Which Attribution Model Should a Small Team Use?
One Model, One Direct Question, and Thresholds for When to Stop Modelling
Short answer
Use last non-direct click as your ledger and a how-did-you-hear-about-us question as the corrective. Below roughly 30 conversions a month, stop modelling entirely and read the self-reported answers as your primary source. Hold whichever model you pick fixed for a full quarter.
1. Which attribution model should a small team actually use?
Last non-direct click as the ledger, and a self-reported question on every form as the corrective. One model held fixed beats a better model swapped monthly, because a changed model makes every earlier number incomparable.
- Attribution model
- A rule that decides how credit for one conversion is divided across the marketing touchpoints that preceded it, which changes which channels look profitable rather than what actually happened.
- Also known as: Marketing attribution, Conversion attribution
- Full definition of Attribution model
Two reasons for that specific pair. Last non-direct click is the model your analytics tool already computes without configuration, so it costs nothing and it is consistent. And it is wrong in a known direction, over-crediting the final search or ad click, which means you can correct for it deliberately instead of guessing.
The self-reported question is the correction. Asking every new lead where they heard about you is the only method that sees channels no tracking can follow: a recommendation in a private message, a podcast mention, a slide in somebody else's conference talk. Those touches are invisible to every model on the market, including the expensive ones.
What a small team should not do is shop for a more sophisticated model. The sophistication is not the constraint. The constraint is that a few dozen conversions cannot support any split at all, which is a counting problem no model solves.
2. What does each attribution model over-credit?
Every model is wrong in a predictable direction. Last click over-credits search and retargeting, first click over-credits discovery channels, linear over-credits anything high-frequency, and data-driven models over-credit whatever your tracking happens to see.
| Model | What it over-credits | When a small team should use it |
|---|---|---|
| Last non-direct click | Paid search, brand search and retargeting, which sit closest to the decision | As the default ledger, at any volume, because it is free and consistent |
| First click | Whichever channel introduces people, usually organic search or paid social | As a second read alongside last click, once per quarter, never as the ledger |
| Linear | High-frequency channels such as email, which appear in almost every path | Rarely. It flatters whatever touches most often rather than what persuades |
| Time decay | Late touches, though less brutally than last click does | When your sales cycle runs one to three months and you want a softer late bias |
| Position based | The two ends of the path, deliberately, at the expense of the middle | When both discovery and closing matter and you can hold it fixed for a year |
| Data driven | Whatever your tracking can see, since untracked touches get no weight at all | Only above a few hundred monthly conversions, and never as your only read |
| Self-reported answers | Memorable channels, and whatever the buyer happened to think of first | Always, as the corrective on every other row in this table |
3. How do I choose based on conversion volume and sales cycle?
Volume decides whether to model at all, and cycle length decides which model. Below roughly 30 conversions a month no split is stable, and above a three-month cycle no model captures the path.
Three volume bands, with a rule for each:
- Under about 30 conversions a month: do not model. A fractional split across 30 conversions swings by whole percentage points when a single deal lands, so the model output is noise wearing a decimal point. Report raw counts by channel and read the self-reported answers as your primary source.
- Roughly 30 to 300 a month: one rules-based model, chosen by cycle length, held fixed for a full quarter. This is the band where a model earns its place, and also the band where switching models does the most damage to your ability to compare quarters.
- Above roughly 300 a month: keep the same ledger, and re-run the same report under a second model before any budget shift over ten percent. If the two models disagree about which channel to cut, you do not yet have an answer.
Then three rules based on how long your sales cycle runs:
- Under a week: last non-direct click is close enough to the truth to act on, because there are rarely enough touches for a split to matter.
- One to three months: time decay or position based, with a lookback window at least as long as your cycle. A 30-day window on a 70-day cycle silently drops the first half of every path.
- Over three months: accept that no model will be right, and shift your weight to self-reported answers plus what prospects say on calls. At that length the tracking gaps dominate whatever the model computes.
4. Why does one direct question beat a better model?
Because it is the only method that sees untracked channels. A model can only distribute credit among touches your tracking recorded, and for most small teams the largest influence never produced a recordable click.
Add one required field to your demo, trial and contact forms: where did you first hear about us, as free text or a short list with an other option. Then read the answers monthly and count them by hand. That is the whole method.
It will not reconcile with your analytics, and that is not a defect. The two sources are measuring different things: your analytics measures recorded clicks, and the answers measure remembered influence. When they disagree about a channel, the disagreement is the finding.
