Tuesday, spend review. One line in the sheet has a CPA roughly double everything else. It has looked like that for six weeks. You pause it, move the budget to the line with the best CPA, and feel like you did your job.
Three weeks later blended CPA is up, total orders are down, and the channel you moved the money into has quietly gotten more expensive too. Nobody in the meeting can explain it, so it gets written off as seasonality.
That sequence is common enough to be a genre. It usually is not a media problem. It is that the report you cut from was never able to see the thing you cut.
"Direct" is not a channel, it is a shrug
Here is Google's own definition, which is more honest than most of what gets built on top of it:
A session is processed as direct traffic when no information about the referral source is available, or when the referring source or search term has been configured to be ignored.
No information available. That is the whole bucket. Direct is not people who love you enough to type your URL from memory. It is every visit that arrived without a trail, and there are a lot of ordinary ways to arrive without one:
- Someone pastes your link into a Slack message, a text thread, or a DM.
- A link opened from a desktop email client, a PDF, or a notes app.
- A podcast mention, a group chat, a private community, a forwarded screenshot.
- A redirect chain that dropped your parameters somewhere in the middle.
Every one of those is real demand that something you paid for created. All of it lands in a bucket labeled with a shrug, and none of it gets credited back to whatever caused it.
The tag can be perfect and the trail still goes cold
The other half is the browser, and this part is worth knowing precisely because it gets misdiagnosed as a tracking bug.
Safari's tracking prevention caps how long your measurement survives. When a visitor lands from a domain Safari has classified, on a URL carrying a query string, persistent cookies written in JavaScript on that page expire in 24 hours. Non-cookie browser storage gets cleared too: after seven days of Safari use without the visitor interacting with your site again, it is deleted.
Now put a normal buying pattern against that clock. Someone taps your ad on a phone Tuesday during lunch. They think about it. They come back the following week on a laptop and buy. Nothing about that is unusual, and there is no mechanism left to connect the two. The purchase gets filed under direct, and the ad that started it shows one more impression with nothing to show for it.
That is not a misconfiguration you can fix by re-checking your tags. It is the browser doing exactly what it was built to do.
Last click pays the closer
Stack those two on top of a last-click model and the bias always points the same direction. Credit collects at the bottom of the funnel, on the branded search click, the retargeting ad, the discount code someone searched for on the way to checkout. Those things did not create demand. They caught it.
Meanwhile whatever created the demand shows a bad CPA, because most of the people it moved came back later through a door the model cannot see. So you cut it. And a few weeks on, branded search volume softens, direct softens, and the "efficient" channels get less efficient, because the thing feeding them stopped.
Cutting the channel that fills the top and keeping the one that harvests the bottom is how a funnel starves quietly.
What to do before you pause anything
None of this means attribution is useless or that you should run on vibes. It means a channel report is evidence, not a verdict. Four things that cost close to nothing:
- Ask. Put a "how did you hear about us" question on the post-purchase page or the form confirmation, with a free-text option. It is imperfect and biased toward what people remember, which is exactly the bias you need, since what people remember is what the trail lost. Compare it against platform numbers monthly and watch where the two disagree.
- Turn it off on purpose. Kill the channel in a handful of matched markets for three or four weeks and watch total orders, not that channel's CPA. If nothing happens anywhere, you have your answer and you can cut with confidence. If total demand dips in the test markets, the channel was doing something the report never showed you.
- Track direct and branded search as outputs. Chart them against upper-funnel spend instead of reading them as their own channels. When they move a few weeks after spend moves, that relationship is the credit that last click refused to assign.
- Read the lag before you judge. Pull time-to-conversion and assisted conversions for the channel in question. If most of its conversions land well outside your attribution window, the CPA in your sheet is a rounding error on the real number.
Where we land on this
All four of those are ways of getting at one question: who actually showed up, and what put them there. Most measurement tries to answer it by following a trail between sessions, which is the part that keeps breaking.
The visit itself does not break. Someone lands on your site whether or not the cookie survived the week, and whether or not your UTMs made it through the redirect. Knowing more about who arrived is a different way at the same question, and it does not ask you to change anything about how you buy media or where your data goes afterward. You are already paying for the traffic. You already have the CRM and the email tool and the retargeting audiences. This is about the arrival, not a new motion to learn.
Whether that is worth doing at your volume depends on numbers you already have, and we are happy to go through them with you and say so if it is not.