Can you compare conversions reported by two different ad platforms?

Not directly, and adding them together is the most common way the error enters a board deck. Each platform writes its own definition of a conversion: the attribution window differs, the rules about which interactions earn credit differ, and the decision about whether post-view credit enters the main total at all differs. Two platforms can both be entirely accurate about the thing each one defined and still produce numbers that cannot be summed, in the same way that two accurate measurements in different units cannot be added. The practical fix is not to pick a winner between them. It is to record, per platform, what the unit is, what window it uses, and whether post-view credit is inside or outside the total, and then to reconcile against one source of truth you control, such as orders in your own system, before anything is blended.

What is a view-through conversion and should you count it?

A view-through conversion is credited when someone converts within a set window after seeing an ad without clicking it. Whether to count it depends on a question most teams skip: what the platform itself does with the number. Some platforms wall post-view credit off from bidding, billing and cost per acquisition, so it appears in a reporting column and influences nothing. In that case a rising view-through count is not evidence that automation is working, because the automation cannot see it. Others fold post-view credit into the main total, where it does move bids and budgets. Before using the metric, find out which of those two situations you are in, and never compare a post-view number from one platform with a click-based number from another as though they measured the same event.

Who should sign off on how a platform defines a conversion?

A named person, before the spend starts, and in practice this usually belongs with whoever owns the blended acquisition number rather than with the person running the account. The definition of a conversion is not a media setting; it is the denominator of every efficiency figure the company reports, so accepting a platform's default definition is a finance decision that has been delegated by accident. The lightweight version is a single row per seller recording the unit as that seller defines it, the attribution window, whether post-view credit is included, the reviewer's name and the date of the last check. If no one signs off, the definition still gets accepted, just without anyone deciding to accept it.

Does using more ad platforms make your measurement more reliable?

No, and the belief that it does is one of the more expensive misreadings in marketing. Adding a platform feels like adding a second opinion, but each platform sets its own bids, selects its own surfaces, defines its own conversion, counts it, and takes your outcome data back to optimise against. A second platform reproduces that arrangement rather than checking it, so what you gain is a second closed system with its own vocabulary, not an independent estimate. Reliability comes from holding at least one position outside the loop: a measurement produced by a method you chose and can run again, such as a geo holdout or a mix model on inputs you control, or simple reconciliation against orders in your own system of record.

Who Counts the Conversion?

The slide said blended cost per acquisition and it had one number on it. Underneath, in the working file, that number was a sum: conversions from one advertising platform plus conversions from another, divided into total spend. Nobody in the room had asked whether the two halves of the numerator meant the same thing. I am not sure anyone had considered that they might not.

That is not a failure of diligence. It is what happens when a company adds a second advertising platform and files the decision under media buying, which is where it has always belonged. You test a channel, you get a conversion count back, you add it to the others. The arithmetic is so ordinary that the question underneath it never surfaces: who decided what a conversion is on each side of that plus sign, and did anyone at your company sign off on either definition?

The obvious reading, and why I am conceding it first

In August a second advertising platform, one that did not have a self-serve auction eighteen months ago, shipped automated bidding, surface-level targeting, and post-view conversion reporting inside a few weeks of each other. The obvious reading is that a challenger is building the features the incumbent already has. That reading is correct. Any competent product team would ship exactly these things in roughly this order, and treating a product roadmap as a revelation is how trade coverage fills a slow week.

So let me take the objection seriously rather than set it up to knock down, because it is the one I would raise. This is catch-up. Nothing here is novel. If the argument depends on the features being surprising, the argument is dead.

It does not depend on that. It depends on the opposite. Everything I have written on this site about sellers and units and who owns which decision was derived by watching one company, and the fair response has always been available to anyone who wanted it: that is a Google problem, and Google is unusual. Largest ad system in the world, twenty-five years of accumulated defaults, a regulatory posture nobody else has. Maybe the structure I kept describing was just one firm's idiosyncrasy dressed up as a law.

I never had a way to test that. This month I got one, and it did not come from Google.

A shape one company arrives at is a strategy. A shape two unrelated companies arrive at independently is a structure.

