Skip to content
Faisal Mahmud TusherDigital marketing, SEO & analytics
with Faisal Tusher

Programmatic Advertising › Troubleshooting

Floodlight Conversions Counted Twice

Why Floodlight reports more conversions than orders: a sales tag that repeats, two activities for one action, a doubled tag, long windows or shared credit.
Troubleshootingintermediate15 min readUpdated

First, read the raw activity count against the business's records

The Floodlight tags guide gives the principle for sorting any disputed count: the tally in the ad server is the source, and what Display & Video 360 or Search Ads 360 attribute is a slice of it. A count that is too high is either a raw tally that exceeds reality or a credited figure that was read as if it counted real orders. The two need different repairs, and hours spent in the container cannot find a fault that sits in a window or a bid goal.

As general practice, start with a comparison against the business's records. Take a closed period, read the activity's count in a Floodlight report from the ad server, and set it beside the business's records for the same period. The records need their own care: the method in the reconciliation of platform conversions with sales records records refunds and cancellations as a line of their own, sets aside conversions the business system cannot tie to an ad, and lines dates up by conversion time; applying the same care to a Floodlight sales activity is general practice. An order that was placed and later cancelled is a real firing and a real order, not a duplicate, so count cancelled orders on both sides of the comparison and keep them as their own line before reading the gap. What remains is the excess to explain.

Then place the excess. If one activity's own count exceeds the records, look at causes one and three in the table below, which duplicate the firing at the source. If each activity's count is close to the records and the complaint is about a sum across activities, credit, windows or a goal, causes two and four to six apply. Name which reader the excess touches, as the measurement plan does for each number, since the same excess shows up differently in a list, a strategy and a report. When the opposite is true and the activity reads low or zero, the diagnosis for an activity that reads zero is the page to use instead.

Diagnosis table

Diagnosis

SymptomCheckLikely causeFix
Raw count exceeds the orders, and test orders show a rise on every refresh of the confirmation pageThe order parameter in the request leaving the browser for a test order, and the activity's counting methodA sales tag with a random order identifier, or a confirmation page that fires the tag on every loadCarry the order identifier in the order parameter of a sales activity with transactions counting, and choose unique counting for a lead counter where a second submission is not a second lead
One action is counted at group level in two places, or two activities share a triggerThe configuration register and the activities whose tags fire on the same eventTwo activities recording the same action, from an earlier build or a migrationKeep the one in the right group type with the right counting method, move its dependents across, remove the other tag and retire it in stages
One test order produces two requests, and the count rises by more than onePreview mode's summary of tags fired for the event, and the network requests for the pageThe same tag fired twice by the site: a second trigger on the tag, or a copy written into the page beside the containerRemove the second trigger or cut the copy in the page over to the container, then retest with a real journey
The raw count matches the orders, but post-view conversions dominate and cluster on cheap placementsThe click-through and view-through split by activity, and the window set on the activityA long view-through window crediting display for conversions it did not causeSet the window as a decision, read post-click first and post-view beneath it, and test incrementality before giving post-view a value
Buying products each report a healthy count and the sum exceeds the raw countThe raw activity count beside each product's attributed count, and the engines' conversion columns summed against the Floodlight totalSeveral products each claiming the same conversion under their own rules; the general case belongs to the cross-channel diagnosisRead the engines and the buying platform as shares of the one tally and take the report-level repair from the cross-channel diagnosis
A search portfolio's conversion column reads above the Floodlight sale count for the same periodThe activities and actions included in the portfolio's conversion goalA bid goal holding both the Floodlight activity and the engine's own action for the same saleExclude one source from each goal, change it once and let the strategy relearn before judging it

Fix, verify and prevent each cause

A sales tag with a random order identifier, or a confirmation page that fires on every load

The Floodlight tags guide sets out the counting methods for counter and sales activities, and the method an activity uses dictates what goes in the order parameter. A sales tag puts the order number there, and that is what lets the ad server (deduplicating repeat loads) recognise a refreshed confirmation page as the same order. A random value there makes every refresh of the confirmation page a new order, a pattern the guide lists among its common mistakes. The same logic runs for leads: a form that reloads on submission is a second lead under standard counting.

