iOS attribution modeling: choosing a method that fits your channel count
Apple's removal of the identifier for advertisers pushed iOS measurement into two camps: SKAdNetwork, which is deterministic but delayed, and modeled conversions from ad networks, which arrive quickly but are estimates. Neither gives a complete picture on its own. What follows is a working method for reading both, split by whether you run one paid channel or several.
Why iOS measurement is hard to read
SKAdNetwork (SKAN) changed how the whole app industry measures installs. Before it, value-optimized campaigns could be judged on deterministic attribution within a month. After it, measurement partners largely display what ad networks send them, enriched from Apple's postbacks, rather than attributing installs themselves. One alternative that has been floated is Appstack, a project in beta that aims to receive raw postbacks from Apple directly; treat that as a dated reference rather than a settled market picture.
The practical problems with SKAN are well known to anyone running iOS campaigns:
- Results take time to settle. Stable performance often does not appear until the third or fourth day of spending.
- SKAN totals never match App Store Connect downloads, which makes campaigns look like they are missing users they actually generated.
- Because postbacks can take up to 41 days to complete, teams build predictive models for long-term value and behavior, and those models are not fully accurate.
- The daily budget you set determines the data tier you reach, so smaller advertisers receive very little post-install detail.
Marketers want same-day feedback. That demand is what pushed ad networks and measurement partners toward modeled, or probabilistic, attribution.
What modeled attribution offers, and what it costs you
Probabilistic attribution reports conversions in near real time and labels some of them as modeled by the provider's technology. It removes two SKAN pain points: there is no data delay, and there are no data tiers tied to daily budget. The trade-off is that the numbers are approximations, and measurement partners still bill for each modeled attribution.
The psychological effect matters as much as the accounting. Modeled figures tend to run higher than SKAN, which can make campaigns look stronger than they are. Treat any modeled number as an estimate that needs a reality check against deterministic and blended data.
How the main ad networks model iOS conversions
The large channels share one advantage: they hold enormous volumes of first-party data, which makes accurate modeling easier. Each has packaged the capability under a different name, and the mechanics differ in ways that affect how you should read the numbers.
- Meta and AEM. Aggregated Event Measurement measures web and app events from people on iOS 14 and later. Its processing removes identifiers, adds differential privacy and aggregates data across users before delivery or measurement. AEM runs independently of SKAN, so you can use it with or without a SKAN schema. Since Meta began sending AEM data to measurement partners in October 2024, SKAN numbers for AEM campaigns often have to be pulled from raw postback CSVs or custom reports rather than default dashboards. Enabling AEM also enables probabilistic attribution with Meta by default, even if it is switched off in the app settings.
- TikTok and ADC. Advanced Dedicated Campaign is an iOS campaign type aimed at improving reporting and delivery. Unlike AEM, it sits on top of SKAN and requires a SKAN schema to be in place. Expect a modest lift in reported conversions rather than the larger gap AEM can produce. TikTok has also been developing its own SDK, which could eventually support an AEM-style solution.
- Google modeled conversions. Google documents the data points behind its modeling in detail and recommends On-Device Measurement as an additional accuracy layer. In that setup, the Google Analytics for Firebase SDK performs on-device attribution using the identity from your app's sign-in experience, and no personally identifiable information leaves the device.
- Snapchat and Advanced Conversions. Advanced Conversions follows the same general idea, with one difference: conversions can still arrive up to 36 hours late because of the obfuscation used to protect privacy. Running it with the same attribution windows as SKAN is worth testing against pure SKAN campaigns, since it can improve signal quality.
{{OWNER_AUTHOR_BYLINE}} reported, from work across more than ten apps over a year, that Google's modeled conversions were the least accurate relative to SKAN, and that AEM is the better starting point because Custom Product Pages give you an extra comparison layer. The same practitioner has shared AEM-versus-SKAN case studies using Custom Product Pages and TikTok ADC case studies publicly. Treat all of this as one operator's experience, not a benchmark.
Choosing a method: a quick decision checklist
The right approach depends mostly on how many paid channels you run, which in turn tracks your budget and company size. Work through these questions in order:
- Do you run paid campaigns on exactly one channel? Use the baseline uplift method below.
- Do you run two or more channels at the same time? Use the four-layer framework below, because a single uplift figure cannot be split reliably across channels once CPM, CTR, CPI, conversion rate and budget all move together.
- Can you keep every other marketing activity stable during the test window? If not, the uplift method will attribute the wrong gains to your campaign.
- Do you have historical conversion-rate and lifetime-value data you trust? Both methods depend on it to turn a cost per trial into a profitability judgment.
- Are you willing to revise projections against actual revenue data? If not, neither method will tell you when a campaign is attracting lower-quality users than expected.
Single-channel method: baseline uplift
With only one channel running, you can measure the campaign's effect by comparing your conversion baseline before and after launch. The difference is the uplift the campaign produced. This only holds if everything else in your marketing mix stays constant during the window; otherwise you cannot tell what drove the change.
Once you have the uplift, divide spend by the uplifted conversions to get a real cost per conversion. Then apply your historical long-term value estimate to see whether the campaign reaches break-even.
