Tracking did not die on iOS; it got noisier. This hub covers what ATT changed, modelled conversions, CAPI, aggregated events, why platform ROAS is not Shopify, and how to test creative when you cannot see every purchase.
01
What Changed
1.What actually changed with iOS tracking?
Apple's App Tracking Transparency prompt let people refuse the identity join that made pixel-perfect user matching easy. A large share of iOS traffic still buys; a smaller share of those buys are observed, click-for-click, inside Meta or TikTok. Platforms responded with modelling, aggregated event limits, and a push to server-side events (CAPI). Your Ads Manager did not become fiction overnight, and it did not stay 2019-accurate. Practically: expect gaps versus Shopify, expect view-through and modelled numbers, and stop treating last-click as a morality test. Creative still matters. The scoreboard got blurrier.
#What Changed
2.Did ATT mean paid social stopped working?
No. It meant optimisation and reporting got harder, especially for small, high-consideration purchases and for teams that only trusted user-level pixel stories. Plenty of ecommerce still scales on Meta and TikTok; they do it with CAPI, broader targeting, modelled reporting, and creative volume instead of 50 tiny interest stacks. If performance fell off a cliff the week you “noticed iOS”, check the offer, the page, and creative fatigue (frequency 2.5–3.5) before you blame the prompt. ATT is a real constraint. It is also a popular alibi.
#What Changed
3.What should I stop doing because of iOS?
Stop building your whole account on microscopic retargeting pools and 20 exclusion stacks that need perfect identity. Stop judging a new hook on six hours of 1-day click purchases. Stop expecting platform ROAS to equal Shopify last-click. Stop delaying CAPI because the pixel “still shows numbers”. The replacement habits: broader prospecting, a 6-creative test grid, 48–72 hour / ~1,000-impression diagnostics on hook and CTR, blended revenue as the business score, and server-side events as hygiene. Complexity was a tracking luxury. Most accounts cannot afford it now.
#What Changed
02
Modelled Conversions
4.What is a modelled conversion?
A conversion the platform infers happened, using the events it can see plus patterns, when it cannot match the person end to end. It is not a random number generator, and it is not a receipt. Modelled conversions exist so optimisation does not go blind on iOS; they also mean a creative's CPA can include inferred results. Rank ads inside one platform, one window, one campaign type. Then sanity-check with blended Shopify (or your store) over the week. If you treat modelled CPA as a court-admissible fact, you will overfit. If you ignore it entirely, you will under-invest in ads that actually work.
#Modelled Conversions
5.Why do modelled numbers move after the fact?
Because models update as more events arrive and as the system revises earlier guesses. A Tuesday ROAS that looks heroic on Wednesday morning can settle by Friday. That is uncomfortable for daily creative kill meetings. Use a lag: diagnose hook and outbound CTR on the 48–72 hour clock, and treat same-day CPA as weather. Weekly blended is for whether the business worked. Chasing intra-day modelled CPA is how you fire a good hook and keep a noisy one. The model is doing its job. Your readout cadence has to change with it.
#Modelled Conversions
6.Can I turn modelled conversions off to see “real” CPA?
You can often view more conservative columns or shorter click windows, and you should for ranking new UGC. You cannot restore pre-ATT identity by toggling a report. Observed-only views will look worse and will starve you of signal on iOS-heavy accounts. Use a strict click window to rank creatives, and a business blended number to rank the channel. Hunting for a dashboard with no modelling is nostalgia. The honest pair is “strict for creative, blended for finance”. Anything else is shopping for a kinder column.
#Modelled Conversions#Platform vs Shopify
03
CAPI & Server-Side
7.What is CAPI, and what does it actually do?
Conversions API (and TikTok's server-side equivalent) sends events from your server — purchase, checkout, lead — so the platform is not relying only on a browser pixel that iOS and browsers can block. It improves match quality when you send the right identifiers (hashed email, phone, order id) and when the pixel and server events are deduplicated. It does not reconstruct the old tracking graph, and it does not fix a bad offer. Think of CAPI as hygiene and fuel for modelling, not as a time machine. If you do not have it, that is the first practical setup job, before another avatar test.
#CAPI & Server-Side
8.Does CAPI double-count purchases?
It can, if pixel and server events are not deduplicated with a shared event id. That inflation looks like a creative miracle. Set event_id (or the platform's equivalent) so the same order is one conversion. Then verify with test events and a known order. If Ads Manager purchases wildly exceed Shopify over a quiet week, assume a counting bug before you assume a breakout hook. Dedup is not optional. Until it is done, do not run a sensitive creative test you intend to scale — you will scale a reporting error.
#CAPI & Server-Side
9.Is a Shopify CAPI app enough?