What actually shipped, stated precisely

Precision matters here more than usual, because the sloppy version of this essay is wrong. The new bid strategy is called Maximize results. It sets and adjusts bids automatically toward the campaign goal already selected, clicks or conversions, and aims to return as many results as possible from the available budget. Its documentation carries an update stamp from around August 12, 2026, and a product update email listed it first among six changes during the week beginning August 17.

Two details in that release are doing far more work than the headline.

The first is that Maximize results holds no cost target at all. The documentation says plainly that it prioritises result volume and does not guarantee delivery against a specific CPA, CPC, ROAS or other cost-efficiency target at this time. This is not the incumbent's target-based bidding with a new label. Target CPA expresses a number the system is supposed to hold. This expresses volume, with the budget as the only constraint. Whatever cost per outcome falls out of the auction is the cost per outcome. Manual bidding remains the only documented route to a hard ceiling.

The second is that it arrives preselected. For eligible new ad groups, Maximize results is the default, reviewable in the bid strategy section if you go looking. There is no account-level preference that makes manual stick. The choice is presented per ad group, every time, to an advertiser base that six months ago required a two hundred thousand dollar minimum commitment and now requires nothing at all.

The default is where the decision actually moved

Put the bid strategy next to two other changes at the same platform in the same six weeks. Fixed daily budgets were converted into seven-day averages during July with no advertiser opt-out. Automatic advanced matching became the default on existing web pixels on August 17, after a ten-day window to decline it.

Three defaults. Three decisions that used to sit with the advertiser: what a bid is, what a budget means on any given day, and what identity data leaves your site. None was announced as a transfer of control, because none of them is one. Each was a sensible product decision, and in each case the advertiser kept the theoretical ability to choose otherwise.

This is the same mechanism I wrote about when the incumbent rewrote its terms of service and nothing changed in anybody's account, which was exactly the problem. A default owner is not someone who took a decision from you. It is whoever ends up holding a decision that nobody in your organisation ever made. The transfer does not require an announcement. It requires a preselected radio button and a person building their fourth ad group at four in the afternoon.

The objection I find hardest

Here is the one that comes closest to ending this. Advertising platforms have self-reported their own effectiveness for thirty years. Panels, post-logs, walled gardens, view-through windows chosen by the seller. Nothing about a party grading its own homework is new to anyone who has bought media, and a writer who has just discovered it is late by about three decades.

That is right, and I want to concede it fully rather than partially. Self-reporting is not the change.

The change is where the self-reported number terminates. For most of those thirty years, a seller's account of its own performance ended in a human being reading a report, forming an impression, and going to a meeting. That is slow, and slowness is a kind of safety: a person who reads a suspicious number can sit with it. Now the number terminates in a bidder that acts on it continuously, and increasingly in an agent that reads it overnight and adjusts. I made this same distinction a few days ago about whether you are permitted to hold evidence at all, and it holds here in its plural form. A self-issued number read once a quarter is a reporting problem. The same number wired into an automated decision is a control problem.

A second seller is not a second opinion

Now the part that changes what a marketing organisation should actually do, and it needs no Google in it whatsoever.

When a company adds a second advertising platform, the instinct is that it has acquired a check on the first. A benchmark. Somewhere to look when the incumbent's numbers seem too good. That instinct is wrong, and it is wrong in a way that is hard to see because it feels like prudence.

You have not acquired an independent estimate. You have acquired a second closed loop. The second seller sets the bid, selects the surface, defines what counts as a conversion, counts the conversions, and takes your outcome data back through its own conversions API to optimise against. Those are the same positions the first seller occupies. Nothing about the second one is external to the first one's problem; it simply reproduces it, using its own definitions, on its own window.

And then the numbers get added together. This is the part I would put in front of a finance leader before anything else in this essay. The seller's unit, the thing you are actually buying and being billed against, is now written independently by each platform you use. A conversion at one seller and a conversion at another are not the same object: different attribution windows, different rules about what earns credit, different eligibility for what enters the total at all. Neither number has to be wrong for the sum to be meaningless. This is unit drift in its plural form, and where a single seller's drift happens slowly over years, adding a seller does it to you in an afternoon.

