The research behind orbt

What this is

This research aims to explain why growing consumer brands struggle to turn information into decisions, why the struggle gets worse the moment a brand crosses a border, why the existing categories of help fail for structural reasons rather than execution reasons, and what the evidence says a real answer has to do.

We published it for two reasons. First, most of what it contains is useful whether or not you ever talk to us. A founder can take the checklist in part four and use it to evaluate any tool, agency, or consultant, including orbt. Second, we think a company asking for trust should show its work. This is ours.

Every number in this report traces to a named source, listed at the end. Where a figure comes from a vendor’s own research, we say so. Where we could not confirm something to a primary source, we cut it.

The findings, in brief

1. Brands do not have a data problem. They have a decision problem. In a 2023 Oracle-commissioned survey of over 14,000 people across 17 countries, 72% said the sheer volume of data, and their lack of trust in it, had at some point stopped them from making any decision at all. 77% said the dashboards and charts they get do not always relate directly to the decisions they need to make.
2. More dashboards will not fix it. Business-intelligence adoption has sat at roughly a quarter of employees for about seven years (BARC and Eckerson Group). The tools exist, but people do not use them, for reasons that are psychological, not technical.
3. Trust in data is falling. 67% of organizations say they do not fully trust the data they use for decisions, up from 55% a year earlier (Precisely and Drexel University).
4. Cross-border operation multiplies the problem mechanically. A brand selling in three countries runs separate marketplace accounts, separate fee structures, and separate currencies, with no consolidated view of global profit. Exchange rates quietly distort which market looks profitable. Syndicated retail data is priced for companies that are already big.
5. The winners in cross-border CPG follow a repeatable pattern, and it is not spend. Samyang took Buldak from roughly $50M to $419M in US revenue in about two years by concentrating on one hero product. Hi-Chew’s US breakthrough came from switching distributors, not from marketing spend. Across the 160-plus deep research reports our engine has produced on consumer brands, the same three bottlenecks repeat: diffuse brand architecture, the wrong distributor structure, and earned media spread across too many channels.
6. Raw AI cannot be trusted with this job yet, and the industry knows it. On the Spider 2.0 benchmark, built from real enterprise data tasks, the best AI framework in the paper solved 21.3% of problems, versus 91.2% on the academic version of the same test. AI fails fluently: the wrong answer looks exactly like the right one. Any system that hands AI output straight to an operator without verification is selling risk.
7. The evidence converges on what a real answer must do: decide what matters for the operator instead of showing everything, recommend with visible reasoning while leaving the call to the human, show a source for every number, arrive in the operator’s existing workflow instead of waiting to be visited, remember every decision and outcome, and get paid in a way that aligns it with the brand’s result.

orbt is our attempt to build exactly that. The rest of this report is the evidence.

Part one: the decision problem

The consumer-brand founders we work with are not under-informed. They have Shopify analytics, Amazon Seller Central, Meta Ads Manager, TikTok analytics, a Google Analytics property, maybe a Triple Whale or Klue subscription, and an inbox full of newsletters. What they do not have, at nine on a Tuesday morning, is an answer to the only question that matters: what should we do next?

The scale of this gap is documented. Oracle, working with researcher Seth Stephens-Davidowitz, surveyed over 14,000 employees and business leaders across 17 countries in 2023. 72% admitted the sheer volume of data, combined with their lack of trust in it, had stopped them from making any decision at all. 85% of the business leaders had suffered what the study calls decision distress: regret, guilt, or second-guessing over a decision made in the past year. 77% said the dashboards and charts at their disposal do not always relate directly to the decisions they need to make. And 70% of the leaders, remarkably, said they would prefer a robot just made their decisions for them.

Hold that last pair of numbers together, because the whole product question lives between them. People are drowning enough to wish a machine would decide, and simultaneously unable to act on the tools they already have.

The tools are not underpowered. They are unused. A BARC and Eckerson Group survey of 214 data and analytics leaders found business-intelligence adoption stuck at roughly 25% of employees, essentially flat over the seven years they had tracked it. The growth that did happen came from insight pushed into tools people already use, not from more logins to analytics platforms.