Worked example. Your baseline is 100 trials per day, or 700 per week. You run Meta with AEM for a week and the campaign reports 500 trials. Your actuals show 1,400 trials for that week. The extra 200 trials are conversions Meta could not attribute (700 baseline + 500 reported + 200 unattributed). Adding those 200 to Meta's total gives you the real cost per trial. From there, use your conversion rate and the lifetime value you can see in {{OWNER_PRODUCT_MENTIONS}} to project what a trial is worth at 30, 60 or 90 days, and decide whether to cut cost per trial, scale, or move to another channel.
A method note worth keeping: if your projections were too optimistic, the campaign is bringing in lower-quality users than expected. If they were too conservative, it is bringing in better ones. Either way, you can re-run the estimate against actual numbers. This approach has been used to scale campaigns from $20K to over $100K in monthly spend across more than five projects in two years.
Multi-channel method: the four-layer framework
When several modeling technologies bring users at once, no single uplift figure can be assigned to each channel. The four-layer framework instead tracks four views of the same conversions and uses their gaps to guide spending decisions.
The four data layers
- Blended. Total trials tracked in {{OWNER_PRODUCT_MENTIONS}} or your internal database. This is the bottom line.
- SKAN. Deterministic but delayed and incomplete. Use it to validate the other layers.
- Modeled. Conversions reported by each ad network (Meta AEM, TikTok ADC and so on). Fast, but often inflated.
- Adjusted. Modeled figures multiplied by an adjustment factor derived from the gap between modeled and SKAN data.
Step 1: establish a baseline for discrepancies. Track total trials reported by SKAN per day or week, total trials reported by ad networks, blended total trials, and organic trends from before you started advertising. Compare them over a couple of weeks to estimate the average gap between SKAN and modeled conversions.
Step 2: calculate an adjustment factor. Because modeled conversions tend to run above SKAN, build a ratio from the observed gap and apply it to any campaign using modeled conversions. The factor corrects in both directions: it pulls down inflated modeled trials and lifts underestimated ones.
Step 3: apply the factor and decide. Read the four layers together:
- SKAN shows profitability on conversion rate and LTV per trial: scale, because deterministic data supports it.
- SKAN does not support scaling, but adjusted and modeled data do: raise spend and watch blended data closely. If blended profitability falls after the increase, stop, because the adjusted and modeled figures were too inflated.
- SKAN and adjusted data do not support scaling, but modeled data does: the network is probably over-attributing and possibly taking credit for organic conversions. Check blended data; if profitability holds, scale cautiously, since rising cost per trial on the paid side can quickly reverse the picture.
- Only blended data shows profitability: organic traffic is large enough to carry the overall cost per trial. Do not scale without control, because every campaign-level view is negative. Improve creatives, setups or audiences first.
Two failures that look alike but are not
When the layers disagree, the cause matters more than the size of the gap.
- Modeled inflation. Modeled conversions run high while SKAN and adjusted figures stay flat. The network is over-counting, and scaling on modeled data alone will not hold up in blended results.
- Organic cannibalization. Modeled conversions look profitable while blended data stays healthy for reasons unrelated to paid spend. The network may be claiming conversions that would have happened anyway. Blended profitability can mask this for a while, but a rising paid cost per trial will eventually expose it.
The distinguishing test is the same in both cases: raise spend modestly and watch whether blended profitability survives the increase. If it does not, the modeled layer was not telling you the truth about incremental conversions.
Conclusion
No single iOS data source can be trusted on its own. SKAN is deterministic but slow and partial; modeled conversions are fast but approximate; blended data reflects actual revenue. The work is in cross-referencing them.
- Use SKAN to validate modeled data, not to judge campaigns alone.
- Apply adjustment factors to modeled figures before making budget decisions.
- Anchor every decision in blended revenue and conversion data from {{OWNER_PRODUCT_MENTIONS}} or your own database.
- Single-channel advertisers can rely on baseline uplift; multi-channel advertisers need the layered framework.
Because ad networks do not disclose exactly how their models estimate conversions, these methods give you a clearer direction rather than a perfect answer.
Frequently asked questions
Why does SKAN data never match App Store Connect downloads?
The two count different things. SKAN reports postbacks tied to ad-attributed installs, while App Store Connect reports all downloads. The mismatch is expected and does not mean a campaign failed to track the users it generated.
Does enabling AEM turn on probabilistic attribution in my measurement partner?
With Meta, yes. Enabling AEM enables probabilistic attribution by default, even if it is switched off in the app settings, because AEM would not function otherwise. That also means you are billed for those modeled attributions.
Can I use TikTok ADC without a SKAN schema?
No. ADC sits on top of SKAN and requires a SKAN schema to be in place. It enhances the signals SKAN already tracks rather than replacing it, so expect a modest lift in reported conversions rather than the larger gap AEM can produce.
How long can SKAN postbacks take to complete?
Up to 41 days for all postbacks, depending on the delay settings in place. That delay is why teams build predictive models for long-term value, and why modeled attribution became attractive in the first place.
What is the adjustment factor in the four-layer framework?
It is a ratio built from the observed gap between modeled conversions and SKAN conversions over a couple of weeks. Multiplying modeled figures by that ratio gives a more realistic conversion count, correcting both inflated and underestimated numbers.