The competitor side of this is worth checking too. If the field you compete with is publishing and advertising heavily on a channel your model gives no credit to, your buyers are being touched there whether or not anything records it. Oppira tracks what competitors run on each channel, which turns that suspicion into something you can look at.
5. What lookback window should I set, and how often should I report?
Set the lookback to at least the length of your longest typical sales cycle, default 90 days if you do not know it. Report monthly, and never re-cut the same period under a different model.
The lookback window is the setting most teams never touch and it changes the answer more than the model choice does. A window shorter than your sales cycle amputates the beginning of every path, which systematically deletes discovery channels from the report.
Ninety days is a reasonable default for a business-to-business team with a cycle of a month or two. If you sell in a day, thirty days is plenty. Whatever you pick, write it down next to the report, because a number produced under a 30-day window and one produced under 90 are not comparable.
Three cadence rules that keep the numbers usable:
- Report monthly, not weekly. Weekly attribution at small volume is a chart of coincidences.
- Compare against the same month last year as well as last month, so a seasonal dip does not read as a channel failing.
- Change the model at most once a year, and when you do, restate the previous four quarters under the new model before drawing any conclusion.
6. How do I set this up in a week?
Five steps: fix your tagging, add the self-reported field, set the lookback to your cycle length, write down the model you chose, then read both sources side by side once a month.
None of this needs a new tool. It needs a decision written down and a form field added.
- Fix the tagging before anything else. Agree one naming convention for campaign tags and apply it everywhere, including email and partner links. Inconsistent tags produce channels that appear twice under different names, and no model can recover from that.Keep the convention in the same document as the model choice. A convention nobody can find gets reinvented every quarter.
- Add the self-reported field to every form. One question on demo, trial and contact forms asking where the person first heard about you. Free text plus a short option list works best, and making the field required is what gets it answered at all.
- Set the lookback window to your cycle length. Find the typical time from first touch to closed deal in your own records, round up, and set the reporting lookback to at least that. Default to 90 days if your records cannot tell you.
- Write the model choice down where the report lives. One line naming the model, the lookback window and the date you chose them. This single line prevents the most common attribution failure, which is comparing two numbers computed under different rules.
- Read both sources together, monthly. Put the model output and the counted self-reported answers side by side once a month. Where they agree, act. Where they disagree, treat the gap as the question worth investigating rather than as an error to reconcile.
7. Which attribution mistakes cost the most?
Switching models mid-year, running a lookback shorter than the sales cycle, and cutting a channel on a single month of small-volume data. Each one produces a confident number that points the wrong way.
Four failures, in descending order of what they cost:
- Cutting a channel on one month of data at low volume. At 40 conversions a month, a channel can look dead for four weeks purely by chance. Require two consecutive periods and a year-on-year comparison before cutting spend.
- Switching models to find a flattering answer. Every switch invalidates your history, and a team that switches twice can no longer say whether anything improved.
- A lookback window shorter than the sales cycle. This is invisible and systematic: it removes early touches from every path, which makes discovery channels look worthless.
- Treating a modelled number as a measurement. Credit assigned by a rule is an accounting convention, not an observation, and the deck should say which rule produced it.
The unifying rule is that attribution is a comparison instrument rather than a measuring one. It is useful for asking whether this channel improved against itself, and unreliable for asking what this channel is truly worth.
Key Takeaways
One model held fixed beats a better model swapped
Last non-direct click plus a self-reported question is the right default, because consistency is what makes any comparison possible.
Under 30 conversions a month, stop modelling
A fractional split across 30 conversions moves by whole percentage points when one deal lands. Report raw counts and read the self-reported answers as primary.
Every model is wrong in a known direction
Last click over-credits search and retargeting, linear over-credits high-frequency channels, and data-driven models over-credit whatever your tracking can see.
The lookback window matters more than the model
A window shorter than your sales cycle deletes the start of every path, which systematically makes discovery channels look worthless. Default to 90 days.
One direct question sees what no model can
Asking every lead where they first heard about you catches private shares, word of mouth and mentions that never produced a recordable click.
Never cut a channel on one small-volume month
Require two consecutive periods plus a year-on-year comparison before moving budget, because at low volume a month of noise looks exactly like a decline.
Frequently Asked Questions
Explore More
Related analyses, benchmarks, and industry insights
Related Guides
Glossary Terms
Turn reading into a reaction
Oppira watches your competitors, keeps your playbook current, and drafts the response. Start free and see your market clearly by tomorrow.