Which Marketing Channel Is Working? Why Every Platform Claims the Same Sale

Why every platform claims the same sale

Each advertising platform measures its own performance, using its own definition of a conversion, its own attribution model, and its own lookback window. Within those rules, each one is reporting honestly. The problem is that the rules were written by the party being graded, and a platform counts a sale it touched as a sale it caused.

One customer, used here only as an illustration, moves through the month as follows. On Monday she watches a video advert on a social platform. On Wednesday she searches a comparison phrase and clicks a paid search advert. The following week she sees a retargeting banner, then returns directly and buys. One customer, one order, one payment.

The social platform reports a conversion, because its advert was seen inside its view window. The search platform reports one, because its click sits inside its click window. The retargeting campaign reports one for the same reason. Three reports, three claimed sales, one actual customer, and every platform can defend its own figure in isolation.

Each platform is accurate about its own contribution and silent about everyone else's. The distortion appears the moment the reports are added together.

Across a full month the arithmetic becomes visible. Where several channels overlap on the same customers, the reported total commonly runs well above the true order count, and the reported revenue does the same. A business that allocates next quarter's budget from those dashboards is dividing money according to a set of numbers that does not reconcile with its own bank statement.

What attribution actually measures

Attribution is the process of assigning credit for a sale to the marketing that preceded it. It answers a narrow question: among the interactions this platform can see, which ones occurred before the purchase, and how the credit should be divided between them.

That is a weaker question than most owners believe they are asking. Attribution describes a recorded sequence. It does not establish that the sale would have failed to happen without the advert, and a customer who had already decided to buy, then passed an advert on the way, produces an attributed conversion and no additional revenue.

The distinction has a name worth learning, because it governs every decision in this article. Attribution is bookkeeping over observed touchpoints. Incrementality is the additional business a channel produced that would not have arrived otherwise. Only the second one justifies a budget.

Platforms do offer randomised lift studies, and those are real experiments with a control group that is held back from the advertising. The limitation is commercial. The interested party designs the test, defines the control group, and reports the result, so a platform lift study should be read as evidence and not as proof.

Attribution models and lookback windows

Two settings decide almost every reported number. The model decides how credit is divided between the interactions. The lookback window decides how far back in time the platform will reach to find an interaction to credit.

• Last-click. All credit goes to the final click before the purchase. The method is simple and reproducible, and it is systematically generous to whatever sits closest to the checkout.
• First-click. All credit goes to the first recorded interaction. It favours demand creation and ignores whatever closed the sale.
• Linear. Credit is divided equally between all recorded interactions. The approach is even-handed and rarely true, since interactions are not equally persuasive.
• Time decay. Interactions nearer the purchase receive more credit than earlier ones. This is a softer version of last click, and it carries the same bias.
• Position-based. A fixed share goes to the first and last interactions, and the remainder is split between the middle. The rule is arbitrary, although it is presented as a compromise.
• Data-driven. The platform builds a statistical model from its own observed paths and assigns fractional credit. The method is more sophisticated, and it is still calculated inside the platform being assessed, on the data that platform holds.

Four of these, first click, linear, time decay and position based, were withdrawn by Google in 2023, leaving data-driven and last click in its own products. The taxonomy is still worth knowing, because other platforms and third-party tools continue to use them.

Lookback windows work alongside the model and receive far less scrutiny. A click window of thirty days means any purchase within thirty days of a click is credited to that click, whatever happened in between. Google Ads uses a thirty-day click-through window by default, adjustable between one and ninety days. Meta withdrew its twenty-eight-day options in 2021 and now defaults to a seven-day click window and a one-day view window. A Google report and a Meta report for the same month are therefore counting over spans that differ by more than three weeks, so the two figures were never comparable.

Defaults are the practical issue, because most accounts never change them. A platform selects the model and the window that best represent its own contribution, and it is under no obligation to select the ones that describe the customer's decision most faithfully. Longer windows and more inclusive models produce larger numbers, which read as better performance.

