AI-Powered Performance MarketingAttribution, incrementality and mix modeling · Lesson 10 of 15
Attribution reality: what each number can and cannot tell you
Video lecture
Attribution reality: what each number can and cannot tell you
The narrated lecture is in production
Every chapter is scripted and ready. Browse the chapters and read the full transcript now — the video will appear here when it’s published.
Chapters
Transcript of the narration, chapter by chapter.
0:00 Attribution reality
Here's a scene that plays out in marketing meetings every week. Meta says it drove a thousand sales. Google says eight hundred. TikTok says four hundred. The finance team says you sold one thousand two hundred in total. Who's lying? Nobody. In this lecture you'll learn what each of those numbers actually measures, why they never add up, and how mature teams give every number a clear job so the arguments stop.
0:31 Why it matters
Why does this matter? Because arguments about which dashboard is right waste enormous amounts of time, and they usually end with the loudest person winning. Here's an analogy. Imagine a football goal. The striker scored, the midfielder made the pass, and the defender won the ball back. Who gets credit? Each player's highlight reel shows them as the hero. Ad platforms are highlight reels. Each one shows the goals it touched. None of them is lying, and none of them is the full match. Your backend is the scoreboard, and experiments tell you who actually changes results.
1:13 Platform numbers
Start with platform-reported conversions. Each platform counts conversions it can link to its own ads, through clicks and often through views, within its attribution window. It also models conversions it can't observe because of consent or browser limits. That's legitimate, but notice the pattern: every platform grades its own homework, and none can see what the others did.
1:38 Analytics and backend
Then analytics, like GA4. It sees sessions and conversions on your site and distributes credit across channels with a model, usually data-driven attribution. But it can't see ad impressions that didn't lead to a click, and it loses people who decline consent or block scripts. Your backend, the shop or CRM, sees the truth about orders, revenue, refunds and new customers, but has no idea which ads anyone saw.
2:08 Why totals overshoot
So why do platform totals add up to more than your real orders? Four reasons. Overlapping credit: the same customer saw a TikTok ad, clicked a Google ad and opened an email, and all three claim her. View-through: an impression before a purchase gets credit even if it did nothing. Modeled conversions add estimates. And baseline capture: ads get shown to people who would have bought anyway.
2:37 Simple example: one Abu Dhabi customer (illustrative)
Here's a simple worked example. A customer in Abu Dhabi sees a TikTok ad on Monday, clicks a Google Shopping ad on Wednesday, and buys on Thursday after opening a promotional email. TikTok, with its view window, claims the sale. Google, with its click, claims the sale. The email tool claims the sale. GA4 with data-driven attribution splits credit between Google and email, and doesn't see TikTok at all because there was no click. One sale, four claims, four different stories. None is wrong in its own terms. Only a test can tell you what would have happened without TikTok.
3:20 Attribution models
A word on attribution models. Last click gives everything to the final click, which flatters brand search and retargeting. Data-driven attribution, the default in Google Ads and GA4, spreads credit based on observed paths, but only across touchpoints it can see. Here's the key point: none of these models is causal. They divide credit. They don't tell you what would have happened if the ad had never run. For that, you need experiments.
3:52 Triangulation stack
The solution is triangulation. Platform data steers day-to-day decisions inside each channel. Analytics explains site behavior and paths. Your backend is the scoreboard. Experiments give causal answers for key channels. And a marketing mix model sets the big budget split, calibrated with those experiments. Then create an incrementality factor. If Meta reports a thousand conversions and a lift test says six hundred were truly incremental, your factor is zero point six. Use it for planning until the next test.
4:26 Example: US brand selling to UK/Gulf (illustrative)
Here's an illustrative example. A US modest-fashion brand selling into the UK and the Gulf found its dashboards summed to about one point six times actual orders. They made the Shopify backend the single source of truth, used GA4 for funnel analysis rather than channel credit, ran one lift study on Meta and one geo test on Google, and reported weekly on blended new-customer acquisition cost and incrementality-adjusted conversions. The arguments about which dashboard was right simply stopped.
