obsessed · outcomes
internal operating site
the operating site · live · june 2026
We are agnostic to where data comes from. A vendor panel, a geo experiment, a Northbeam export, GA4, a CSV from a DSP: it all feeds one computation layer that we own. We triangulate because we are integrated media strategists first, and we lean on agentic, modular tooling to land sophisticated reads in hours and days, not the weeks and quarters this category accepts. Vendors fill slots. None of them is the program.
outcome first
name the result that matters per client: sale, site visit, branded search, awareness that ladders down. then prove lift against it.
triangulate
multiple methods, forced to argue. when they agree, that is the read. when they disagree, that is the finding.
modular
every vendor is a swappable slot. swap the engine, nothing upstream changes. the analysis layer is ours.
days, not quarters
agentic ingestion + synthesis per source. course-correct while the budget is still live, not at the annual readout.
01 · the operating model
Every engagement runs the same skeleton: three lenses hardened into targets, reconciled against a geo-holdout spine. Expand each lens for the method and the Stable layer that differentiates it.
method
survey lift inside exposed geos, never a standalone national panel. two-stage model: brand media moves brand metrics, brand metrics feed revenue as a regressor.
continuity
synthetic brand index from search, social, and LLM-presence signals bridges between survey waves. directional only, anchored to first-party. the read never goes dark.
horizon
quarterly wave reads with lag tracked. unaided awareness is a secondary trend line, never the KPI. benchmark reality: UBA moves +0.6 pts on average vs +4.5 aided and +10 ad awareness.
the stable layer · halo visibility cut and re-attributed halo make the brand-to-revenue link explicit in the model rather than a hand-wave, which is the cleanest answer to MMM undercrediting long-tail brand.
method
correlation plus ad-stock diagnostics joined to GA4 funnel and search signals, with predictive on-site behavior reading intent before conversions mature.
validation
geo-holdout on the driving channel, respecting the 10 to 14 day ad-stock lag and the 15 to 25 day decision window. exposed geos should out-signal controls or the brand read is suspect.
horizon
leading indicator at 3 to 4 weeks. explicitly labeled directional, never causal, never a headline metric. it pressure-tests the other two lenses.
the stable layer · this is where CTV's downstream value becomes visible: branded search uplift two weeks post-flight inside the exposed geo, not a last-click line item.
method
MMM plus MTA with incrementality tests injected as priors to calibrate channel coefficients. transactions as the KPI, never blended revenue. holdout sequencing TOF to mid to bottom.
action layer
marginal CAC and marginal ROAS by channel feeding budget-neutral weekly reallocation. significance thresholds before any lift claim ships.
horizon
15 to 25 day read on a 10 to 14 day ad-stock half-life. quarterly incrementality-coefficient refreshes keep the annual model honest between fits.
the stable layer · TOF out of the denominator, dual-MER, marginal CAC by channel. proven live on the Northbeam loop and portable to any engine in the slot.
Exposed markets against matched controls. It is the only method that cleanly captures OOH, radio, and cinema, which are inherently geographic, and it unifies offline with digital in one read. Parent-brand halo sits in both cells and cancels in the delta. Surveys run inside the exposed geos to add attitudinal depth. Footfall corroborates. Everything reconciles back to the holdout, and the cross-cutting fix for every client is the same: quarterly incrementality refreshes layered on the annual MMM so nobody flies blind between models.
02 · productization
Depth is set by spend and by which outcome the client actually cares about. We buy the lightest combination that proves outcome lift, then add one corroborator. Never more than the outcome requires.
03 · the stack living · team-editable
The engine slot can be held by any attribution, MMM, or incrementality vendor; our conventions sit on top of whichever is in place. Statuses are live as of June 10. Expand a row for the full read and team notes.
The proven loop: MMM plus MTA with incrementality as priors. Source of the decision-window, ad-stock, and transactions-not-blended rules. Good, known quantity at a known price. Daily pacing read that the quarterly geo refreshes recalibrate.
Great. Managed incrementality and attribution. More expensive than Northbeam and probably better. Not modular: you buy their methodology, not components. Price it as the managed engine option for clients who want a brand-name methodology.
Good. Credible newer entrant for the same engine slot. AI-native pipeline, leans ecom.
May be an option. Josh joins the Stella demo; pressure-test against Northbeam, Measured, and WorkMagic before it earns a shortlist spot.
Great. Daily Bayesian MMM with 80 to 90 percent week stability, geolift experiments injected as the causal anchor. Geolift $1k/seat with unlimited studies, full MMM ~$150k, operator $20k plus $10k per business, roughly 5-week onboarding. Call with Drew Seman week of June 15: standard lift, deeper outcomes, the integrated holistic read, and confirming the LA + Dallas two-market design.
Good. Gold-standard geo experiments for the LA/Dallas design. For CTV and video they recommend 4 to 6 week minimum flights plus a 2 to 4 week post-treatment window to capture delayed lift. Call with Andrew R week of June 8. Run in parallel with Recast and compare.
Confirmed: W25-54 targeting, trust as a measurable KPI, deterministic tracking for CTV, linear, streaming audio, and podcasts, opportunity-to-see for social, YouTube, and newsletters. Floor of 7 to 10M impressions per readable cut, 1 to 2 day tag setup. Pricing contingent on impression volumes. Contact Christina Holstein.
Scalable representative panel with custom geo cells. Confirmed it can include low-impression channels like Uber (1.95M) and Lyft (750K) in overall campaign measurement, grouped given the geo targeting. Active exchange as of June 8. Contact Al Gaviria.