The conversions counted without a click

View-through crediting is the least understood setting in advertising reporting. It counts a conversion when an advert was displayed to a person who later purchased, even though that person never clicked it.

There is a legitimate argument for it. Advertising that is seen and not clicked can still influence a decision, which is the premise of the broader formats covered in the inefficient channels wasting your budget. There is also an obvious problem. An advert displayed low on a page, or in a feed the person scrolled past, may never have been looked at, and it will still be counted if a purchase follows inside the window. Meta withdrew its twenty-eight-day view window in 2021 and left a one-day default, which reduces the most extreme cases without removing the mechanism.

• Check whether view-through credit is included in each report before comparing channels, because a click-only figure and a view-inclusive figure are different measurements.
• Read view-credited conversions separately from click-credited ones, so the mix is visible.
• Treat retargeting figures with particular care. Retargeting shows adverts to people who already visited, so a high attributed conversion rate is partly a description of who was targeted.
• Compare like with like. If one platform is set to a long window with view credit and another to a short click-only window, the gap in reported performance may be entirely a difference in settings.

The commercial damage of last-click credit

The largest platforms have moved away from last click as a default, but the bias it describes has not gone away. Data-driven models still assign credit only among the interactions one platform can see, and they still weight the interactions nearest the purchase most heavily. Credit lands on whatever the customer touched shortly before purchasing, which is usually the channel closest to an intention that already existed.

Branded search is the clearest case. Somebody who has decided to buy from a company frequently reaches it by typing its name into a search engine. If an advert appears above the organic listing and the customer clicks it, the sale is attributed to paid search. That campaign will report an excellent return, because it is collecting credit for demand created elsewhere and often for visits the business would have received without paying, a point that connects directly to the organic visibility discussed in why the majority of search traffic fails.

Last-click reporting gives the credit to the channel closest to the purchase, and removes funding from the channels that persuaded the customer to buy.

The consequence compounds over quarters. Budget drifts towards the channels reporting the strongest return, which are the ones closest to the purchase and the easiest to measure. The channels that create demand, including brand building, public relations, and broad reach formats, report weakly under these rules because their influence happens early and often without a click.

If those channels are cut, the reported returns remain strong for a period, because the demand they created is still in the system. Some months later the volume of people arriving ready to buy declines, and the efficient channels start reporting worse results with no change to their settings. By then the cause is several quarters behind the symptom, which is why the compounding described in building a strong brand is so often dismantled by a measurement system rather than by a decision.

Modelled conversions and the missing record

Reports present two different kinds of number in the same column, with no visual distinction between them. Some conversions were observed. Others were estimated by a statistical model that infers how many conversions probably occurred among the people the platform could not follow.

Two ordinary behaviours remove part of the record. The first is consent. Where a visitor declines tracking, the platform loses the ability to connect that person's later purchase to the advert that preceded it, and the resulting gap is filled by modelling. The second is movement between devices. A person who sees an advert on a phone during the evening and completes the purchase on a laptop the next morning has broken the chain, unless a sign-in or a similar identifier joins the two sessions.

Modelling is a reasonable response to both, and better than ignoring the gap. The issue is presentation. An estimate delivered as a whole number in a table reads as an event that happened, and nothing in the interface says which part of the figure was seen and which part was inferred.

Every platform figure therefore contains some modelled volume, and the share varies by market, by consent behaviour, and by how the site is built. The practical response is to reconcile the scale and to ignore the detail. A platform reporting twice the orders the business recorded needs investigation. A platform reporting three fewer is within normal variation.

None of this makes platform reporting useless. Platform reporting manages the inside of a campaign. It does not settle whether the campaign was worth running.

The repairs that cost nothing to start

Four habits repair most of the damage, and none of them requires software, a data project, or a budget. They can begin this week.

