Which Marketing Channel Is Actually Bringing Me Customers?

Add up the sales Google Ads, Meta and TikTok each claim for last month, and the total often beats your real order count. GA4 hands a large slice to "direct," a channel nobody runs. Ask which marketing channel is working and every report gives a different answer, and your budget follows whichever one you believe.
No dashboard can settle it, because attribution records which channels touched a sale, not which ones caused it. Use your order system as the only total and compare channels on one basis. Then run holdout tests where the biggest spend and the most flattering numbers sit. A better attribution model won't tell you where customers come from. Those tests will.
Why does every ad platform claim the sale?
Each ad platform measures conversions from its own side of the wall. A purchase comes back through the pixel, the tag or a server-side integration. The platform then searches its own logs for any click, engaged view or impression from that person inside the attribution window you set. If it finds one, the sale is its sale. It never checks whether a rival touched the same buyer, because the walled gardens don't share user-level data with each other.
So one ordinary journey produces two claims. A shopper clicks a Meta prospecting ad on Monday and watches part of a YouTube video on Wednesday. On Thursday they search your brand, click a Google ad and buy. Meta counts the order, and so does Google. Both are reporting correctly by their own rules. The error only appears when you treat their figures as shares of one total.
Three things widen the gap between claims and orders:
- Audience overlap. The more platforms you aim at the same audience, the more journeys include touches from more than one of them.
- View-through credit. A platform claims sales after an impression that no other tool can verify.
- Long windows. A longer window lets an ad claim more sales without causing a single extra one.
That is why adding Google and Meta conversions together produces a number that means nothing. Your order system is the only total that can't double count, whether that's Shopify or a CRM counting closed deals.
One related trap is worth separating out. If Meta's browser pixel and its Conversions API both log a purchase without a shared event ID, Meta counts that sale twice. That is a tracking bug with a tracking fix, and the fix only works inside Meta. It does nothing about Google claiming the same order.
Is GA4 the neutral referee?
It's closer, but no. GA4 sees paid and organic channels and gives each purchase one unit of credit, so its channel totals never exceed your real sales. That makes it the better basis for comparison.
It has blind spots of its own, though. It can't see ad views, so campaigns that work mainly through impressions look weak there. It also leans toward whatever click sat nearest the purchase. When branded search closes a journey that Meta opened, branded search takes the credit.
So the platforms over-report, and your analytics under-reports paid influence. The truth sits somewhere between them, and switching attribution model won't find it. First click vs last click, position-based, data-driven, Shapley or Markov: every flavor of multi-channel attribution redistributes credit among the touches that happened. None of them asks whether the sale would have happened without the ad. An attributed sale is not an incremental one.
What to do with platform numbers
Use each platform's figures for what they're good at, which is ranking campaigns, ad sets and creative inside that platform. Comparing across channels needs three habits:
- One total. Pick one source of truth for totals.
- One basis. Compare every channel on the same window and the same model.
- A separate column for views. Keep view-through conversions apart, so the share nobody can verify stays visible.
The hard part is doing that reconciliation every week without a spreadsheet that's stale by Friday. SproutMe Companion reports spend and claimed conversions from Google Ads and Meta Ads alongside orders from Shopify and pipeline from HubSpot. The gap between what was claimed and what was sold sits in one report. The full mechanics of windows, views and overlap are in why every ad platform says it brought the sale.
Why does so much traffic show as direct?
If the platforms claim too much, direct is where your analytics quietly loses credit. Nobody runs direct as a channel. GA4 assigns every session a source from the campaign parameters in the landing URL. Failing those, it uses the referrer the browser sends. When neither exists, the visit becomes `(direct) / (none)`. Direct counts an absence of evidence, not loyal customers typing in your URL.
What actually lands in direct
Most of it is traffic that came from somewhere and lost its label on the way:
- Untagged email, since most mail apps open links without sending a referrer.
- Messaging and social apps that strip referrers for privacy. A link shared in WhatsApp or Slack looks identical to a typed visit.
- AI assistants that pass no referral. Buyers who read a recommendation and then type your brand in land here too.
- Redirects, shorteners and URL rewrites that drop UTMs or the referrer before the page loads.