Why do capable people ignore expensive tools full of relevant information? Behavioral economics has an unflattering answer. George Loewenstein at Carnegie Mellon and his colleagues documented the “ostrich effect”: people avoid information even when avoiding it degrades their decisions, because looking feels like a test they might fail. Investors check their portfolios less when markets fall. A founder whose ad costs are climbing opens the dashboard less at exactly the moment it matters most. Any product that assumes an anxious operator will voluntarily visit a wall of charts is built against human nature.

And when they do look, one error ends the relationship. Databox, which studies dashboard usage, puts it plainly: “a single stale number confirms a stakeholder’s suspicion and permanently removes that dashboard from their workflow.” Distrust is already the norm and growing. In a Precisely-commissioned survey with Drexel University’s LeBow College of Business, covering over 565 data professionals, 67% said they do not completely trust the data their organization uses for decision-making, up from 55% the year before.

So the baseline picture, before anyone crosses a border: the average growing brand has more information than any operator in history, uses about a quarter of it, trusts it less every year, and still ends the week asking what to do next.

Part two: what a border does to the problem

Everything above gets mechanically worse when a brand starts selling in a second country, and the brands we serve usually operate in three or more.

Start with the bookkeeping reality. A brand on Amazon in the US, the UK, and Japan runs three separate Seller Central accounts, one per region, with three fee structures and three currencies. There is no native consolidated view of global profit; the gap is well enough documented that a category of analytics vendors exists to patch it. Exchange rates make it worse than inconvenient: compare GBP revenue to USD revenue without normalizing and you can believe your most profitable market is one that actually loses money.

The measurement industry does not rescue smaller brands here. Syndicated retail data contracts from firms like Circana or NielsenIQ typically run five to six figures a year, which puts a direct license out of reach for most emerging brands. This creates a trap we see constantly: a brand must wait until it is big enough to afford to measure whether it is big enough. Meanwhile the channels multiply underneath it. A Skai and Stratably survey of 166 retail media advertisers found brands running an average of six retail media networks, expecting eleven by the end of 2026, each with its own login and its own version of the truth.

The cost side squeezes at the same time. Across Triple Whale’s aggregated data from the Shopify brands it serves, Meta CPMs rose 22.6% year over year in 2025 and Google cost-per-acquisition rose 17.3%. Paid acquisition keeps getting more expensive while the data needed to spend it well stays fragmented.

Against that backdrop, look at who actually wins cross-border, because the pattern is the most useful thing we have found. Our engine has now produced over 160 deep research reports on consumer brands, many of them Asian CPG companies entering Western markets and Western brands going the other way. Three bottlenecks repeat so reliably we treat them as defaults until proven otherwise:

Diffuse brand architecture. The instinct when entering a new market is to bring the whole portfolio. The winners do the opposite. Samyang Foods put its entire US push behind one product, Buldak, and grew US revenue from roughly $50M in early 2023 to $419M in 2025, expanding from about 15,000 to 30,000 retail doors. CJ did the same with Bibigo. Concentration wins; the diversified entrants stall on crowded shelves with no single thing to be known for.

Distributor structure as an invisible ceiling. Morinaga’s Hi-Chew spent years in the international aisle of US grocery stores. The breakthrough was not an ad campaign; it was switching from a Japanese distributor to an American candy distributor around 2012, which unlocked Walmart, Target, and Kroger. Brands routinely spend on marketing to fix what is actually a distribution-structure problem, because the distributor question never shows up on a dashboard.

Earned media spread thin. The winners pick one earned channel and saturate it. Hi-Chew took Major League Baseball dugouts. Kopiko took K-drama product placement. Lao Gan Ma took chef advocacy. The losers run thin campaigns on six channels and own none of them.

None of these three bottlenecks is visible in Shopify analytics or a media-monitoring feed. They are judgments formed from reading a market whole: filings, distributors, retail sets, social movement, category history. That kind of reading has traditionally been sold one way, which brings us to the consultants.

Part three: why the existing answers fail structurally

Four categories of help exist for a brand facing all this. Each fails for a structural reason, meaning a reason that better execution inside the category cannot fix. We built orbt only after convincing ourselves these failures were structural. Here is that case.

Consultants: right depth, wrong clock

A top-tier strategy engagement produces genuinely deep work. It also costs six to seven figures (industry estimates put a typical McKinsey, BCG, or Bain project at $500K to $1.25M and up; the firms keep fees confidential), takes months, and ends: the deliverable is a deck, and the deck is a snapshot. A market read in March is delivered in June to be executed against through December. In categories where a competitor’s TikTok can move shelf velocity in weeks, strategic information now has a half-life measured in days. Whatever a deck cost, being wrong for ninety days costs more. The engagement model, brilliant people billed by the project, cannot deliver continuously. That is not a flaw of any firm, but the economics of the format.