1. Ask every enquirer how they found you, and record the answer. One free-text question on the enquiry form, or one question asked on the call, produces a record that no platform controls. Free text works better than a list of options, because a list teaches people to pick the first plausible answer.
2. Treat the company's own sales records as the single source of truth. The order system, the invoices, and the accounts hold the only count of customers that exists. Every marketing report is a claim about that count, and the count itself is not negotiable.
3. Divide the claimed conversions by the recorded orders, and track the ratio each month. Add together the conversions every platform claims for a defined period and divide by the orders in the company's own system. A ratio of 1.7 means the dashboards are claiming seventy per cent more sales than the business made. A stable ratio means the reports stay comparable month to month even though they are inflated, and a moving ratio means something changed in the settings or in the channel mix. Applying that ratio as a discount before two platform reports are compared is the reconciliation the dashboards never perform.
4. Compare total spend against total revenue over a matched period. Add every marketing cost for a defined period, add the revenue from new customers in the same period, and compare. This single ratio survives every attribution argument, because neither number was produced by a platform.

The self-reported question deserves a defence, because analytically minded people dismiss it. It is imprecise, memory is unreliable, and people credit the last thing they remember. All true. It is also the only measurement in the business that can name a channel no tracking system can see: a recommendation from a colleague, a podcast, a conference, a printed advertisement, a conversation. Those channels are invisible to every dashboard, and self-reporting is the only place they ever appear.

Used together, the habits triangulate. The sales records say how many customers arrived, the self-reported answers say what those customers believe brought them, and the platform reports say what each channel claims. Where all three agree, the conclusion is safe. Where they disagree, the disagreement is the most useful finding in the report.

Revenue instead of conversion counts

A conversion is whatever the account was configured to call a conversion: a form submission, a newsletter sign-up, a click on a telephone number, a page view of sufficient duration, an add-to-basket. Several of these can occur for a single person in a single visit, and none of them is money.

Revenue has no such flexibility. It is recorded by the business, reconciled by the accounts, and cannot be redefined by a platform to improve a report. Moving the primary measure from conversion counts to revenue removes most of the ambiguity in one step.

• Value differs by customer, and averages hide it. Two channels producing identical enquiry volumes can produce very different revenue if one attracts larger orders.
• Enquiry quality is part of the cost. A channel that generates enquiries which cannot be converted consumes selling time as well as media budget, and that time rarely appears in any report.
• Lag varies by channel. Considered purchases close over weeks or months, so a channel judged inside a short window will always look worse than one serving people ready to buy today.
• Repeat business belongs in the assessment. Judging acquisition on the first order alone undervalues channels that bring customers who return.
• Margin outranks turnover. High revenue on discounted, low-margin orders may be worth less than a smaller channel selling at full price.

For businesses with long or complex sales cycles, the practical step is joining the enquiry record to the eventual outcome, so that each closed sale carries the origin recorded when the enquiry arrived. That connection is the substance of serious analytics and performance reporting, and it changes the monthly conversation from volume to value.

Holdout and geographic testing

Only one method establishes whether a channel added anything: stop it somewhere, keep it running elsewhere, and compare the difference in results. A controlled experiment compares actual results against the results produced where the channel was withheld. Attribution cannot do this, because it only describes sequences that were recorded.

1. Geographic testing. Select comparable regions, continue the channel in some and withhold it in others, and measure the difference in sales between the groups. Region-level testing is unaffected by consent choices and device switching, because it never needs to follow an individual person.
2. Audience holdout. Withhold the channel from a randomly selected share of the audience and compare that group's purchasing against everyone else's. The mechanism is the same, applied to people rather than places.
3. Scheduled pauses. Stop a channel entirely for a defined period, hold everything else constant, and observe what happens to total enquiries and revenue. The least precise version, and often the only one available to a smaller business.
4. Platform lift studies. The large platforms run randomised tests with a control group that is not shown the advertising. These are genuine experiments, and the platform still designs the test and reports the outcome, so the result belongs alongside the other evidence and not above it.

The honest caveats matter as much as the method. Testing requires volume, because a difference of a few sales across two regions proves nothing. It requires patience, because the test period has to cover the normal delay between first contact and purchase. It requires stability, since a price change, a seasonal peak, or a competitor's campaign in one region contaminates the comparison. It also has a cost, in the revenue deliberately forgone in the withheld group. Where volume is genuinely low, a scheduled pause with careful record keeping is still worth doing, provided the conclusion is treated as an indication and not as a proof.