- Bots that execute your tag and announce nothing.
This matters for attribution because every misfiled session is credit taken from the channel that earned it. An untagged newsletter makes email look weak and direct look strong. A podcast mention that sends people to type your URL never shows up as podcast at all. Your channel report doesn't just get noisier. It systematically flatters whatever you did tag and penalizes whatever you didn't.
Shrinking it, and what you can't fix
You can't recover a lost referrer, but you can usually tell which direct sessions were never direct:
- Landing page. Visits that land on a deep blog post or a campaign page followed a link from somewhere.
- Timing. A spike that lines up with a send date came from the send.
- Engagement. Near-zero engagement clustered on a handful of URLs points to bots.
Then fix the plumbing:
- Tag every link you control, using one naming convention.
- Audit your redirects so parameters survive to the landing page.
- Store parameters at landing. Put UTMs and click IDs in a first-party cookie so a later page can't lose them.
- Filter known bots.
There's a tension here worth knowing about. UTMs fix what you send, not what people pass on. If you tag a link and someone copies it into a group chat, that conversation gets credited to your campaign. If you leave it untagged, it stays direct.
Only one method sees dark social at all: a "how did you hear about us?" field at sign-up or checkout. It's self-reported and imperfect. For marketing attribution for small business, it's often the highest-value change available. It costs nothing and covers exactly the word of mouth no tag can see.
Read direct as "unknown," not "brand strength." The full diagnostic checklist is in why so much of your traffic shows as direct.
Is retargeting creating sales or claiming them?
Direct absorbs credit other channels earned. Retargeting has the opposite problem: it collects credit for sales that were already on their way. In most accounts, it's where reported and real results sit furthest apart.
The reason is structural. Retargeting audiences are chosen for a single property, which is intent. A cart abandoner is far more likely to buy than a cold prospect, with or without an ad. A high conversion rate among people who saw the ad therefore tells you about the audience, not the ad.
Last click then credits whatever sits at the end of the journey, which is exactly where retargeting is built to sit. Take a buyer who found you through organic search, read reviews and compared prices. They see one final banner, and the sale lands on the banner. View-through windows then add orders from ads that were merely scrolled past.
How much of it is real?
There's no benchmark you can plan against. Published retargeting holdouts are mostly single-account anecdotes with no method attached. A blanket discount like "assume it's worth half what Meta says" just swaps one unmeasured number for another.
What does hold is a rule of thumb that follows from the mechanism. The closer the audience sits to purchase, the higher the reported return and the smaller the share that's incremental.
- Close to purchase. Checkout abandoners from an hour ago post spectacular ROAS and probably caused little of it.
- Further away. People who read one blog post three weeks ago look mediocre in the dashboard and may be doing more real work.
Retargeting tends to create sales when:
- the purchase is considered and takes weeks
- the segment is cooler
- recent buyers and in-checkout visitors are excluded
- click windows are short and view-through credit is minimal
It tends to capture existing demand when:
- the product is impulse-priced
- the audience is pulled warm from checkout pages
- view-through windows are long
- frequency runs high against a small pool
Branded search follows the same logic. A shopper who searches your name and clicks the paid ad would often have clicked the organic listing directly below it. Treat both as the first suspects whenever the platform totals outrun your orders.
When a test comes back weak, the right response is usually to shrink retargeting, not kill it. Keep the segments that showed lift and exclude recent buyers. Then move the freed budget toward prospecting, which feeds the pool retargeting later harvests. The conditions and the decision are covered in depth in is retargeting making sales or taking credit for them.
How do you test what a channel really adds?
Every section above ends at the same wall. Attribution answers a bookkeeping question: which touch gets the credit. You need the causal one: would the sale have happened without the ad?
A holdout test builds that counterfactual. Withhold the channel from a matched group of people or regions, keep it running for everyone else, and compare conversions over the same period. Both groups face the same season, promotions and competitors, so the gap between them is what the channel caused. Experiments with control and exposed groups are the way to find a channel's true incremental value, per WARC's 2022 best-practice guide. The same guide warns that over-reliance on last-touch attribution breeds short-term thinking and poor budget allocation.