Dashboards and intelligence SaaS: the what-engine problem

Every analytics and competitive-intelligence product we studied, from Palantir at the enterprise summit down through Triple Whale, Similarweb, Klue, and Crayon, shares one shape: it is a place data goes to be looked at. We started calling them what-engines. They compete on how much data they gather and how well they display it, and they all leave the user with the same homework: you figure out what it means, you figure out what to do, you find the time.

The record of that shape is the record from part one: a 25% adoption ceiling, two-thirds of organizations distrusting the numbers, and the most repeated complaint across every review corpus we mined being some version of “buried under mountains of reports, unable to discern what matters.” Tools that need a dedicated in-house analyst to stay useful drift to shelfware in a quarter when that person is busy, which in a twenty-person brand is always. The category’s honest job description is “we show you what happened.” The job the operator is hiring for is “tell me what to do about it.” Those are different products.

In-house hires: the right answer most brands cannot buy

A seasoned cross-border operator as a full-time hire genuinely solves this, and for later-stage companies it is often the right call. But the fully-loaded cost starts around $250K a year before you learn whether the fit is right, ramp takes a quarter, and one person covers one market’s worth of hours. The brands that need the judgment most, the ones between roughly $2M and $50M, are priced out at exactly the stage where a wrong market entry can end the company.

Raw AI: fluent, confident, and wrong too often

The obvious 2026 answer is to point a large language model at the problem. We build with these models daily, and the evidence says: not without heavy scaffolding. On Spider 2.0, a benchmark built from 632 real enterprise data tasks rather than academic exercises, the best-performing AI framework in the paper solved 21.3%. The same framework cleared 91.2% on the older academic version of the test. The gap between the demo and the enterprise is that wide, and worse, the failures are fluent: plausible SQL that means the wrong thing, an answer to a different, easier question presented as if it were yours. Microsoft’s own documentation for Power BI Copilot warns, as of this writing, that it “can produce inaccurate or low-quality outputs, including incorrect answers to data questions.”

The market has already run the experiment of skipping the verification step. The AI sales-agent startup 11x sold autonomous execution before earning basic accuracy. TechCrunch reported in March 2025 that ZoomInfo found the product performed significantly worse than its own human SDRs during a trial, that companies including ZoomInfo said their logos were displayed as customers without authorization, and that a former employee described churn around 70 to 80%. Meanwhile a 2026 BCG survey of 300 CMOs found only 8% run campaigns where multiple AI agents act autonomously, which strikes us as sensible caution. Researchers Berkeley Dietvorst, Joseph Simmons, and Cade Massey documented years ago that people abandon an algorithm faster than a human after seeing the same mistake. One hallucinated number in a Monday briefing does not lose that briefing a point of credibility. It ends the briefing’s career.

The failure the four categories share

Notice what none of the four sells: accountability for the outcome. The consultant is paid whether the deck works or not. The dashboard collects its subscription whether anyone logs in. In The Trusted Advisor, David Maister, Charles Green, and Robert Galford formalized why this matters: trustworthiness rises with credibility, reliability, and intimacy, and is divided by self-orientation, the degree to which the advisor benefits regardless of the client’s result. A subscription product can maximize the numerator forever and can never touch the denominator. Its revenue does not know whether your brand grew. That, more than any feature gap, is why operators heed a trusted advisor and ignore a tool showing the same information.

Part four: what the evidence says a real answer must do

Strip out everything vendor-specific and the research above converges on six requirements. We offer them as a checklist for evaluating anything in this category, including us.

1. It decides what matters, and says so. The operator’s scarce resource is attention. Oracle’s 77% figure is a design instruction: the product must do the prioritizing, surfacing the few things that changed and matter, ranked, with everything else a level deeper. A wall of tiles is an abdication.

2. It recommends, with the reasoning visible, and leaves the call to the human. The demand for delegation is real (70% of leaders would prefer the robot decide) and full delegation is punished (the 11x lesson; algorithm aversion). Dietvorst, Simmons, and Massey found the resolution in a 2018 Management Science paper: people will happily use an imperfect algorithm when they can even slightly modify its output. The stable form is: here is what we would do, here is why, you make the call.

