All articles Fundamentals

Checkout Abandonment Analysis by Device and Source

Engineer reviewing floating translucent checkout funnel panels in a dark studio with electric blue lighting

A checkout abandonment analysis is only useful if it tells you where money actually left and who left it. Most reports stop at one percentage — sessions that started checkout minus orders completed — and that number is wrong in both directions. It counts people who are still deciding as lost, counts bots as shoppers, and counts a second payment attempt as a second abandonment.

The fix is not a better dashboard. It is a stricter model: four states per flow, a value that comes from the event itself, and a split by device, source, country and page. Once the model is honest, the splits tell you which step to fix and which traffic is worth fixing it for.

Here is that model, in the order you would build it.

Model each checkout flow as four states

Start from the event stream of your own checkout, not a third-party pixel. Every session that reaches your cart sits in exactly one state at any moment:

  • Started — the session fired your checkout_started event.
  • Completed — an order confirmation event carrying an order id.
  • Abandoned — started, no completion, window closed.
  • Still in window — started, no completion, window still open.

Define the window before you look at data

Choose a window that matches your buying cycle and keep it fixed for the period you compare. Twenty-four hours suits same-day purchases; a longer cycle needs a longer window. The discipline matters: when the window drifts between reports, month-over-month movement is your own definition moving, not customer behavior.

Attach only the value you observed

Value lost should come from the cart total the session actually carried, in your store’s currency. Do not multiply abandonment count by average order value — that produces a figure nobody can reconcile against the ledger. A session that abandoned with no cart value contributes a count, not revenue, and that distinction matters when you rank fixes.

Exclude traffic that cannot buy

Crawlers, preview bots, scrapers and automated agents reach checkout pages more often than most teams expect, and each one lands in “abandoned” unless you filter it. Revenue AI works from first-party events, and Prynt Analytics excludes bot sessions from its counts; if you write your own query, apply the same filter before you compute a rate. Our guide to bot detection covers the signals worth checking.

Analyst arranging four glowing frosted glass blocks on a black table to model checkout funnel states

Make device a dimension, not a guess

Split by device class first — mobile, tablet, desktop — then by browser and OS. A gap that appears only on mobile at the payment step usually points to a payment method, a keyboard, or a redirect that desktop never sees.

Then add identity. Sessions are the wrong unit when the same person opened checkout in two tabs. Grouping by visitorId counts one shopper once, across tabs and short returns, which is what device fingerprinting gives you.

Deduplicate retries and multi-tab sessions

A declined card followed by a successful one is one completed checkout, not one abandoned plus one completed. Key the funnel on a checkout id when you have one, and on visitorId plus time when you do not.

Read source and country without fooling yourself

Source splits go wrong for boring reasons: self-referrals from your own payment provider, “direct” traffic that was a stripped referrer, and last-touch credit handed to a retargeting click. Pick one attribution model — first-touch or last-touch — and label it in the report. Revenue AI supports both, plus linear, so the same cohort can be read three ways and you see which conclusion survives.

Developer holding a phone and tablet connected by thin blue lines to a translucent floating panel in a dark lab

Country splits need context before they need action. A drop at the payment step in one country can come from currency, a missing local payment method, or shipping cost revealed late. Compare step-level drop-off across countries before treating geography as an intent signal.

Split by page to find the step that leaks

Aggregate abandonment tells you there is a leak. Step-level abandonment tells you where. Instrument cart, shipping, payment, review and confirmation, then compute the drop between consecutive steps. The step with the largest observed value lost and enough volume is your first candidate.

Volume matters. A step that lost three checkouts is not a finding, however large the percentage looks. Revenue AI compares converters and non-converters with significance tests, so a difference is flagged only when the sample supports it.

Turn the split into a decision

Rank by observed value lost, then by how directly you can change the step. A mobile payment drop you can fix outranks a source-level difference you cannot control. Keep the reason attached to each opportunity — device, source, country, step — because that reason is what makes the next change testable.

Two colleagues reviewing a glowing blue funnel shape hovering above a dark meeting room table

Revenue AI puts this in one view: abandonment with the value lost, scored opportunities with reasons, and recommendations with the evidence behind them. The AI summary is off by default, so you read the numbers first.

FAQ

How long should the abandonment window be?

Long enough to cover your normal purchase cycle, short enough that most sessions resolve inside it. Choose it once, document it, and hold it constant across comparisons.

Should a failed payment count as abandoned?

Count it as abandoned, but tag the reason. A declined card and a voluntary exit are the same state and different problems, and merging them sends you to the wrong fix.

Why does the abandonment rate change when bots are excluded?

Automated sessions almost never complete an order, so they inflate both the rate and the count. Removing them leaves sessions that could have bought.

See it on your own checkout

Revenue AI computes abandonment, the value lost, and the split by device, source, country and page from your first-party events. Start a 14-day trial of Prynt Revenue AI and point it at your checkout events.

Try it free

Prynt is device intelligence with a free plan: 5,000 identifications a month, visitor IDs with confidence and bot detection — no credit card. Upgrade to Pro for VPN, proxy, tampering and behavioral signals.

Keep reading