The site already has the single-seller version of this: a number the seller produces about its own contribution is a seller's counterfactual, and it may be perfectly accurate while still being a quote from an interested party. What changes with two sellers is not the trustworthiness of either quote. It is that you now have two, in different currencies, and a board deck that adds them.

Yes, it is a rounding error, and yes, it still counts

The finance objection arrives immediately and it is fair: this platform is two percent of my spend. You are asking me to build governance for a rounding error.

Right about the money. The unit is wrong. Governance cost does not scale with dollars spent, it scales with the number of definitions you have agreed to accept, and a seller at two percent of budget contributes one hundred percent more conversion definitions to a blended figure that has one. The cost of a second seller does not appear in the media budget. It appears in the reconciliation, and it appears again every time someone asks why two systems disagree.

This is the same trap as budgeting an agentic stack in tokens: the meter you are billed on is not the unit your economics run on, so the number that matters never lands on anyone's line item.

It is also the answer to a question I raised about what happens when the transaction moves onto the seller's surface. That essay asked what a cart is worth when the cart is not yours. This one asks the prior question, and it is duller and more consequential: what a conversion is when the definition is not yours either, at more than one seller at once.

Where I might be wrong, with a date on it

The strongest evidence against my own argument is in the same release, and I would rather present it myself than have it presented to me.

When post-view conversion reporting appeared at this platform in mid-August, it arrived with a fixed one-day window and it arrived walled off. Post-view credit is excluded from the main conversions total, from cost per acquisition, from billing, and from bidding. It shows in a column and does nothing else. That is a genuinely conservative choice, more conservative than the industry standard, and it cuts against the closed-loop reading: the seller built a number and then declined to let it touch the loop.

It also creates something odd that nobody seems to be discussing. There are now two conversion counts inside one platform, the one that governs bidding and the one in the reporting column, and they diverge by construction. An advertiser watching post-view credit rise and concluding the automation is working is reading a figure the automation cannot see.

A claim you can check

Thirty essays here argue, and an argument cannot be scored in retrospect. So, with a date on it:

On April 30, 2027, open the ChatGPT Ads help documentation and check whether post-view conversions are still excluded from the main conversions total and from bidding. I claim the wall will be gone or materially lower by then, and that the change will be presented as a reporting improvement rather than as a change to what counts as a conversion.

Ninety seconds to verify, and it can go against me in public, which is the point. If the wall is still standing in April, my read on which direction seller-defined measurement travels was wrong. A smaller one already resolved: on July 17 I wrote that stale targets were about to start binding, and on August 17 budget-limited campaigns began delivering to the target you typed rather than the better number you had been getting.

What to write down before the next platform

The instrument for this already exists and it needs one more column. The Unit Ledger in the seller's unit was built to keep custody of denominators at a single platform. Give it a row per seller: the unit as that seller defines it, the attribution window, whether post-view credit is inside or outside the total, who at your company reviewed the definition, and the date they last checked it. Five fields. It fits on one screen and it takes an afternoon.

That row is the difference between a blended number your finance partner can defend and one that merely exists. It is also the only version of platform diversification that produces independence, because independence was never about how many sellers you buy from. It is about whether any position in the loop is held by someone who is not being measured by it.

And it is where the ownership question finally gets concrete. You own a decision class, not a tool, and the decision class here is not which platforms to run. It is what your company will accept as evidence that something worked. That class currently has no owner at most companies, which means it has a default owner, and the default owner is whichever vendor's dashboard loads first on Monday morning. The rest of the vocabulary for this, decision classes, kill conditions, counterfactual audits and the ledgers underneath them, lives on the hub.

I have been wrong before about how fast this consolidates, and I could be wrong about the direction too, which is why there is a date on the last claim. But the thing I am confident about is smaller and duller than a prediction. Before a second seller's numbers enter a blended figure, write down what that seller's unit is, what window it uses, and whose name is against the sign-off. Do it while the spend is still a rounding error and the exercise takes an hour. You cannot add two numbers defined by the people being measured by them and call the total evidence. Write down what a conversion means at each seller before you add them, because after you add them, nobody can tell.