Marketing mix modelling answers the same question from a different direction. It works from aggregate spend and revenue over time, so it survives consent loss and covers offline channels that no tracking code can see. The limits are real and should be stated before the work starts. It needs two to three years of history and meaningful variation in spend, it produces estimates with wide ranges instead of exact figures, and it is most useful for allocation between channels, and weak at judging any single campaign. Experiments and modelling work best together, because the experiments calibrate the model.

What honest reporting looks like

Honest reporting is a single reconciled view instead of a stack of platform screenshots. It begins with the numbers the business owns and works outwards to the claims made about them. Total marketing spend for the period is stated as one figure, including media, production, and fees. Total revenue from new customers in the same period is taken from the sales records. Enquiry volume by self-reported source is shown as recorded at the point of contact, and the platform claims sit alongside all of it, with the overlap acknowledged instead of summed into a false total.

The report then states one judgement per channel: what the evidence supports, what any test showed, and what remains unknown. That last line separates a report from a sales document. Any supplier can produce a figure, and a supplier willing to write down what the evidence does not establish is one worth listening to. Where a report does not tie back to invoiced revenue, a short list of questions usually settles the matter.

1. The attribution model and the lookback window behind these figures, and whether both were left at the defaults.
2. The proportion of the figure that is modelled rather than observed.
3. The reconciliation between these conversions and the orders recorded in the company's own system, with an explanation of the gap.
4. The expected effect on sales if the channel stopped, and the evidence behind that expectation.
5. The separation between the platform's claim and the supplier's own judgement.

Confident, specific answers indicate a supplier who has done the work. An answer that only repeats the dashboard figure indicates a supplier who has not examined it. Reporting of this kind is part of how we work.

Doing it yourself versus specialist help

A great deal of this is within reach of any business without assistance. Adding one question to an enquiry form costs nothing, and so does recording the answers in a spreadsheet. Comparing total marketing spend against revenue from new customers over a matched period requires arithmetic and discipline, and it will usually reveal more than a year of dashboard reading. Checking the attribution model and lookback window on each account is a short task, and it permanently improves how every future report is read.

Specialist help earns its fee further along. Designing a valid experiment is technical work, and two judgements decide whether the result means anything: sizing the test so the difference can be separated from normal variation, and setting the duration against the real purchase cycle. Cross-market work is harder still, since consent behaviour, channel mix, and seasonality differ between countries. Several channels across several markets is usually the point at which the internal version stops being sufficient, and the context in which paid media management and measurement should be commissioned as one piece of work.

One filter applies when choosing that help. Ask any prospective supplier how they will prove that their work added revenue. Answers involving reconciliation against the company's own records, and controlled testing, describe a measurement practice. Answers involving platform-reported return on ad spend describe a reporting habit, and the fee structures that make one answer more likely than the other are examined in the fee models that decide whose interests win.

Key takeaways

• Every platform grades its own work, so summed dashboards claim more sales than the business made.
• The attribution model and the lookback window decide most of the reported figure, and the defaults favour the platform. Google Ads counts thirty days of clicks, while Meta counts seven days of clicks and one day of views.
• Credit concentrates on the channel nearest the purchase, so the channels that created the demand lose their funding first.
• View-through credit counts adverts that were displayed but never clicked, so click-only and view-inclusive figures are different measurements.
• Consent choices and movement between devices remove part of the record, and modelling fills the gap without saying so.
• Self-reported source, sales records, and a monthly claimed-to-recorded ratio cost nothing and cannot be manipulated by a platform.

The question of which channel is working is answerable, but the answer never arrives from the channels themselves. It comes from putting their claims alongside the company's own record of customers and revenue, and from occasionally turning something off to see what happens. Businesses that do this consistently spend less and know more, while their competitors continue to move budget on the strength of numbers that were never designed to be added together.

For a reconciled view of what each channel is genuinely contributing, and an honest account of what the available evidence cannot yet establish, book a strategy call with Reachford.