You are arguing about the attribution model. You already lost the taxonomy.

There is a real fight underway about the future of web measurement, and I want to state it fairly before I say why I think it is happening one layer too high.

A W3C Working Draft published in late July specifies a browser API for measuring advertising performance. A Recommendation is expected by the close of the year. The design goal is aggregate conversion statistics without exposing records of any individual’s browsing. And unlike the sandbox work tied to a single browser, the W3C ambition is a standard adoptable across browsers — which makes it a serious candidate for the default definition of a measured conversion on the open web. The reports it hands back are pre-aggregated and carry differential-privacy noise, rather than the event-level records the industry has built two decades of practice on.

The readiness numbers are worth sitting with. As of a W3C meeting in March, only 12% of the top-1,000 advertisers by Chrome ad spend had completed integration testing on the related reporting API. That API’s own timeline has already slipped, citing advertiser readiness. Twelve percent, on the thing that may define what a conversion is.

Most coverage frames this as a privacy story. The sharper coverage frames it as a governance story, and asks whether the companies best resourced to shape a browser standard — which are also the companies selling the ads — end up tilting the model in their favor. That is a legitimate worry and I share a version of it. But it is still an argument about the model.

The layer underneath

Every position in that debate presupposes something: that the touch was correctly identified as a paid touch in the first place. Model choice, noise tolerance, lookback length — all of it operates on a set of touches that someone already labeled.

In the work I actually see, that labeling is where it breaks. And it breaks on things nobody writes standards about.

The identifier that is missing. For certain placements the traffic tracks off URL parameters like anything else, but a program identifier has to be in the landing URL or the activity cannot break out on a report at all. The principle generalizes past that one case: the identifier a report will later slice on must be present at tagging time, or no attribution model recovers the breakout later. I have watched automated checks flag the same account for the same missing identifier in clusters, weeks apart. It is not a modeling failure. It is a naming convention, and it recurs monthly.

The two URLs that are one page. A trailing slash, a casing difference, a parameter order, www against apex. The click logs against one string, the landing-page tag fires against another, and the two never join in the conversion path. The trailing slash is the most common offender for a precise and irritating reason: it is the one difference ad platform interfaces silently normalize behind the scenes, so the mismatch is invisible exactly where the work is being done. Computers see two URLs as two pages. They are right, and it costs you the conversion.

The parameter the network will not accept. A major network blocked event parameters on multiple client pixels inside a handful of weeks, hitting several accounts, one of them repeatedly. Read as incidents, that is a run of unrelated client mistakes. Read as a pattern, it is one shared tagging template carrying a single field the network’s classifier reads as something it will not take — which means the fix is one audit of the template, not a queue of per-account patches. Downstream, continued blocks constrain custom conversions and audiences, so the cost lands on both optimization signal and retargeting.

The rollup that drops a grain. City-level spend that mostly failed to aggregate up into the state breakdown, so an entire geographic view reported a fraction of reality. Understated by a multiple, not a margin. The reason that class of bug survives is the reason it is dangerous: the totals look reasonable, so nobody queries them.

Not one of those four is a model dispute. Each one decides whether a touch exists.

The definitions you do not control

Then there is the version where you did everything right and the definition moved anyway.

Earlier this year a major analytics platform shortened its default acquisition-conversion lookback from 90 days to 30. Some accounts reported search conversions down roughly 18% while their sales teams saw steady leads. That is the unmistakable signature of a definition change wearing performance clothes, and the remedy is to put the window back where the actual buying cycle is rather than where a default chosen for someone else’s account shape put it.

I had the quieter version of the same thing in my own house. A headline metric that ignored the date picker sitting directly above it, reporting a fixed trailing window no matter what range you selected, with a standing disclaimer as the only signal that it worked that way. We fixed it. And the consequence of fixing it is the part worth generalizing: every prior explanation anyone had given about what that number meant became wrong on the day it started behaving correctly.

A definition the reader cannot see, underneath a control the reader believes they hold. That is the same shape as the lookback change, and the same shape as all four tagging failures above.

What I would actually watch for

So I do not think the browser API’s most important consequence is the noise. Noise you can reason about; you can model error bars and stop pretending you ever had certainty.

The consequence is that the definition moves somewhere you cannot inspect. Today, when a touch is misclassified, I can go find the URL string, or the campaign name, or the blocked parameter, and see exactly why. Pre-aggregated, pre-attributed browser output removes the artifact I would have inspected. I get a number and an assurance. The 12% integration-testing figure suggests very few people have noticed this yet, and I think I know why: you only notice it the first time you try to debug a number and discover there is nothing left to debug.

And that is the honest version of the Big Tech worry. The concern people voice is that the giants will tilt the model. The more durable power is definitional. Whoever specifies what event a browser is willing to call a conversion has drawn the boundary of the argument, and every model dispute afterward happens in a vocabulary they wrote.

The takeaway

Follow the standards process — it is real, it has a deadline, and the governance concern is not paranoid. But do not let it pull your attention off the layer where your measurement is actually being decided this quarter. Pick one canonical URL form per account and enforce it before publish. Put the identifier in the URL at trafficking time, because nothing downstream recovers it. Audit the tagging template once instead of patching accounts as the notices arrive. Check your totals against a grain you did not aggregate, because a number wrong by a multiple can still look perfectly reasonable. That is unglamorous work and none of it will be in a standard. It is also what determines whether the touch exists at all — and every sophisticated argument about which model deserves the credit is downstream of that.