with category demand and no distribution of your SKU, pack sizes that do not move, cannibalisation inside your own range.
See it live →The one that pays for itself
This is the measurement most brands are missing at these retailers, and the one with a budget line attached. Everything below is a working screen on synthetic data — click into it.
The campaign produced $33.9K you would not have had
420,000 exposed households against 420,000 matched controls, at one 212-store regional chain.
Sales in the eight weeks. The gap is the result; the screen carries the confidence interval, the channel breakdown and an honest note that one channel did not clear break-even.
Where the quarter's change actually came from, and who took volume back.
Open →Targeting who will buy is not the same as targeting who you can change. One group of your own buyers responds badly, and conventional targeting walks straight into it once the budget is large enough.
Open →How an audience could be activated inside an AI assistant with nothing disclosed to anyone. We publish the design because nothing here is standardised yet. Marked as unbuilt on the screen itself.
Open →Being straight with you
You are being asked to join early, so you should know exactly which parts are proven and which are ahead of us. This table is the honest version.
| Capability | Status | What that means for a pilot |
|---|---|---|
| Confidential environment and attestation | In production | The underlying technology runs in regulated industries today. Not a research demonstration and not ours to invent. |
| The four analyses shown here | Built | Methods and outputs are defined and run. They have been exercised on synthetic data at full scale, not yet on your retailer's file. |
| Verifiable run manifest | Built | Sample package and procedure are published and checkable. Production signing is in place; the published sample is unsigned and says so. |
| Activation into retailer media | Partner-dependent | Built, but each retailer's media stack needs its own connector. Which one we build first is decided by the first partners. |
| Offsite activation (DSP, walled gardens) | In build | Architecture is settled and documented; the connectors are being built now. Expect this later in a first pilot, not at the start. |
| Model-selected audiences | In build | Working on synthetic data. Real predictive quality cannot be claimed until it has seen a real loyalty file, and we will not claim it before then. |
| Uplift (persuadable) targeting | Needs your first campaign | The method is built and demonstrated here, but uplift cannot be estimated without a campaign that ran with a randomised holdout. Your first quarter creates it; your second uses it. |
| Predicate activation in AI assistants | Concept | A published design, not a feature. No assistant can run it today, and we say so on the screen. Not part of any pilot. |
| Federated network across retailers | Later | Needs several retailers before it means anything. Described on this site, not sold, and not part of a first pilot. |
Why your retailer says yes
Suppliers who can measure what happens in their stores shift budget toward their stores. That shows up in the retailer's retail media revenue rather than in an invoice they have to chase.
And the alternative on offer to a mid-sized chain today is an intermediary that sees every retailer at once — including what their customers buy from their competitors — and competes with them for the shopper. This is the other route.
They stay in control the whole way. Their data never leaves their environment and is never copied. They approve each analysis before it can run, set the thresholds themselves, see every result you see, and can revoke access at any point — with no copy anywhere to retrieve, because none was made.
A page written for them, ready to forward → — you do not have to explain any of this yourself.
We show the difference rather than hiding it behind one identical button. Hiding it is how you lose a retailer — usually at the contract stage, after everyone has spent three months.
Founding partner
Honestly: it depends on the two of you, not on us. The parts we control — environment, mapping, analysis — are weeks. The parts that set the pace are your retailer's legal review and how quickly their data team can produce an export. We have seen both move in two weeks and we have seen both take two months, and no vendor who promises you a fixed date controls either.
What we will do is tell you at the start which step is likely to be slow for your particular retailer, and we will say so before you spend relationship capital on the introduction.
Context, if you want it
Every chain below the leaders is individually too small to justify a data integration: the signal is modest, the integration costs the same as a large one. So brands integrate with the few that clear the bar and write off the rest — and the rest is not one competitor, it is hundreds of chains.
You do not need the whole board. You need the chains where your volume actually sits — and each one is individually too small to justify its own integration, which is exactly why nobody has built it.
Shares are of U.S. food and beverage spend from Numerator's Q2 2026 market share report. Kroger's 8.7% assumes its pending Giant Eagle acquisition clears. The residual is our own calculation — the remainder after the named shares — and is shown as a picture of concentration, not as an addressable-market claim.
Each side loads its own data separately. Matching and computation happen inside a hardware-sealed environment whose contents are unreadable to the retailer, to you, to the cloud provider and to us. Decryption keys are released only after the processor proves the running code is the exact code the retailer approved. Only aggregates come out, with any cell below 100 households suppressed.
Activation sends identifiers, never records, and the enclave opens the connection itself using the retailer's own credentials, to destinations on an allowlist inside the policy they signed. Every run produces a manifest your security team can verify against the chip vendor's attestation service rather than against our word.
The full mechanism is on the retailer page, and a real verification package is published here.
That is all a first conversation needs. Thirty minutes, no deck, no preparation on your side — and if the sample at that chain will not support the question you care about, we will tell you then rather than sell you a pilot.
One of us reads it and replies personally, usually with a question about the retailer and the category before we propose anything.
Lift Mosaic. The example screens on this site run on synthetic data with the structure and distributions of a regional grocery loyalty file; the retailer, brands and figures in them are invented. The methods are not: the matched-control design, the decomposition, the suppression rule, the disclosure allowance and the manifest format are what we run. What is built and what is still ahead of us is set out in the status table above.