Every number on your dashboard cites its source. Even the ones that didn't move.
You've been sold "data-driven marketing" before. What you
got was a dashboard with numbers nobody could explain,
and no clear path from those numbers to a decision you
could act on next Monday morning.
Here, every measurement has a plain-English name,
the method behind it, and the tier where it unlocks.
No hand-waving — published methods, explained simply. So
next time someone asks "why?", you've got an answer.
29
published methods behind every number
250
measurements across 5 business-profile reports
50
named specialists across 6 teams
0
made-up numbers: every claim sourced
What we actually measure.
Eight of the load-bearing measurements behind every
engagement, in plain English. The full 250-measurement
catalogue across five reports sits on
/profiling.
How much it costs to win a new customer
If someone asked tomorrow what each new customer costs you to win, you'd be guessing. The accountant could tell you what they earn. You couldn't tell them what they cost.
What it tells you: Whether you can afford to grow your marketing.
How we quantify it: Customer Acquisition Cost (CAC), calculated per channel from spend + organic effort ÷ customers won.
Why it matters: Without it, you can't know if spending more leads to making more.
How it compounds: Spark gives you CAC by channel. Bloom predicts it forward. Apex breaks it apart causally with multi-touch attribution.
Source
Standard direct-response economics. Refined via Google + Meta attribution methodologies (2020-2024).
You land on:Pay for outcomes, not hours.
What a customer is worth over their lifetime
You sense some customers come back. You couldn't say which ones, or for how long, or whether you can afford to chase the new ones at the rate you've been paying for them.
What it tells you: How much you can afford to spend acquiring one.
How we quantify it: Customer Lifetime Value (LTV), via the BG/NBD + Gamma-Gamma model from the Wharton Customer Analytics literature.
Why it matters: It's the ceiling of your CAC. Go above it and you lose money quietly.
How it compounds: Bloom predicts LTV per segment. Premium models per-customer LTV with causal forests. Ultra Studio uses it as input to dynamic pricing.
The cash leaves on Monday. The customer's first payment lands months later. You haven't drawn the line on a page to see whether those two numbers actually meet.
What it tells you: How many months it takes a new customer to cover their acquisition cost.
How we quantify it: Payback period, CAC ÷ monthly contribution margin per customer.
Why it matters: Long payback = cashflow risk + slow compounding. Short payback = headroom to reinvest.
How it compounds: Spark shows you payback per cohort. Apex shows you payback per causal channel contribution.
Source
Standard SaaS + DTC unit-economics convention; popularised by Reforge essays + a16z growth benchmarks.
You land on:Pay for outcomes, not hours.
Where the next pound should go
Three channels claim credit for the same customer. The dashboards don't agree with each other. You end up moving budget by whichever report you looked at last.
What it tells you: Which channel produces the highest *incremental* return, not just the highest reported ROAS.
How we quantify it: Marketing Mix Modelling. We integrate Meta's Robyn and Google's LightweightMMM, peer-reviewed implementations of decades of Hanssens-Parsons-Schultz market-response theory.
Why it matters: Without it, you split budget by gut. With it, you split by evidence, and the evidence often disagrees with the dashboard.
How it compounds: Thrive runs lightweight MMM monthly. Apex runs full Bayesian MMM with saturation curves + geo-experiments. Ultra Studio adds hierarchical pooling across the agency portfolio.
Source
Hanssens, Parsons, Schultz, Market Response Models (2nd ed., 2003). Robyn (Meta, open source). LightweightMMM (Google, open source).
You land on:Know what works. Do more of it.
Whether your creative is actually working
You ran an A/B test for six weeks. By the time the answer landed, the season had changed. The lift went to whoever guessed faster, not whoever guessed right.
What it tells you: Which ads + posts + emails drive conversion vs. which are just running.
How we quantify it: Multi-armed bandits, Thompson sampling, that allocate spend adaptively toward winners while still exploring.
Why it matters: Most agencies A/B test for weeks then switch. Bandits give you the lift WHILE testing, typically 20-40% more efficient than fixed splits.
How it compounds: Thrive runs bandits on creative rotation. Apex adds contextual bandits per audience. Ultra Studio closes the loop, winning creative atoms feed back into new variant briefs.
Source
White, Bandit Algorithms for Website Optimization (O'Reilly, 2012). Chapelle & Li, Thompson sampling for contextual bandits (2011).
You land on:Know what works. Do more of it.
Whether your customers are about to leave
A customer disappears. You only know weeks later, when the next renewal doesn't land. The first thing you ever hear about their dissatisfaction is the silence.
What it tells you: Which active customers are quietly about to churn, before they tell you.
How we quantify it: Churn models, logistic regression or gradient-boosted machines, on recency / frequency / monetary + engagement features.
Why it matters: Win-back costs roughly a fifth of new-acquisition. Identifying churners early lets retention catch them while it's still cheap.
How it compounds: Bloom flags churn risk per segment. Premium scores individual customers continuously. Ultra Studio drives the whole lifecycle as a Markov decision process.