Platform lift study or geo test?
The quickest route is the one the platform already offers. At the time of writing, Google's Conversion Lift documentation describes holding a control group of users back from your ads. It compares their conversions with the group that saw the ads. It ignores your normal attribution settings and counts every conversion across the study, not just those inside a window. Meta and TikTok run similar studies.
It's worth holding two views of these studies at once, because both are true:
- The design is clean. Randomizing users is the cleanest design available inside one platform.
- The referee is also the player. The platform designs the test, runs it on data you can't audit, and is the channel being graded. It can only test itself. A Meta lift study tells you what Meta added. It can't tell you whether Meta or Google deserves your next dollar.
A geo-lift test gives you a result nobody else controls. It also works for channels with no lift product, such as TV or a podcast sponsorship. The method:
- Match the markets. Pair test and control regions on sales history and size.
- Check the baseline. Confirm that four to eight weeks of pre-launch sales move together.
- Keep the controls dark. Any spend that leaks in shrinks the gap you're measuring.
- Measure against your own orders. Compare test and control in your order system, not the platform's.
On duration, two to four weeks is a reasonable floor for fast purchase cycles. If your buyers take longer to decide, run past your normal purchase cycle so delayed conversions land inside the test.
The cost is real. You give up the sales the channel would have driven in the control markets. You also need enough regional volume to separate signal from noise.
What if you're too small for a geo test?
Then start cheaper:
- Check claims against orders. If what the platforms claim clearly exceeds the orders you received, something is being double counted.
- Black out branded search. Pause it for a period and see whether organic sales make up the difference.
- Run a platform lift study on your biggest channel, accepting its limits.
- Ask customers how they found you, if you haven't already.
None of these is as clean as a well-powered geo test. Each one replaces an assumption with an observation, which is the whole point.
Reading the result
Wait for the study to close, because early reads swing. Then turn lift into money. Divide spend by incremental conversions to get incremental CPA. Google's own hypothetical is a $10,000 campaign that produces 800 incremental conversions, an incremental CPA of $12.50. Set that next to the CPA your dashboard reports. The distance between the two is how much credit attribution was giving the channel for sales it didn't cause.
Compare channels on incremental cost, not relative lift percentages. Those balloon when the control group is small. Decide before launch which result would change your budget, because a test nobody committed to acting on is just an expensive report. The full design and reading guide is in how to test what a marketing channel really adds.
Conclusion
No single report will tell you which channel is bringing you customers. Each platform claims every sale it touched. Analytics loses credit to direct and favors the closing click. Retargeting sits exactly where credit is easiest to claim and hardest to earn. Changing attribution model only reshuffles the same touches.
What works is plainer:
- one source of truth for totals
- one basis for comparison
- cleaner tagging so your analytics loses less
- holdout tests on the channels carrying the most spend and the most flattering numbers
Let measured lift set the budget, read every dashboard with that discount in mind, and re-test when your market moves.
See how SproutMe Plan models budget scenarios across Google, Meta, TikTok and LinkedIn against one business objective before any of it is committed.
Frequently Asked Questions
Neither is better; they flatter different channels. Last click credits whatever sat nearest the purchase, which favors branded search and retargeting. First click credits whatever introduced the buyer, which favors prospecting and discovery channels. Comparing the two shows which channels open journeys and which close them. Neither tells you whether a sale would have happened without the ad, so use them for diagnosis, not budget.
No. The two work best together. A mix model estimates each channel's contribution from aggregate spend and sales over time, and open-source options such as Meta's Robyn and Google's Meridian exist. On its own, a model can mistake correlation for effect, especially for channels that track demand closely, like retargeting. Calibrating it with holdout results anchors its estimates to measured lift.
Re-test whenever something that drives demand changes: your prospecting mix, pricing, creative, competition or season. Incrementality isn't a fixed property of a channel, so last year's lift figure describes last year's account. A regular cadence on your major channels keeps the discount you apply to dashboard numbers honest. In geo tests, rotate which markets serve as test and control.
Get a complimentary audit to uncover AI opportunities hidden in your data.
Put these strategies to work