5:00 The weekly blended table
The lesson includes a SQL query that builds the most useful weekly table in marketing: total ad spend, new customers from the backend, blended new customer acquisition cost, and margin after ads. It's simple, it's hard to argue with, and it keeps everyone honest while the channel-level debate continues. Avoid the classic traps: judging one platform with another platform's numbers, and changing attribution windows mid-quarter.
5:28 Mistakes + try this now
Common attribution mistakes. Judging TikTok with Google's numbers, or Google with Meta's. Treating data-driven attribution as proof of cause. Switching attribution windows mid-quarter and comparing periods. And spending months searching for a perfect multi-touch model instead of running one good test. Try this now: build the simplest table in marketing. Four columns for last month: total ad spend, new customers from your backend, blended cost per new customer, and margin after ad spend. Share it with your team and agree it's the scoreboard.
6:04 Quick self-check
Quick self-check. Your Meta dashboard says return on ad spend is five. Your finance team says total marketing spend rose twenty percent last quarter while new customers rose only five percent. Which number should drive next quarter's budget conversation? Pause and think. The answer: the backend numbers set the scoreboard, and they're telling you the extra spend bought few new customers. Meta's five isn't wrong, it's measuring credit, not cause. The right next step isn't to argue about dashboards. It's to run a lift or geo test on the channel that grew most, so you replace the argument with evidence.
6:47 Watch me do it: the scoreboard (illustrative)
Watch me do it. Let's build the one-table scoreboard for an illustrative Karachi fashion brand, using the SQL pattern from the lesson. First, spend: I union daily cost from Meta, Google and TikTok into one table, converting everything to rupees on the same date. Second, orders: from the store database, only paid orders, with a new-customer flag based on the customer's first order date, and contribution margin after returns. Third, the join by week. The first result is eye-opening: last month, spend rose fifteen percent, new customers rose four percent, and blended cost per new customer rose ten percent, while every platform reported better ROAS. Fourth, I add one more column: incrementality-adjusted conversions per channel, using the factors from our last lift and geo tests. Meta's factor is zero point six, TikTok's zero point eight from a recent geo test. Now the meeting has one table everyone accepts, and a clear question for next month: which channel's extra spend actually bought the new customers?
7:58 Recap and next step
Recap. Every number has a job. Platforms steer, analytics explains, the backend keeps score, and experiments and mix models decide budgets. Totals overshooting is normal. Attribution divides credit, it doesn't prove cause. Your next step: write a one-page measurement charter saying which number is your scoreboard, which numbers steer each channel, and which tests you'll run this quarter.
Four numbers, four truths
A single purchase can be claimed by Google Ads, Meta, TikTok, your email tool and GA4 at the same time. None of them is lying; each answers a different question with different data.
| Source | What it measures | Blind spots |
|---|---|---|
| Platform-reported conversions | Conversions the platform can link to its own ads (clicks and, often, views) within its attribution window, including modeled conversions | Credits itself; cannot see other platforms; view-through inflates |
| Analytics (e.g., GA4) | Sessions and conversions on your site with a cross-channel attribution model (data-driven or last click) | Cannot see impressions without clicks; consent and blocking gaps; cross-device limits |
| Backend / CRM | Actual orders, revenue, refunds, qualified pipeline | Usually no ad exposure data |
| Incrementality / MMM | Causal effect or modeled contribution of channels | Slower, coarser, requires design and statistics |
Why platform totals exceed reality
- Overlapping credit: a customer who saw a TikTok ad, clicked a Google ad and opened an email is counted by all three.
- View-through: impressions that preceded a conversion get credit even if the view had no effect.
- Modeled conversions: platforms estimate conversions they cannot observe due to consent or browser limits. This is legitimate but adds uncertainty.
- Baseline capture: ads shown to people who would have bought anyway.
Summing platform-reported conversions and comparing with backend orders often shows a total well above 100% of real orders. That is expected.
