Why this matters
The advanced Google Ads level gave bidding richer signals, each counted under a model and a window often left as setup chose them, in a total the business's records never match. The guide to conversion windows and attribution sets both from evidence and changes them without breaking the trend; the tutorial on reconciling Google Ads conversions with sales explains the gap line by line and turns it into a ratio. The intermediate exercise on reconciling Google Ads and Google Analytics compared two measuring systems; this one compares the platform with what the till took.
The piano retailer and the disputed month
Trivaero Pianos is an invented retailer of new and restored pianos. Digital pianos and accessories sell online; acoustic pianos sell in the showroom, mostly after a visit booked on the website. The checkout and the booking form both store the Google click identifier, and showroom sales are uploaded weekly against the booking's click under the offline conversion upload SOP. The finance director's complaint is one month: Google Ads says it sold far more than the shop did.
Trivaero Pianos: settings, lag and the disputed month
| Conversion action | Role | Attribution model | Click-through window |
|---|---|---|---|
| Online purchase (tag passes the order ID) | Primary | Data-driven | 30 days |
| Showroom booking (form) | Primary | Last click | 90 days |
| Showroom sale (uploaded weekly) | Secondary | Last click | 90 days |
Days from ad click to outcome, for records carrying a click identifier over six months:
| Days from click | Online purchases | Showroom bookings | Showroom sales |
|---|---|---|---|
| 0 to 7 | 51% | 74% | 9% |
| 8 to 30 | 30% | 23% | 38% |
| 31 to 60 | 14% | 3% | 33% |
| 61 to 90 | 4% | 0% | 13% |
| Over 90 | 1% | 0% | 7% |
March, reconciled at the start of July once every window had closed. Spend £35,000. Platform purchases by conversion time 212, value £148,400, including 18 view-through and engaged-view (£11,200) and 26 cross-device, enhanced-match or modelled that no stored identifier confirms (£17,900). Order IDs show 9 finance orders (£12,600) counted twice: the tag fired again, without the order ID, on return from the lender's page, and 7 finance applications (£9,800) declined after the tag fired. Orders with a Google Ads identifier from inside the window: 149 (£95,400), of which 6 (£4,900) were returned or cancelled. Another 21 orders (£19,600) carried an identifier older than the window.
Bookings: 96 in the platform, 81 with an identifier in the booking system; 11 platform bookings were second submissions from customers changing their slot, and 6 were cross-device or modelled.
Target ROAS 400%; platform ROAS 424%. From its margin the business needs 300% on net confirmed value.
Windows from lag, reconciliation by identifier
A window follows the business's own Conversion LagThe time between an ad interaction and the conversion it leads to, which leaves the most recent reporting days incomplete., holding the bulk of each action's delay and leaving the tail to the business's records. A model follows the action's use. Reconciliation runs on a closed month, by conversion time and by identifier, sizing each difference as structure or fault until only a ratio is left.
Task
- Set a click-through window for each action from the lag table, and say what handles each tail.
- Choose an attribution model for each action, with the reason.
- Plan the order of the changes, how the months either side are read and when targets move.
- Reconcile March's purchases to net confirmed value, labelling each line structure or fault.
- Reconcile March's showroom bookings the same way.
- Say what each side cannot see, and set the ratio the business will watch and the target it implies.
Worked answer
Windows. The purchase window is too short: nearly a fifth of purchases arrive after it closes, and the orders carrying older identifiers prove the loss. Lengthening it to hold all but the thin tail credits what the ads started. The booking window is the opposite: almost every booking arrives within a month, so a Lookback WindowThe time after an ad interaction in which a conversion can still be credited to it. several times longer only credits late bookings that owe the ads little. Showroom sales spread too far for any window; the import keeps a long one and the shop's records carry the rest. View-based settings stay as they are, reported on their own line.
Models. Purchases and bookings are primary for the same campaigns, so a data-driven purchase beside a last-click booking sums two kinds of credit. The booking moves to data-driven. The showroom sale, secondary and checked against the shop, stays on last click, which changes only when behaviour does.
Trivaero Pianos: settings after the change plan
| Conversion action | Attribution model | Click-through window | Share of its lag inside the window |
|---|---|---|---|
| Online purchase | Data-driven | 60 days | 95% |
| Showroom booking | Data-driven | 30 days | 97% |
| Showroom sale | Last click | 90 days | 93%; the tail stays in the shop's records |
The change plan. The double counting and the declined applications are tag faults, fixed first and logged as corrections, and the lender's return page is made to pass the order ID, as the checkout already does. Then one setting per date, weeks apart and clear of the autumn trade: the purchase window, then the booking window, then the booking model, each through the conversion goal change checklist with targets and budgets left alone. Each date is a break in the series: a month straddling it is provisional, and the fair comparison is a month wholly under the new setting against a similar one under the old, each read once its own window closed. Targets are reset from the first full month under the new purchase window, read once its window has closed, because a longer window lets an old target be met with nothing improved.
Trivaero Pianos: March reconciled
| Line | Purchases | Value | Kind |
|---|---|---|---|
| Google Ads purchases, by conversion time | 212 | £148,400 | |
| Less view-through and engaged-view | −18 | −£11,200 | Structure |
| Less cross-device, enhanced-match and modelled | −26 | −£17,900 | Structure |
| Less finance orders counted twice | −9 | −£12,600 | Fault |
| Less declined finance applications | −7 | −£9,800 | Fault |
| Unexplained residual | −3 | −£1,500 | Watch |
| Orders with an identifier inside the window | 149 | £95,400 | |
| Less returns and cancellations | −6 | −£4,900 | Structure |
| Net confirmed | 143 | £90,500 | |
| Not in the platform: identified orders older than the window | 21 | £19,600 | Window too short |
Bookings: 96 − 11 second submissions − 6 cross-device or modelled = 79, against 81 in the booking system, a residual of 2 the other way. Count ratio 81 ÷ 96 ≈ 0.84.
Purchase value ratio as reported: £90,500 ÷ £148,400 ≈ 0.61; with the faults removed, £90,500 ÷ £126,000 ≈ 0.72. Confirmed return: £90,500 ÷ £35,000 ≈ 259%, short of the 300% required. Platform target on the corrected ratio: 300% ÷ 0.72 ≈ 417%, so about 420% rather than 400%, once the first month under the 60-day window confirms it.
What each side cannot see. The platform cannot see returns, the lender's decisions, its own duplicates, orders after its window or showroom buyers who never booked online. The shop cannot see view-based credit, cross-device journeys, enhanced matches or Modelled ConversionsConversions a platform estimates statistically for activity it could not observe directly, such as unconsented or cross-device journeys., set aside as unconfirmable, not false. The second booking submissions are a DeduplicationRemoving double counts when one conversion is reported by more than one source. fault: a tracking fix at the form, not just a reconciling line, and set aside each month until it is fixed.
The ratio to watch. Finance watches the purchase value ratio monthly, with the booking ratio beside it. A steady ratio describes how the platform counts for Trivaero; a jump means a fault, a settings change or a new campaign type. Targets are the required business return divided by the ratio, moved in steps.
Common mistakes
- Reconciling March in April. The purchase window had not closed, so the platform figure was still growing.
- Calling every difference an error, or none. View-based and modelled credit is structure; double counting is not.
- Lengthening every window to be safe. The booking window was already too long.
- Changing windows and a model on one day. One break in the series, several causes.
- Reporting platform totals to finance as sales. Report the ratio and the trace behind it.
The checkpoint that closes this module tests the same judgements and adds the routine that keeps imported outcomes honest.