It is often a good start and not always a complete one. Apps vary in which events they send, how they hash identifiers, how they handle subscriptions, and whether they dedup cleanly with the browser pixel. After you connect one, still verify: test a purchase, check event match quality, confirm purchase value, and compare a week's platform count to Shopify without demanding a perfect match. If the app cannot send the events you optimise for, you do not have CAPI, you have a logo in the integrations tab. Server-side is a system. Treat the app as a component.
#CAPI & Server-Side#Practical Setup
04
Aggregated Events
10.What are aggregated events, and why do they limit me?
On Meta, Aggregated Event Measurement capped how many conversion events you could usefully send from iOS web in a domain, with a priority order. The practical effect is that you cannot treat eight equally important events as if it were 2018. Pick a purchase-shaped primary event, keep the funnel events you truly need, and do not invent a custom event zoo. If your domain's event setup is messy, creative tests inherit that mess: the system optimises to the wrong thing, then you fire the video. Event choice is a tracking decision that feels like a media decision.
#Aggregated Events
11.Which event should I optimise creative tests on?
The event closest to money that you can still generate enough of — usually purchase. If volume is too low, a higher-funnel event is a labelled compromise, not a secret better KPI. Do not optimise the test on view content and scale on purchase, then argue the creatives “stopped working”. Put the same event in test and scale. Value optimisation (highest-value purchases) is a different job from counting orders; do not switch it mid-grid. Write the event name on the brief. Ambiguous events make honest kill rules impossible.
#Aggregated Events#Testing Under Noise
12.Do I still need the browser pixel if CAPI is on?
Usually yes. The pixel still catches some browser events, helps with matching, and covers setups your server might miss (certain landing-page views, some in-app browsers). CAPI plus pixel with dedup is the standard, not CAPI instead of pixel. Turning the pixel off to “simplify iOS” is a good way to lose observed events you were still getting. The opposite is also a mistake: pixel-only because CAPI felt like engineering. You want both, deduped, verified. Simple to say. Still the right shape.
#Aggregated Events#CAPI & Server-Side
05
Platform vs Shopify
13.Why is Meta ROAS higher than Shopify ROAS?
Because Meta may count modelled conversions, view-throughs, and multi-touch credit that last-click Shopify will never show, and because windows differ. Shopify may also miss paid clicks that arrived without parameters. A gap is normal. A gap that widens every week without a blended revenue increase is a warning. Use Meta to rank creatives; use Shopify plus total ad spend to rank the business. Neither number is the enemy. The enemy is picking whichever one makes this week's meeting easier. See /knowledge-base/en/faq/cpa-and-roas-faq for the break-even maths on top of this gap.
#Platform vs Shopify
14.Should I trust UTM last-click in Shopify more than the platform?
As a conservative floor, it is useful. As the only truth, it under-credits video prospecting, especially on iOS. People see an ad, bounce, come back via email, branded search, or a naked visit. Last-click will call that “organic” and you will starve the creative that started the path. Watch both: last-click so you do not get drunk on modelled ROAS, blended so you do not fire the top of funnel. If last-click is your only religion, you will over-invest in branded and retargeting ads that look efficient and under-invest in new hooks.
#Platform vs Shopify
15.Can a creative “win” on Meta and lose in Shopify?
Yes. A clickbait hook can harvest modelled or view-through credit and send junk sessions that never become Shopify orders, or bounce so fast they barely session. That is a click-quality story as much as a tracking story. Check outbound CTR versus landing-page views, then Shopify conversion rate on that traffic if you can segment it. High platform ROAS with dead sessions is not an iOS mystery — it is an ad that does not match the page. See /knowledge-base/en/faq/ctr-and-click-quality-faq. Fix the match; do not rebuild CAPI first.
#Platform vs Shopify
16.How should agencies report iOS-era results to clients?
Show the pair every time: platform performance with the window named, and blended store revenue over spend. Name modelling, name CAPI status, name the creative grid you ran. Do not send a single Meta ROAS screenshot as the month. Clients who were promised pre-ATT precision will be disappointed anyway; surprise is worse. If tracking is not verified, say that before you debate the avatar. Honesty here is a retention strategy. A pretty modelled ROAS with a falling bank balance is how agency relationships end, and no render quality will save that conversation.
#Platform vs Shopify#Testing Under Noise
06
Testing Under Noise
17.How do I test AI UGC when conversion data is noisy?
Lean on metrics that are less modelled: hook rate (3-second views ÷ impressions, 30%+ target, weak <20%), hold (12–20%), outbound CTR (1%+ cold). Those tell you whether the file is doing creative work before the purchase model argues. Use the 6-creative grid (3 hooks × 2 avatars) so you are comparing siblings in the same noise. Read diagnostics at 48–72 hours or ~1,000 impressions; wait longer to rank CPA. Do not add more targeting complexity to “get cleaner data”. Cleaner structure, fewer variants, better events — that is the iOS testing kit.
#Testing Under Noise
18.Should I run lift tests for every new avatar?