Source
Linoff & Berry, Data Mining Techniques (3rd ed., 2011). Sutton & Barto on MDPs (RL: An Introduction, 2nd ed., 2018).
You land on:Customers who fit your business.
What customers say (in their own words)
Your headlines were written for the customer you imagined. The reviews speak a completely different language. Both versions are live. Both can't be right.
What it tells you: The exact language to put in your ads, your landing pages, your follow-up emails.
How we quantify it: Voice-of-customer mining, topic modelling + sentence-transformer embeddings, across reviews, Reddit threads, niche forums, and interview transcripts.
Why it matters: Customer voice typically converts 2-3× better than marketing voice. The good language already exists; we surface it.
How it compounds: Bloom mines public reviews quarterly. Thrive adds forums + interviews. Ultra Studio runs continuous drift detection on emerging themes.
Source
Robinson & Silge, Text Mining (2017). Reimers, sentence-transformers (2019-).
You land on:A brand that gets remembered.
Whether your brand stays consistent at scale
Three months of work look like the work of three different brands. By the time you notice the drift, you've spent the budget reinforcing the wrong one.
What it tells you: Whether your ads, posts, and assets carry the same brand identity, or drift quietly into a different brand.
How we quantify it: A typed 10-dimension aesthetic ontology, about 200 named atoms across visual style, lighting, colour, typography, composition, motion, sound, voice, narrative, and archetype, annotated on every asset.
Why it matters: Drift kills brand equity slowly + invisibly. By the time you notice, you've spent six months reinforcing the wrong brand.
How it compounds: Good audits drift quarterly. Apex flags drift in real-time. Ultra Studio generates new variants from your authentic strategy + runs causal tests on the aesthetic axes.
Source
Internal taxonomy (ADR-0248). Influenced by Ricci et al., Recommender Systems Handbook (3rd ed., 2022) for the atom-vector approach.
You land on:A brand that gets remembered.
These eight are illustrative, the full Studio Crew engagement
instruments approximately 250 typed metrics across business,
customer, market, marketing, and brand axes.
See all 250 →
How depth scales with tier.
Every tier ships a true subset of the next. Spark is honest
about depth, descriptive numbers, reported weekly. Ultra
Studio is the full surface, adaptive systems with HITL
approval, refreshed continuously, one named agent per metric.
1
Basic
Spark £500/mo
Foundational numbers, CAC, ROAS, LTV (historical), payback. The numbers a competent in-house analyst would deliver, reported monthly.
2
Great
Bloom £1,000/mo
Prediction. LTV moves from historical to projected. Churn risk gets scored. Attribution moves beyond last-click.
Causal evidence. Full MMM with saturation curves. Quarterly geo-experiments. Uplift modelling. Per-customer treatment effects. Synthetic controls for one-off moments.
5
World Class
Ultra Studio £10,000/mo
Adaptation plus auditable discipline. RL-bidding within policy bounds. Lifecycle journeys as Markov decision processes. Closed-loop creative, winning aesthetic atoms feed back into new briefs. Quarterly fairness audits. Quarterly board-ready business reviews. A named regulator-pack auditor per jurisdiction the engagement touches. Thirty-day post-engagement support window after any exit. Portfolio-level inference compounds across all World Class clients with strict tenant isolation.
How we cite research.
Most agencies cite research as a vibes signal. We cite it
because we use it, and because clients verify it.
If we get the science wrong in the sales deck, what does
that say about the rest of our work?
Two examples of the discipline we hold ourselves to.
Aggarwal et al., GEO: Generative Engine Optimization (KDD 2024)
The paper is widely misquoted as "+41% / +38% / +34% / +28%
per-method uplift." That's not what it says. What it
actually says: up to ~40% citation-rate uplift,
domain-dependent. Best result comes from a
combination, Fluency optimisation + Statistics
addition together: +5.5% over any single method. We cite
it that way. arXiv 2311.09735.
Liu et al., Lost in the Middle (2023)
The empirical phenomenon (U-shaped recall in long-context
prompts) is real. The pseudo-mathematics often layered on
top, "factorial dead zones", "primacy/recency deltas",
isn't. We state the phenomenon, not the confabulated
mechanism. arXiv 2307.03172.
TCAV (Kim et al. 2018)
TCAV needs white-box activation access. Closed frontier
models (ChatGPT / Claude / Gemini) don't expose those.
We do not sell "latent-space sentiment probing" against
those models. Some agencies do; the technique is
structurally incompatible with closed APIs.
Every metric on every report we hand you carries its
source, when it was fetched, and a confidence band. If
we cite a paper, we cite it correctly. If a number can't
be sourced, we say so.
What this isn't.
It isn't a stack of dashboards. It isn't a list of vendors.
It isn't a roomful of strategy decks. It's a typed,
versioned, reproducible method library wrapped in an
accountable team.
Every claim cites a method + a data source + a confidence
band. If we can't measure something defensibly, we say so.
That promise is the brand.