Attribution models, briefly
- Last click — credits the final click. Simple, biased toward bottom-funnel channels like brand search and retargeting.
- Data-driven attribution (DDA) — used by default in Google Ads and GA4; distributes credit based on observed paths, but only among touchpoints it can see.
- Multi-touch (third-party tools) — similar idea across channels; increasingly limited by privacy restrictions on user-level tracking.
None of these is causal. They distribute credit; they do not tell you what would have happened without an ad.
Attribution windows and settings
Every platform lets you choose a click window and, on some platforms, a view window (for example, "7-day click, 1-day view"). Longer windows claim more conversions; shorter windows claim fewer. Choose windows that reflect your real purchase cycle, keep them consistent across reporting periods, and document them in your measurement charter. When you compare platforms, note that their default windows differ, so "conversions" rarely mean the same thing across dashboards.
The triangulation approach
Mature teams use a measurement stack:
- Platform data for daily steering (bids, creative, budgets inside a channel).
- Analytics for site behavior and cross-channel path insight.
- Backend truth for the scoreboard: orders, revenue, new customers, margin.
- Experiments (lift and geo tests) for causal answers on key channels.
- MMM for the big-picture budget split, calibrated with experiment results.
Then create a simple incrementality factor per channel from tests: if Meta reports 1,000 conversions and a lift test suggests 600 were incremental, the factor is 0.6. Apply factors to platform numbers for planning until the next test.
Worked example: a US e-commerce brand selling into the Gulf
A US-based modest-fashion brand sells to customers in the US, UK and the Gulf. Platform dashboards summed to about 1.6× actual orders (illustrative). The team:
- Set a single source of truth: the Shopify backend for orders, new-customer flag and margin.
- Used GA4 for landing-page and funnel analysis, not channel credit.
- Ran one Conversion Lift study on Meta and one geo test on Google Demand Gen; derived incrementality factors.
- Reported weekly: blended new-customer CAC (backend), channel spend, and "incrementality-adjusted" channel conversions.
Debates about "which dashboard is right" stopped because each number had a defined job.
Hands-on: a blended reporting table
-- Weekly blended metrics (BigQuery-style SQL; adapt table names)
WITH spend AS (
SELECT week, SUM(cost) AS total_spend FROM ads.daily_spend GROUP BY week
), orders AS (
SELECT DATE_TRUNC(order_date, WEEK) AS week,
COUNTIF(is_new_customer) AS new_customers,
SUM(contribution_margin) AS margin
FROM shop.orders WHERE status = 'paid' GROUP BY week
)
SELECT s.week, s.total_spend, o.new_customers,
SAFE_DIVIDE(s.total_spend, o.new_customers) AS blended_new_cac,
o.margin - s.total_spend AS margin_after_ads
FROM spend s JOIN orders o USING (week)
ORDER BY s.week DESC;Pitfalls
- Using one platform's numbers to judge another platform.
- Treating data-driven attribution as causal.
- Chasing a "perfect" multi-touch model instead of running tests.
- Changing attribution windows mid-quarter and comparing periods.
How to measure success
Leaders can answer, with one table: what did we spend, what did we get (backend), what is each channel's incrementality-adjusted contribution, and when is the next test?
Key takeaways
- Platform, analytics, backend and incrementality numbers answer different questions — assign each a job.
- Summed platform conversions usually exceed real orders due to overlap, view-through, modeling and baseline capture.
- Attribution models distribute credit; they are not causal.
- Triangulate: platforms steer, backend scores, experiments and MMM decide budgets.
- Use incrementality factors from tests to adjust platform numbers for planning.
Check your understanding
Quick questions to lock in the lesson. They don’t count towards your certificate.
Put it into practice
Write a one-page 'measurement charter' for your business: which number is the scoreboard, which numbers steer each channel, which tests you will run this quarter.
Enrol for free to save your progress
Reading is always free. Enrol to keep your place, take the final assessment and earn a verifiable certificate.