Most founders cannot answer this honestly, and the dashboard is the reason. The default report credits the effort that stood closest to the sale, which is usually the demand you already had, not the work that created new demand. So the busiest channel looks like the best one, and the founder doubles down on the thing that was merely nearby.

The cleanest proof is a famous experiment. When eBay ran a controlled test on its own paid search, the causal returns came in a fraction of what the attribution models claimed, and brand-keyword ads showed no measurable short-term benefit at all. The ads were getting credit for buyers who were already coming. The dashboard saw a win. The holdout saw nothing. That gap, between what looked busy and what caused the lift, is the whole subject of this piece.

When researchers compared standard attribution against fifteen large advertising experiments, the attribution methods often failed to recover the real causal effect, even after controlling for everything they could measure.

Gordon, Zettelmeyer, Bhargava and Chapsky, Marketing Science, 2019

A vanity metric goes up whether or not your work caused it. Real lift is the difference your work actually made, measured against what would have happened without it. The only reliable way to tell them apart is comparison: run the effort for one group and not for another, then look at the gap. Everything else is correlation wearing a confident face.

This is why a rising number is not proof. Clicks, impressions, attributed conversions, and last-click revenue all climb when demand climbs, and demand climbs for reasons that have nothing to do with the campaign. The honest standard has a name, incrementality, and it isolates true lift by comparing a test group against a holdout. The pressure to adopt it is real: in a 2026 survey, three out of four marketers said their measurement was not delivering the accuracy and trust they needed. The number on the dashboard is not the lift. The gap between the test and the holdout is.

Under the default tools, all three claim it. Last-click hands the sale to whatever the buyer touched last, and each platform’s own dashboard counts the same conversion as its own, so the three reports add up to more than the sales you actually made. Credit gets multiplied, which feels generous and tells you nothing.

The way out is not a smarter model that splits the credit more cleverly. It is a test that removes one thing and watches what happens. The biggest spenders learned this by accident. When Procter and Gamble cut roughly two hundred million dollars of digital ad spend it judged ineffective, its reach went up, not down. The spend had been busy, and the busy-ness had been mistaken for impact. Turning it off was the only honest test, and the result was nothing lost.

Hold one slice back. Pick a portion of the audience, the territory, or the spend, run everything else exactly as normal, and deliberately leave that slice untouched. After a set period, compare the held-back slice against the rest. The difference is the honest measure of what your effort caused. That is a holdout, and it is the cheapest honesty available to a founder.

Don’t credit the busy thing. Credit the thing that caused the lift. The cheapest honesty is a holdout.

The objection is always “I am too small for that.” It is no longer true. The minimum budget for an incrementality test has fallen from around a hundred thousand dollars to roughly five thousand as the methods got cheaper, which puts the honest version within reach of the mid-market for the first time. And a holdout on your own audience costs nothing but the discipline to wait and not peek. The hard part was never the budget. It was being willing to find out you were wrong.

So the bar any honest answer has to clear is this: stop crediting the effort nearest the sale, run a comparison that isolates real lift, and keep a record clear enough that the result is legible rather than spun. The holdout is the test. The record is what turns the test into something you can act on instead of argue about.

JynAI built Works, an AI Business OS, to apply that discipline to everything the AI does, not just ad spend. The proof of the work is tied to the result, so “busy” becomes a number you can stand behind.

• Job is knowing which effort moved a line, not which one ran.
  Works logs every run, action, and outcome and rolls outcomes up at the area and workspace level, tying an action to a revenue or cost line.
  Gain is credit assigned to cause, not proximity.

• Job is running the comparison without building a measurement team.
  Works lets you scope a workflow or a sequence to a slice and track the held-back slice against the rest in the same record.
  Gain is a holdout you can actually run.

• Job is trusting the result enough to act.
  Works keeps the activity log append-only, so the numbers behind the comparison cannot be quietly reshaped to flatter the channel someone already believed in.
  Gain is a measure you can defend.

• Job is showing a partner or the board what actually worked.
  Works exports the outcome rollups to a board deck in docx, xlsx, or pdf.
  Gain is an honest story instead of a hopeful one.

The price makes the claim honest. The full capability set unlocks on the $49 Pro tier, not behind an enterprise measurement contract, which is the proof that a founder-led business gets honest measurement without an analytics hire. And we run on it: at Machintel, AI went live across six teams in ninety days, and the outcomes were tied to the work from the start.

Get honest attribution. Get early access. Or see the shape first with the proof-of-work example, a worked holdout you can copy.

What is honest attribution?

Honest attribution is the discipline of crediting the effort that caused the lift rather than the one that stood closest to the sale, verified by a holdout rather than a dashboard. The default tools over-credit existing demand and let multiple platforms claim the same conversion. In a 2026 survey, three in four marketers said their measurement was not delivering the accuracy they needed, which is the scale of the default problem.

How do I tell real lift from vanity metrics?

Incrementality is the standard: real lift is the gap between what happened with the effort and what would have happened without it, isolated by comparing a treated group against a held-back slice. A rising number on its own is correlation wearing a confident face, not causal proof. The eBay paid-search experiment is the cleanest illustration: the ads were taking full credit for buyers already coming regardless.

How is this different from measuring AI ROI?

Honest attribution answers which specific effort moved the number. Measuring your overall AI return, whether the spend is paying off across the board, is a broader question covered in the honest AI audit. Attribution is the per-effort truth; ROI is the portfolio view. This piece is the first one.

Do I need an expensive tool to run a holdout?

No. A holdout on your own audience costs nothing but the discipline to wait and compare. And for paid media, the minimum incrementality-test budget has dropped from around a hundred thousand dollars to roughly five thousand, which puts the formal version within reach of a mid-market business for the first time.

What does this have to do with the work AI actually runs?

Everything. You can only attribute what you can see, which is why a readable record of what the AI did comes first, and the lift you are attributing is the output of the marketing function you could run without the hire. Honest attribution is the test you run on top of that work.

1. Blake, Nosko and Tadelis, “Consumer Heterogeneity and Paid Search Effectiveness: A Large-Scale Field Experiment,” Econometrica, 2015 (eBay)
2. Gordon, Zettelmeyer, Bhargava and Chapsky, “A Comparison of Approaches to Advertising Measurement: Evidence from Big Field Experiments at Facebook,” Marketing Science, 2019
3. Adweek (Lauren Johnson), “When Procter & Gamble Cut $200 Million in Digital Ad Spend, It Increased Its Reach 10%,” 2018
   
   

Keep reading:

• What the AI actually did
• Marketing ops without the hire
• Provable AI