Someone clicks your TikTok link on Monday. They don't buy. On Wednesday they Google your app name and download it from the App Store. On Friday they open the app and subscribe.
Which marketing channel gets credit for that sale?
That question is the entire reason attribution models exist. And the answer changes depending on which model you pick.
What an attribution model actually is
An attribution model is a rule for deciding which marketing touchpoint gets credit when someone buys. That's it. No mystery. Just a rule.
The problem is that most customers don't see one ad, click once, and buy immediately. They bounce around. They see a TikTok, forget about it, see an Instagram ad, Google it later, and eventually buy through an App Store search. Every step in that chain could claim credit, so you need a rule to decide who wins.
Different models give different answers. None of them are objectively correct. They're just different ways of looking at the same journey.
First-touch attribution
The first interaction gets all the credit.
In the example above, TikTok gets 100% of the credit because that was the first time the customer encountered your app. Everything after that — the Instagram ad, the Google search — counts for nothing.
When it's useful: You want to know which channels introduce new people to your app. Good for understanding top-of-funnel awareness. If you're trying to figure out where your audience first hears about you, first-touch tells you that.
When it misleads: It ignores everything that happened between discovery and purchase. Someone might see your TikTok, forget about you for three weeks, and only buy because of a well-timed email. First-touch gives TikTok all the credit and the email none.
Who uses it: Early-stage apps trying to find which discovery channels work. Brands focused on awareness over conversion.
Last-touch attribution
The last interaction before the purchase gets all the credit.
In our example, the App Store search gets 100% credit. TikTok and Instagram get nothing because they weren't the final step.
When it's useful: You want to know what closes the deal. Which channel is the tipping point that turns browsers into buyers? Last-touch tells you that.
When it misleads: It gives zero credit to the channels that introduced the customer in the first place. That TikTok video did real work — it planted the seed. Last-touch pretends it didn't exist.
Who uses it: Most small app developers, whether they realise it or not. If you're using basic link tracking without multi-touch support, you're probably running last-touch attribution by default. It's the simplest model and it's what most tools give you out of the box.
Linear attribution
Every touchpoint in the journey gets equal credit.
Three touchpoints? Each gets 33%. Five touchpoints? Each gets 20%. Simple division.
When it's useful: You want a balanced picture of the whole journey. No channel gets unfairly ignored. It's the "everyone gets a trophy" model, which sounds dismissive but is actually reasonable when you genuinely don't know which touchpoint mattered most.
When it misleads: It treats a passing glance at a banner ad the same as a detailed product review that convinced someone to buy. Not all touchpoints contribute equally, and linear attribution can't tell the difference.
Who uses it: Companies that use multiple marketing channels and want a general sense of how they all contribute. It's a decent starting point if you don't have strong opinions about which channels matter more.
Time-decay attribution
Recent touchpoints get more credit than earlier ones. The closer an interaction is to the purchase, the more weight it carries.
So in a three-step journey over two weeks, the touchpoint from yesterday might get 50% credit, the one from last week gets 30%, and the one from two weeks ago gets 20%. The exact percentages depend on the decay curve.
When it's useful: You believe that recent interactions are more influential than older ones. This is often true for impulse purchases and short buying cycles. If someone saw your ad yesterday and bought today, that ad probably matters more than a tweet they saw a month ago.
When it misleads: For longer buying cycles, early touchpoints might deserve more credit than time-decay gives them. A podcast recommendation from three months ago might be the real reason someone eventually bought — it just took them a while to get around to it.
Who uses it: E-commerce and subscription apps with buying cycles of a few days to a couple of weeks.
Position-based attribution (U-shaped)
The first and last touchpoints each get 40% credit. Everything in between splits the remaining 20%.
This model assumes that discovery and conversion are the two most important moments. The middle steps — the nurturing, the reminders — matter less.
When it's useful: You want to reward both the channel that introduced the customer and the one that closed the deal, without completely ignoring the middle. It's a pragmatic compromise.
When it misleads: The 40/40/20 split is arbitrary. There's no particular reason to weight it that way beyond it feeling roughly right. And sometimes the middle touchpoints did most of the heavy lifting.
Who uses it: B2B and SaaS companies with longer sales cycles. Less common in mobile apps, but it shows up in larger app companies with multi-channel marketing.
Data-driven attribution
An algorithm analyses your actual conversion data and assigns credit based on patterns. If customers who saw your TikTok ad are statistically more likely to convert, TikTok gets more credit — even if it wasn't the first or last touchpoint.
When it's useful: You have enough data (thousands of conversions) for the algorithm to find real patterns. At scale, this is genuinely the most accurate model because it's based on your actual customers rather than arbitrary rules.
When it misleads: It needs volume. If you're getting 50 purchases a month, there isn't enough data for statistical significance. The algorithm might find patterns that are just noise. Google Analytics 4 requires a minimum number of conversions before it'll even enable data-driven attribution.
Who uses it: Companies with significant traffic and conversion volume. Google, Meta, and most major ad platforms now default to some version of data-driven attribution for their own reporting.
Which model should you actually use?
Here's what I've seen work in practice.
If you're a small app (under 1,000 purchases a month): Use last-touch. It's the simplest and gives you the clearest signal. When someone buys through a specific tracked link, you know that link did the job. You don't have enough data for anything fancier, and the added complexity won't give you better answers.
If you're a growing app (1,000-10,000 purchases): Consider time-decay or position-based. You're starting to see multi-touch journeys and you need to understand which channels contribute to the middle of the funnel, not just the entry and exit.
If you're at scale (10,000+ purchases): Use data-driven if your analytics platform supports it. At this volume, the algorithm can find real patterns that no rule-based model would catch.
If you're just getting started: Don't overthink it. Pick last-touch, set up tracked links for each marketing channel, and start collecting data. You can change models later. Having imperfect attribution is infinitely better than having none.
The dirty secret about attribution models
No model is perfectly accurate. Every single one is a simplification of messy human behaviour. People don't follow neat funnels. They see things, forget things, get reminded by a friend, and buy three weeks later for reasons even they couldn't explain.
Attribution models are useful anyway because "roughly right" beats "completely unknown." Knowing that TikTok drives about 40% of your purchases (give or take) is much more useful than not knowing where your customers come from at all.
The biggest attribution mistake isn't picking the wrong model. It's not tracking anything.
If you're running marketing campaigns for your app and you can't tell which ones generate purchases, start there. Use LinkOwl or whatever tool fits your setup. Create a tracked link for each channel. See what comes back. Argue about attribution models later, once you have data worth arguing about.