Fix: read the order identifier from the data layer into the order parameter of a sales activity with transactions counting, and carry revenue and quantity in their own parameters rather than in a custom variable; for a lead activity, decide whether a person who submits twice counts as one lead or as two, and choose the counting method that matches, as the guide advises.

Verify: place a test order, refresh the confirmation page several times in preview mode and read the request leaving the browser in Tag Assistant: the order parameter carries the same identifier each time, and the Floodlight report type of the CM360 reporting guide, run across the test window, shows whether any refreshed load became a counted conversion.

Prevent: write the counting method and the source of the order identifier into the register beside each activity, so that a random value carried into a sales activity from a counter template is more likely to be noticed.

Duplicate activities recording the same action

An inherited configuration, or one that survived a migration, can hold a counter and a sales activity for the same purchase, or a replacement activity built beside an old one that was never retired. The cleanup tutorial for an inherited configuration names duplicates that count the same action twice as part of what an inherited configuration brings, builds a register of every activity with its counting method, windows and dependents, and compares the counts with the business's records. Where each activity counts the action once, its own tally is right and the doubling comes from the reading: as general practice, a report that sums at group level adds both activities, and a product that reads both reads the action twice.

Fix: decide which activity is the correct one, the one in the group type that suits the action with the counting method the business question needs, and keep it. Move the dependents across and retire the other activity in the order the tutorial sets out, without renaming the live tag strings of either.

Verify: the count at group level sits within the expected gap of the business's records for a closed period, and each list, column and goal that read the old activity now reads the new one.

Prevent: keep one activity for one action, named so that a report reader long afterwards can tell what fired, and hold the register as the single record of what each activity counts and who reads it.

A tag duplicated on the site

The activity can be configured perfectly and still receive two requests per order because the site sends them. In Tag Manager (GTM), a second trigger on the same tag or a copy written straight into the page beside the container's version both produce it. The guide to debugging and preview in Tag Manager is where a tag fired twice is proved, from the summary view and the network requests. The container governance guide explains why it happens: the same vendor pixel added twice under different names, and its audit asks for duplicates of the same event and of the same condition. Whichever way it arrives, the Floodlight guide ties the order identifier to repeat loads of the confirmation page and says nothing about a second tag, so treat a second tag as its own fault and prove it with the preview and network checks that follow (general practice).

Fix: walk a real journey in preview mode and read which tags fired on the confirmation event; remove the extra trigger, delete the stale tag or, where copies sit in the page outside the container, cut each over with the migration tutorial for tags written into pages, one tag at a time.

Verify: one test order produces one request, preview shows the Floodlight tag firing once on the event, and the count in the ad server rises by one.

Prevent: run the container audit's duplicate check after a release that touches the checkout or the confirmation page, and keep the version note that records which tag replaced which.

A long view-through window crediting display for conversions it did not cause

Here the raw count can be right and the credit can still be inflated. The Floodlight guide describes the windows each activity carries: a conversion is attributed to a click inside the click-through window, and to an impression inside the view-through window when no eligible click exists. It singles out a lengthy window on cheap display as the largest source of inflated counts in an ad server account, and it urges that the window be chosen for each activity rather than accepted as a default. The DV360 online sales setup makes the same point and gives the working rule for the window and for reading post-click orders ahead of post-view ones. The DV360 reporting guide is where Floodlight conversions appear split that way. If the post-view conversions cluster on placements nobody chose, the diagnosis of spend going to low-quality sites explains how the placements got there; the credit is genuine in the report and weak as evidence of influence.

Fix: apply that working rule to each activity and, before a post-view conversion carries any value, run the incrementality test that the view-through against click-through comparison recommends, set out in the incrementality testing guide.