Strong where exposure is logged: CTV, social, audio. Weak for OOH and radio, which are opportunity-to-see only. Confirmed LA + Dallas as best-effort for small geos. Panel skews younger, so a W35-54 quality cut is required before any Midi award. Never the OOH spine. Contact Darin Lentzner.
Ties OOH exposure to a real outcome. One piece of the pie, a corroborator inside the exposed geos. Placer.ai is the heavier alternative or cross-check.
The partners already running media will often run lift studies at no cost. Factor in where credible, with a skeptical eye on self-marking. Also on the explore list: Upwave for an always-on digital brand read, Adelaide for attention quality feeding the interest lens, INCRMNTAL and Lifesight for geo alternatives at scale, Kantar or Nielsen held for when a client demands the brand-name study.
the other path
Everything above is the managed path: you buy a methodology and live inside it. The modular universe is a different angle entirely. Newton runs agentic MMM, MTA, and geolift in one, built on PyMC and Meridian, transparent auditable code in a client-owned AWS container, output-credit pricing. Config and code based, composable, agentic: a candidate hub that ingests every source and runs the triangulation itself.
This is the path that converges with our own engine. Buy the cheap causal anchors (Recast geolift), build the moat: the triangulation hub, the three-lens scorecard, the synthetic interest and brand layer, and the halo-aware re-attribution. No vendor sells that combination. Near term the dashboard is the centralization point and the vision, not production measurement infrastructure: it needs warehouse-backed compute before real client data runs on it, and we do not over-claim it. Build-first, Newton-fallback.
notes on the stack
04 · live engagements
Danny's briefs set the table. The value is the second layer: the design risks and reframes we bring that are not in any vendor's RFI response.
Mid-July flight. We own the geo-native half: OOH, radio, NCM cinema (3.2M), United in-flight (6.4M), TeamSnap (8.6M), Nextdoor (7.7M), Life360 (17.2M), Uber (1.95M), Lyft (750K). Amazon's in-house team owns CTV, Meta, YouTube, podcasts, and DSP. Their stated priority is unaided awareness, and Amazon.com equity bleeds into every absolute number.
the design
LA and Dallas run exposed against synthetic matched controls, with the panel survey fielded inside both cells and the baseline wave completed before first flight. The parent-brand halo hits both cells equally and cancels in the delta, so the decision metric is the exposed-vs-control delta on aided, Autos-specific awareness and consideration. Geo-holdout is the spine; the panel reads the digital slice inside it; Reveal Mobile corroborates OOH; partner first-party studies fold in skeptically. One delta, decomposed across both halves of the plan.
what we add beyond the brief
The live risk is Amazon's own flighting, not the halo
If their regional DSP and CTV weight concentrates in LA and Dallas, controls understate the counterfactual and we over-credit our channels. We need their DMA-level flighting calendar before locking controls. Nobody else is raising this.
Two markets is a small-N design
LA and Dallas are idiosyncratic DMAs. One-to-one matching will not hold; controls should be a synthetic weighted blend of non-exposed markets fit to pre-period trend.
Readability has a floor
Panels need roughly 7 to 10M impressions per cut. The portfolio clears it in aggregate; Lyft, Uber, and cinema do not individually. Group ride-share and cinema into composite cuts and set the expectation now, not at the readout.
The UBA ask is stakeholder management
Keep unaided on the scorecard as a trend line; decide on the aided delta. Cross-study benchmarks: UBA +0.6 pts average vs +4.5 aided, +10.0 ad awareness, +10.9 familiarity. Chasing absolute UBA buys a flat line and an argument in November.
August through December. We run CTV, linear TV, streaming audio, podcasts plus TV and podcast integrations, Reddit, and YouTube. The full plan also carries paid social, newsletters, Kindle, and partnerships outside our scope. Danny's asks to the vendors: trust as a metric, and isolating our channels from the full plan.
the design
Isolation is solved structurally, not with attribution accounting. Our channels are geo-flexible: run them as a geo-weighted heavy-up while Midi's always-on social stays national, a constant background present in both cells. The exposed-vs-control delta then reads our channels specifically without pausing anything. Trust enters the survey as a fixed two-to-three question battery, identical wording every wave: credibility, safety, expert care. Reported only as the exposed-vs-control delta.
what we add beyond the brief
October is a structural confound
Menopause awareness month and its PR cycle land mid-campaign. The geo design absorbs it since both cells get the same October, but the design doc must call it so nobody celebrates or panics at the spike.
Trust moves slower than awareness
In considered healthcare a 4-week trust pop is not a result. The honest read anchors at wave two and we say so upfront rather than overpromise the first readout.
Panel quality is the award fulcrum
Toluna confirmed W35-54 targeting and trust as a KPI. DISQO skews younger. The decision is panel composition verified with a quality cut, not feature checklists.
Sensitive-category screening changes the math
Menopause is a health condition; panels apply screening that shrinks eligible sample and moves cost. Surface it on the vendor calls now. Podcasts leak too: endemic placements outside the plan need UTM discipline or the lift becomes unattributable.
notes on the briefs
05 · the next 30 days
Workstream A locks the program. Workstream B lands Amazon Autos and Midi now. Mid-July flight math makes vendor award and tag placement a June decision.
in motion
owed
notes on the plan
internal operating site · living version · nothing ships externally without josh's written go
obsessed