No. Conversion-lift or geo-lift experiments need spend, time, and a stable offer. They are for channel-level questions, not for whether avatar A or B should say line three. Using a lift test as a creative QA tool is how you ship 2 ads a quarter. Use the cheap grid for creative, and reserve lift for when finance will not believe blended plus platform together. We cannot run a lift test inside the renderer. We can get you the 6 files the same afternoon so the ordinary test is actually powered.
#Testing Under Noise
19.Is a longer attribution window the fix for noisy tests?
It is a trade. Longer windows capture more delayed purchases and also credit more coincidence, especially on view-through. For new UGC ranking, a shorter click window is usually the more honest creative metric. For arguing that video prospecting exists, a longer or blended view may be closer to reality. Do not widen the window mid-test because you dislike the CPA. Pick it, print it, keep it for the whole grid. Window shopping is the quiet cousin of modelled-CPA shopping.
#Testing Under Noise#Platform vs Shopify
20.Does creative fatigue look like an iOS tracking problem?
Often, in the meeting. CPA rises, someone says “signal got worse”. Check frequency (typical cold fatigue 2.5–3.5) and hook rate first. If hook slid from 30%+ toward <20% on the same file, that is fatigue or a dead open, not ATT. If hook and CTR are stable and blended revenue fell across every channel, you may have a tracking or demand problem. Mislabeling fatigue as iOS delays the refresh you actually need. See /knowledge-base/en/faq/creative-fatigue-and-refresh-faq. Diagnosis order: creative health, then tracking, then market.
#Testing Under Noise
21.What is the one tracking mistake that ruins a 6-ad test?
Optimising to a broken or empty event. The auction then hunts for the wrong people, every variant looks expensive or oddly cheap, and you fire AI UGC as a category. Second place: changing the pixel, the window, or the CAPI app on day two of the test. Third: reading 1-day modelled CPA on 200 impressions. Verify events, freeze the apparatus, run 3 hooks × 2 avatars, read hook and CTR at 48–72 hours or ~1,000 impressions, then let CPA mature. That sequence is boring. It is also how you learn anything in public after iOS.
#Testing Under Noise#Practical Setup
07
Practical Setup
22.What is the practical tracking setup before I scale AI UGC?
Pixel installed on the destination, CAPI (or equivalent) sending purchases with value and currency, dedup via event id, a sensible primary event, UTMs on the ads, and a weekly blended report next to Ads Manager. Verify with a real test purchase. Then, and only then, run the 3 × 2 grid. Skipping setup to “just see if the avatars work” produces a maybe that haunts the account. We generate the ads either way; we cannot repair a missing purchase event from the editor. Tracking is not a premium add-on. It is the scoreboard.
#Practical Setup
23.Do I need a server engineer to set up CAPI?
Not always. Many Shopify and common ecommerce stacks can get a working CAPI path from an app or a gateway. You still need someone accountable to verify events — that can be a media buyer who will actually place a test order. Custom stacks, subscriptions, and multiple domains usually do need engineering. Do not wait for a perfect data pipeline to ship the first test; do wait for purchase events that fire. Perfect is the enemy. Silent is worse. If your current setup cannot confirm a $1 test order, you are not ready to interpret CPA.
#Practical Setup
24.What identifiers should I send with server events?
The ones you actually have and are allowed to send: hashed email, hashed phone, order id, IP and user agent when appropriate, and a stable event id for dedup. More good identifiers generally mean better match quality; invented identifiers mean junk. Follow the platform's hashing rules. Do not dump extra personal data “for the model”. Privacy law still applies — see /knowledge-base/en/faq/gdpr-and-data-handling-faq — and over-collection is not a performance strategy. Match quality is a health metric. Treat a sudden drop as an incident, not as a creative dip.
#Practical Setup#CAPI & Server-Side
25.Should I use a third-party tracker as the source of truth?
You can use one as another lens, not as a priest. Third-party tools have their own modelling, session rules and iOS gaps. Adding a fourth ROAS number to the argument rarely produces a winner; it produces a longer meeting. If a tracker helps you pass UTMs cleanly and QA events, good. If it exists to adjudicate Meta versus Shopify versus TikTok to the cent, it will fail. Keep the pair: platform for creative ranking, store blended for business. A third system has to justify itself with operational help, not with a promise of perfect truth.
#Practical Setup#Platform vs Shopify
26.Does Klip Kanvas fix attribution?
No. Klip Kanvas does not sit in your pixel path, replace CAPI, or certify that a given purchase will show up in Ads Manager. What we do is make it cheap to run the creative tests that still work under noise — new hooks, new faces, enough volume to see directional hook and CTR even when CPA is modelled. Anyone bundling “AI avatars plus guaranteed attribution” is selling two products, one of which is imaginary. Get tracking to a verified baseline, then generate. The order matters, and we will not pretend otherwise.
#Practical Setup
Ready to put this into practice?
Create your first AI UGC video ad in minutes — no filming, no actors, no editing.