Verify: after the change the post-view share of the count falls while the post-click count holds, and the raw count is unchanged because the windows decide credit, not whether the activity fired.

Prevent: write the windows for each activity into the measurement plan as decisions with a reason, and read post-view conversions as a separate figure in each report.

Buying products each claiming the same conversion

This is the one cause that is not a fault in the activity or the container. The general case, every platform claiming the same sale and a report that adds the claims together, belongs to the diagnosis for channel conversions that exceed sales, which carries the full fix, verification and prevention habit for it, so only what is particular to Floodlight is set out here. The Floodlight guide notes that one activity can read differently in three interfaces while the tally in the ad server remains the source, which makes each product's figure its own share of that tally under its own Attribution rules. The SA360 Floodlight and reporting guide names the search version: each engine credits the click it served itself, so the engines' conversion columns taken together exceed the Floodlight total, and its reconciliation method is to sum those columns and compare them with the Floodlight figure. The programmatic measurement hub adds that a demand-side platform has no cross-channel deduplication of its own, so its count is not netted against the other channels' counts.

Fix: nothing in the activity or the container changes for this cause; read the raw activity count as the total and each engine's or platform's figure as that product's share, and take the report-level repair from the cross-channel diagnosis.

Verify: once the causes that duplicate the firing are ruled out, the raw activity count sits within the expected gap of the business's records, and the gap between that count and the engines' summed columns is explained by their click-crediting rules and not by a second firing.

Prevent: record in the measurement plan that each engine's and the buying platform's figures are shares of one activity, so that a mismatch with the raw count is read as expected before anyone reopens the container.

A search bid goal holding both the Floodlight activity and the engine's own action

In Search Ads 360 a conversion goal is built from sources, and the bid strategies guide is direct: put a Floodlight purchase activity and the engine's own purchase action into one goal and each sale is counted twice, unless one of them is left out. The goal's conversion column reads double, a target is judged against a figure that is not real, and the tutorial for moving paid search into SA360 places the choice of activities at the step where each future portfolio is set up, with the same warning.

Fix: list the sources in each portfolio's goal, exclude one for each sale, and make the change once, deliberately, because a change to the sources is a change to the data the strategy builds on, and it has to settle again afterwards; the conversion goal change checklist sequences it.

Verify: each sale appears once in the goal's conversion column, the figure falls to the single source's count after the conversion lag has passed, and the strategy's status settles after its learning period rather than being read inside it.

Prevent: record each goal's sources in the portfolio's record with the reason, and treat any later addition of a source as a change to that record.

Prevention: a refresh test, a register row for each activity and a regular read against orders

The habits below are general practice, not platform rules, and each happens before the number reaches a report. After a change to a tag, a trigger, a template or the confirmation page, place a test order and run the refresh test: refresh the page, read the requests leaving the browser and the tags fired in preview mode, and confirm that one order made one request and one counted conversion. The tracking verification checklist is the place to run it from, with consent granted and with it refused, since the Floodlight guide points to the refused-consent path and the return visit as where tags go wrong, and a duplicate can hide on either. Keep one register row per activity with its counting method, windows, order identifier source and readers, so that a second activity for the same action shows up as a conflict before it fires.

Then close the loop on a steady rhythm. Read the raw activity count beside the business's records for each closed period, note the expected gap and investigate only when the gap moves, which is the discipline the programmatic measurement hub describes for each pair of systems. When a fix changes a count, write the date and the reason in the measurement change log and mark it as a break in the trend, so that the fall after a correction is read as a correction rather than as lost performance. Search portfolios and audience lists that learned from inflated counts will see their inputs fall at the same moment, so, as general practice, hold their targets steady and judge them from the fix onward, and change one thing at a time, as the bid strategies guide advises for any change to what a strategy learns from.

