Rook and Sable sells fountain pens from a shop in Aberdeen. In June 2026 it asked three systems how many orders its paid search had produced. Google Ads said 41. Google Analytics said 26. The shop's own order table, counting orders whose first recorded session carried the campaign tag, said 33.
All three are working correctly. Three tools counting the same month gave 41, 26 and 33, and the gap between them is not an error to be fixed, it is three different definitions of the word order. Until somebody reads those definitions, every meeting about that month is an argument between people quoting different questions at each other.
Why each number is the number it is
The ad platform counts a conversion for itself and inside its own window. Nine of Google's 41 were people who were shown an advert, did not click it, and ordered within the following month; the platform can see that because it recognises the browser, and it counts it because the advert was, on its own account, part of the reason. Remove those nine and 32 remain, which is nearly the shop's own figure.
Analytics counts sessions that told it they existed. Around a fifth of Rook and Sable's 2026 sessions never sent an event at all, because consent was declined, an extension blocked the request, or the person left before the script finished. The analytics number is a sample, not a census, and the gap is not evenly spread: it is widest among exactly the privacy-minded, older-browser, ad-blocking readers that a shop selling £200 pens attracts most.
The order table counts money. It knows nothing about cause, which is its weakness, and it can be reconciled against the bank, which is a property neither of the others has. Only the number in your own order table can be reconciled against the bank, and that single sentence decides which of the three should appear on the report that goes to whoever is paying for the advertising.
Deciding, in advance, which number decides
The move that ends the argument is administrative rather than technical. Decide which number decides, and write its definition down before the month starts. Rook and Sable's definition, agreed in July 2026, runs to two lines: an order counts against paid search if the campaign tag appears anywhere in the sessions belonging to that customer in the twenty eight days before the order, and the count comes from the shop's database rather than from any platform.
That definition is not the correct one. There is no correct one. It is consistent, it is checkable, and it will be the same definition in September, which are the only three properties that make a monthly number worth putting on a chart. Choosing between the possible definitions honestly is deciding which touch gets the credit, and it does not have a tidy ending.
A difference you cannot explain is a definition you have not read
The habit worth building is to treat every disagreement between two numbers as a question about wording rather than about accuracy. A difference you cannot explain is a definition you have not read, and the explanation is nearly always in the documentation of whichever tool is higher, because the higher number is usually the one counting something extra on purpose.
The reverse habit, which is the common one, is to pick whichever number is most flattering and stop. That is how an advertising account keeps a budget for eighteen months on the strength of a figure that includes people who never clicked anything.
The audit
What this shop actually knows about last month
Three states, because a shop cannot act on the difference between a measured zero and nobody having looked.
One month of orders
- YesHow many orders were placedFrom the shop's own database, reconciled against the bankYes
- YesWhich orders carried a campaign tagRecorded on the session, for the sessions that reportedYes
- NoWhether the wallet payment route records an orderIt did not for five months in 2025, and nothing said soNo
- NoWhich orders would have happened anywayNo tool holds this. It costs a month with something switched offNo
- Not checkedWhat the fifth of sessions that never reported didNobody looked, and by the nature of it nobody canNot checked
The last row is the reason for the third state. Filed under no, it would read as a measured absence, and it is a hole in the instrument rather than a fact about the shop.
Where this stalls: measurement set up once and never checked again
The stall in analytics is silent, which is what makes it this pillar's worst. Tracking is installed properly on a Tuesday, and then the shop changes its checkout, or moves to a new theme, or adds a cookie banner, and the order event stops firing on one of the three payment routes. Nothing breaks visibly. Numbers keep arriving. They are simply wrong from that day forward, and every comparison against a previous month is now comparing two different measuring instruments.
Rook and Sable ran for five months in 2025 with card payments recorded and one wallet payment route not recorded. The reports were internally consistent the whole time, the totals looked plausible, and a fifth of the shop's orders were invisible. The check that catches it takes ten minutes a month: total the orders in the shop's own database and compare that single figure against the analytics total. Anything other than a stable ratio means something has broken, and the ratio itself never has to be one.
What to open in 2026
Google Analytics 4 is the free floor and it replaced Universal Analytics, which stopped processing data on 1 July 2023; anything written before that date describes a different product with different definitions, which matters here more than anywhere else on this site. Google Tag Manager replaced pasting tracking snippets into a theme file, and its real value is that it keeps a version history, so the question of what changed on the day the numbers moved has an answer.
For anyone who has outgrown sampled, consent-limited data, the category to look at is a warehouse the shop owns, with the platform exports landing in it. That is a genuine step up in cost and in skill, and it is not worth it below the point where a fifth of the orders being invisible would change a decision. Say plainly which side of that line the shop is on, because most vendors in this category are selling to the other one.
Four clusters, and one of them comes first by a distance
Getting the data right before reading it is not first as a matter of taste. Everything downstream of a broken event is a well-formatted wrong answer, and the three clusters after it all assume it was done. Then attribution, which is where the definitions above are argued out properly.
Turning data into something someone acts on is the cluster that decides whether any of this leaves the spreadsheet, and its hardest instruction is that a report with nothing to decide on it should not be sent. Finding out instead of arguing is last because it is the most expensive, and it is the only one of the four that can settle a question nobody has an opinion strong enough to resolve.
Sideways, the numbers being argued about come from buying attention, and not overpaying for it and from being the result someone clicks, without paying for it, and the decisions they exist to serve were set at the decisions that come before any channel. It sits under the channel is the last decision, and most people make it first.
The question Rook and Sable now has is what to do about the eighteen months before the definition was written down. Those months are on a chart on the office wall, they were reported to the shop's accountant, and they were counted under a rule nobody wrote down and nobody can now reconstruct. Restating them honestly means saying that a year and a half of reporting was not measuring what it claimed, and that is a harder conversation than any of the arithmetic above.