Fixing a Year of Inconsistent UTMs Without Losing History
You have decided on a convention. Now you are looking at twelve months of data containing five spellings of the same channel, and the question is whether you have to throw it away.
You do not, and you also cannot fix it in the place you want to fix it. Understanding which of those is which saves a lot of wasted effort.
The one fact that shapes everything
Data already collected is not rewritten when you change how you tag links.
A visit recorded last March carries the values that were on that link in March. Changing your convention today changes what future visits carry and touches nothing that has already happened.
Two consequences follow, and both are useful.
You cannot clean the history at the source. There is no rename that reaches backward. Anyone promising otherwise is describing something else.
And you cannot break the history either. Whatever you do to your tagging from now on, last year's data stays as it is. That removes most of the fear from this project.
Where the fix actually goes
At the reporting layer, meaning wherever you build the view somebody reads.
The variants stay in the underlying data, and you group them into one row at the point of reporting. Every reporting environment supports some version of this, whether as a grouping rule, a filter, a calculated dimension, a lookup table, or a manual consolidation in whatever tool builds the deck.
Which of those your platform offers, and what it is called, changes; check your own current documentation rather than any article's menu path, including this one, which deliberately does not give you one.
The important thing is the concept: you are building a translation from many recorded values to one reported value, and you are leaving the recorded values alone.
The five-step cleanup
1. Get the list of what is actually in there
Pull every distinct value of source and medium, over the whole period, with the volume against each.
Sort by volume. You will almost certainly find that a handful of values account for the great majority of traffic and a long tail accounts for very little. That distribution is the good news, because it means the cleanup is a short list rather than a hundred rules.
2. Group them, on paper first
Put the variants into groups by what they were meant to say. All the spellings of one channel in one group. Pick the canonical value for each group from your new convention.
Do this in a document before touching any tool, because the mapping is the artifact and it needs to outlive whichever tool you build it in.
3. Deal with the ambiguous ones honestly
Some values will not be assignable. A campaign name with no source, a value nobody recognizes, something built by a person who has left.
Put them in an unknown group rather than guessing. A row labeled unknown, with its volume visible, is honest. The same volume silently folded into a channel it may not belong to is a number that will mislead somebody.
4. Build the grouping in your reporting
Apply the mapping. Then check the totals: the sum of your grouped rows should equal the ungrouped total. If it does not, you have missed a variant or double-counted one, and that check is the only quality control available here.
5. Draw the dated line, and write it down
From today, the convention applies. Record the date somewhere the next person will find it, and record it in the report itself.
That line is what makes year-on-year comparison readable. A report crossing it needs a note saying that tagging changed on a named date, because otherwise a change in how traffic was labeled will be read as a change in performance, and that is the exact misreading this whole exercise exists to prevent.
What not to do
Do not delete anything. Data you remove is data you cannot recover, and messy data is more useful than absent data.
Do not rebuild the property or start a new one to get a clean slate. You would lose the history in exchange for tidiness, which is a bad trade, and the new one will drift too unless the convention and the process changed.
Do not attempt to retroactively re-tag links that are already live in old emails, old social posts and old documents. Those links will keep arriving with their original values for years, which is precisely why the grouping at the reporting layer is the durable answer rather than a temporary one.
And do not fix the labels and leave the process. The reason there were five spellings is that five people built links with no shared list. Grouping the history without fixing that produces the same problem again by next year, and this site covers the naming convention that prevents recurrence in its own article.
The maintenance, which is small
Once a quarter, pull the distinct values again and look for anything new that is not on your permitted list.
A new value appearing is a signal, not necessarily a mistake: somebody built a link outside the process, or a new channel started, or a partner tagged something. Any of those is worth knowing about within three months rather than within a year.
And add each new grouping rule as it appears, rather than in a single annual clean-up. Five minutes a quarter beats a day every twelve months, and the quarterly version catches the problem while the person who made it still remembers.
What this does and does not fix
It fixes: duplicate rows, channels that cannot be totaled, campaigns split across spellings, and reports that cannot be compared across periods.
It does not fix: whether the conversions in those rows are counting the right thing, whether attribution reached the site at all, or whether the values were meaningful in the first place. Those are separate and usually larger problems, and our article on choosing what a conversion should actually measure is where to go if the doubt is about the numbers rather than the labels.
The short version
Group at the reporting layer, never at the source. Draw a dated line, note it in the report, and keep everything.
The history is not lost. It is inconsistently labeled, and inconsistently labeled data with a translation table on top of it is a perfectly usable year.