3. Every number carries its receipt. Source, date, and a path back to the origin, on the surface, not in an appendix. Because AI fails fluently, verification cannot be optional, and because one wrong number is fatal, the system must be able to say “we do not know” when the data will not hold an answer.

4. It arrives; it does not wait to be visited. The 25% adoption ceiling is a verdict on pull-based tools, and the ostrich effect explains it. The answer has to show up where the team already works, on a rhythm, and most mornings it should say: you are fine, here is the one thing worth your attention. Lowering the operator’s anxiety is the mechanism that keeps them looking.

5. It remembers. Generic advice is a consequence of amnesia. A system that retains every baseline, every past decision, and every outcome gets more specific to the brand every week, the way a good hire does and a dashboard never has. Benn Stancil, who co-founded the analytics company Mode, states the underlying law: “You can’t tell people something surprising without knowing what is expected; you can’t tell them something interesting unless you know what they think is boring.”

6. Its incentives point at the client’s outcome. The trust equation’s denominator. However structured, the entity giving the recommendation should be paid more when the brand grows and less when it does not. This is the one requirement no feature can retrofit, because it is a business model, and the market is moving toward it: Bloomberg Intelligence estimates outcome-based pricing could grow from about 10% to about 60% of software pricing models over the next decade.

Part five: how orbt is built against that checklist

orbt is an AI growth strategist for consumer brands, and each design decision traces to a line of the research above.

Signal reads the brand’s market daily across roughly two dozen live sources: SEC filings, earnings transcripts, patents, trade data, six social platforms, competitor content, reviews, and search. Its job is judgment, not collection: filtering the day’s noise down to what is relevant to this brand’s active questions (requirement 1).

Strategy is a panel of five specialist analyst agents (financial, leadership and governance, brand and marketing, operations and supply chain, competitive moat), each producing findings that a separate fact-check agent then verifies against the evidence and marks verified, flagged, or insufficient. What survives becomes a recommendation with its triggering data, supporting evidence, expected outcome, and the honest downside if it is wrong (requirements 2 and 3). Before anything reaches a client, a human strategist reviews it. We built the verification layer first because the Spider 2.0 numbers say fluent failure is the default state of this technology, not an edge case.

Memory is a knowledge graph we call the Customer Brain, where every finding, decision, and outcome persists. A fact established once stays established; a hypothesis rejected in March does not get re-proposed in June; recommendations sharpen as the record deepens (requirement 5). As of mid-2026 it holds over 1,700 verified findings across the brands we study, alongside the 160-plus full research reports the engine has produced so far.

Delivery is push, not pull: a few clear moves a week, in Slack, where the team already works (requirement 4). And pricing is outcome-shaped: our engagements are structured around the revenue we help create, so we are paid more when the brand grows (requirement 6). We do not claim this makes us right more often. It makes us accountable when we are wrong, which the research says matters more.

One field lesson shaped the system as much as any study. Early client feedback taught us that even correct analysis fails when it arrives as noise: a client’s creative team asked us to strip detailed scene descriptions out of briefs because the team, in their words, “loses focus when reading long blocks of text.” The three-to-five-moves-a-week format exists because operators told us anything more gets ignored, which is the Oracle finding replayed in miniature, at one brand, in one Slack channel.

Part six: what we do not know yet

A research report that only supports its author is marketing. Three honest edges:

Attribution is genuinely hard. When a brand grows after acting on a recommendation, separating our contribution from everything else moving is an unsolved problem industry-wide. Outcome-shaped pricing forces us to negotiate this honestly per engagement rather than hide behind a platform metric, but it does not make the measurement problem disappear.

Virality is not velocity. Social signal is a leading indicator we weight heavily, and it can mislead: a trend can be loud and commercially empty. Some of our own desk-research conclusions have been reversed by ground-truth checks, which is exactly why the verification layer exists, and why we assume any unverified insight, including ours, might not survive contact with the shelf.

The half-life of this report. The numbers here are point-in-time 2026. The structural arguments (the decision problem, the what-engine shape, the trust equation) have held for decades and we expect them to keep holding. The market figures will drift, and we would rather you check them than quote them blind. Every source is below.