A blacksmithing tool seller traces extra orders (illustrative)

Illustrative example — not a real result

Where the extra orders came from

Brelkunda, an online seller of blacksmithing tools and horseshoes, compares one closed fortnight. Its order system holds 410 orders, cancelled ones counted and none taken by phone; the Floodlight report for the order activity shows 604 counted conversions, and the display team has been quoting a cost per order built on the larger number. A test order with the confirmation page refreshed three times adds three conversions, because the activity's tag was copied from a counter template and carries a random value in the order parameter. In preview mode the same test order shows a second request: a hard-coded copy of the tag on the old trade-customer checkout page, left behind when the container tag replaced it for everyone else, fires beside the container's tag. The team reads the excess as two causes, 68 repeats from refreshes and 126 extra firings from the second tag, which together explain the 194 conversions above the records. They put the order number in the order parameter with transactions counting, cut the hard-coded copy over to the container and run the refresh test with consent granted and refused. The next closed week shows 197 conversions against 199 orders, and the two-order gap is phone orders taken that week, which never reach the confirmation page, recorded as the expected difference. The change log marks the date, the cost per order is restated from that day, and the search portfolios that read the activity are held at their targets for the following weeks.

The programmatic troubleshooting section lists the other symptoms that start from a reading rather than a setting, and the Floodlight tags guide remains the owner of how activities, counting methods and windows are designed in the first place.

Common mistakes

  • Rescaling the history. Dividing the count by the size of the excess can hide the fault from lists, strategies and reports that read the activity, and a later release can break the factor.
  • Treating a cancelled order as a duplicate. The firing was real and the order was real; count cancelled orders on both sides before reading the gap.
  • Fixing the tag when the window is the problem. A raw count that matches the orders cannot be repaired in the container, and a changed window changes credit, not firings.
  • Adding the products' counts together. Each product's attributed figure is its own share under its own rules; because the shares overlap, they can together exceed what the activity recorded.
  • Testing only the happy path. A single test order without a refresh and without consent refused misses the places where a tag fires twice.
  • Changing a goal mid-learning. Excluding a duplicate source while a target and a budget move in the same week leaves no way to tell which change made the difference.
  • Retiring an activity before its readers have moved. A bid goal or audience list still reading it makes the retirement the riskiest change in the configuration.
  • Leaving the correction unrecorded. A fall in conversions with no entry in the change log is read as a performance problem long after the cause is gone.

Frequently asked questions

The ad server count is higher than our orders. Does that prove the tag is firing twice?

No. It proves the count and the records disagree, and the cause may sit in either. Count cancelled orders on both sides, record refunds and test orders as lines of their own, then ask whether the excess is in the activity's own firings or only in what a buying product credits to its media. Repeated firings and a duplicated tag raise one activity's own count; two activities, each counting the action once, leave each tally right and double the sum read across both; lengthy windows, credit shared between products and a bid goal holding two sources leave the raw count alone and inflate what a report or a strategy reads.

Can we divide the Floodlight count by the size of the excess and carry on?

Not safely. A correction factor assumes the cause is steady, and a refresh habit, a release that duplicated a tag and a window that credits cheap display each change on their own timetable. A factor can also hide the fault from audience lists, bid strategies and reports that read the activity. Find the cause, fix it at the source, record the date in the change log and read the trend from the fix onward instead of rescaling the history.

Does Floodlight remove duplicate firings by itself?

Only where the activity is set up to. A sales activity that carries the order identifier in the order parameter lets the ad server recognise a repeat load of the confirmation page, and a unique counter counts a browser once a day. A standard counter counts every firing, and the Floodlight guide ties the order identifier only to repeat loads of the confirmation page, so a second tag, a second activity or a second system claiming the same conversion is found with the checks on this page.

Terms used on this page

Frameworks

Sources

  1. Campaign Manager 360 Help — Google
  2. Search Ads 360 Help — Google
  3. Preview and debug containers — Google Tag Manager Help

Related