Closing the platform reported conversions discrepancy against your registration file

Two weeks after the show closes, the media agency's deck reports 1,480 registrations from paid channels. The registration file, queried the same afternoon, produces 910 rows that can be traced to any paid click. The gap is 570 registrations, or 63 per cent on top of the file.

Every organiser has this argument. Most of them have it as an argument, in a meeting, with two people asserting two numbers. It is a reconciliation, and reconciliations are settled on paper with one adjustment per line, the way anybody would settle a bank statement against a ledger.

The four causes, and how to recognise each

The platform reported conversions discrepancy almost always decomposes into the same four categories.

Attribution windows that differ from yours. The platform counts a conversion if its click happened inside its lookback window. Google Ads help documentation, as it stands in 2026, puts the click through default at 30 days, adjustable to 60 or 90 for Search and Display, and the view through default at one day. If your reporting window is longer or shorter, the two systems are counting different journeys, and the window is a decision worth making from your own data rather than inheriting.

Modelled conversions. Where a user declines consent, the platform estimates conversions it could not observe. Google's consent mode modelling documentation, in the form published in 2026, states that the modelling requires consent mode or the IAB Transparency and Consent Framework to be implemented, plus a threshold of 700 ad clicks over a seven day period per country and domain grouping, and that the modelled conversions then appear in the conversions column, affecting every report that uses conversion columns. They are not rows. They are an estimate presented in the same column as observations.

View through credit. A registration claimed on an impression that nobody clicked. Whether these belong anywhere near a registration count is a policy question with its own answer, and for reconciliation purposes they simply come out on their own line.

Claims from more than one platform. Two platforms counting the same person, neither aware of the other. This is the cause that scales with how many platforms you run, and the only one that gets worse as the media plan gets more sophisticated.

What does the worksheet look like?

One column of numbers, top to bottom, each line an adjustment with a name. Here is the walk from 1,480 to 910.

Start at 1,480 platform reported registrations, summed across all paid platforms.

Subtract 168 view through only claims, where no click exists on any platform. That leaves 1,312.

Subtract 145 modelled conversions, which the platform estimated rather than observed. That leaves 1,167.

Subtract 92 conversions that fall outside the file's edition boundary, meaning people who registered for the following edition or after the reporting cut off. That leaves 1,075.

Subtract 96 registrations claimed by two platforms, counted once each. That leaves 979.

Subtract 41 conversions counted against a different conversion action, typically a form start, a brochure download or a registration page view that somebody configured as a conversion two years ago and nobody has looked at since. That leaves 938.

Subtract 28 rows that failed to match, where a click identifier exists but no registration row can be found for it. That leaves 910, which is the file.

The adjustments total 570. Every line has an owner and a fix, and four of the seven can be reduced by configuration rather than by analysis.

Which line is a problem and which line is just accounting

The lines are not equal, and treating them as one number of shame stops anybody from acting.

View through and modelled conversions are working as designed. They are the platform reporting on a basis you did not choose, and they belong in the media report with the basis named. Neither needs fixing.

The 92 conversions outside the edition boundary usually mean your conversion action does not distinguish between editions, which matters enormously for a portfolio running eight shows through one registration platform. That is a tagging fix and it is worth doing, because it silently inflates every campaign that runs across a boundary between two editions.

The 96 double claims need a precedence rule rather than a fix. Click before view, later click before earlier click if you are running last touch, unresolved where only views compete. Write the rule down and apply it identically every period.

The 41 conversions on the wrong action are the easiest win in the whole worksheet. Audit the conversion actions in every ad account once a year, check which are marked as primary, and remove anything that is not a completed registration from the primary set. Twenty minutes of work removes an entire line.

The 28 unmatched rows are the interesting ones, because they are a measurement failure rather than a definitional difference. Something broke between the click and the record. That number belongs next to the coverage rate for the whole file, since both are describing the same underlying weakness in capture.

What the match is actually joining on

The worksheet only works if the match underneath it is specified, and most disputes about the numbers turn out to be disputes about the join.

Three keys, applied in order. The click identifier captured in a hidden field on the registration form is the strongest, because it comes from the platform itself and needs no guesswork. A hashed, lowercased email address is the second, and it works whenever the person used the same address in both places. A composite of surname, company domain and registration timestamp inside a stated tolerance is the third, and it should be reported separately because it is the one that produces false matches.

Record which key produced each match. A month where 70 per cent of matches come from click identifiers is a healthy month. A month where that falls to 40 per cent and the composite key picks up the slack is a tagging failure in progress, and the total will look unchanged while the quality underneath it falls.

Which number goes to finance?

The 910.

It is a count of rows in a table that reconciles to the badge file, to the revenue line for paid registrations, and to the attendance figure that will appear in the post-show report. Somebody can open it, filter it, and count it again next year and get the same answer.

The 1,480 keeps a legitimate job. Bidding algorithms need a conversion signal within hours, they need it in the platform's own accounting, and they will optimise worse if you starve them of modelled and view through conversions in the name of purity. Feed the platform its signal, let it do its job, and stop repeating its output in documents that get read as financial statements.

The practical arrangement is two reports with two headings and one shared worksheet that ties them together. The media report opens with the platform figure and carries the worksheet immediately underneath. The management report opens with the file figure. Anybody who wants to know why the numbers differ reads seven lines and finds out, which is faster than the meeting they would otherwise have had.

Running it as a monthly routine, not an autopsy

A reconciliation done once, after the show, is a forensic exercise on a period nobody can change.

Run it monthly through the campaign, on the same seven lines, and each line becomes a control. The unmatched row count is a tagging monitor. The wrong-action count is a configuration monitor. The double claim count tells you how much overlap your platforms have this month, which is the number that decides whether a new platform is buying incremental audience or the same audience twice.

Keep the historic worksheets. Six months of them will show you which lines are stable and which are drifting, and drift is the signal worth acting on. A modelled conversion line that doubles between March and June is telling you something changed in consent rates or in traffic mix, and finding that in June is worth considerably more than finding it in November.

The worksheet also settles the recurring question of whether the agency is inflating results. Usually they are not. They are quoting the platform, the platform is counting on its own rules, and once the seven lines are visible the conversation becomes about which line to reduce rather than about whose number is honest.

Where this stops

A reconciliation explains a gap. It does not tell you which side is closer to the truth about what the advertising did.

The file is the auditable number and it is not the causal one. Every one of the 910 rows might have registered anyway, and nothing in this worksheet distinguishes a registration the campaign produced from one it merely preceded. That question needs a test, and the worksheet exists so that the test has a stable denominator to run against.

The 28 unmatched rows also hide an unknown quantity of registrations that never got as far as a claim. If a click identifier is stripped before the registration form, the registration will not appear as unmatched, it will appear as direct, and the worksheet will look cleaner than the tracking deserves. A shrinking discrepancy is therefore not automatically good news, and the coverage rate for the file has to be read alongside it.

Neither does the worksheet resolve which model should divide credit among the surviving 910, which is where the rest of the acquisition and attribution pillar applies.

Build the seven line worksheet this week for one completed month, using whichever paid platform is largest. The line that surprises you will almost certainly be the conversions counted against an action that is not a registration, and that one can be closed before the next report goes out.