Before buying attribution software, ask vendors fifteen questions across four areas: how the model works, how it is validated against real experiments, what data it needs and who owns it, and what buying and living with it actually costs. Good vendors answer with specifics and admit limits. Weak ones promise a single perfect ROAS number. Full disclosure: we sell attribution software at Admira, and every question below is one you should ask us too.
Methodology questions
- What attribution methodology do you use, and why? "Proprietary AI" is not an answer. You want to hear whether it is rule-based, Markov chains, Shapley values, regression-based, or a hybrid, and why that choice fits your business.
- How does the model handle channels without click data? Podcasts, influencers, TV, out-of-home. If the answer is "promo codes and surveys," that is honest. If the answer implies clicks capture everything, the model will quietly starve those channels.
- How do you treat view-through conversions? View-through credit is the easiest way to inflate paid social. Ask how impressions are weighted and whether you can see results with view-through stripped out.
- What happens to your model when consent rates drop or cookies disappear? Ask what percentage of journeys the vendor can actually observe today and how the model degrades. Every vendor loses data to privacy changes; the difference is whether they measure and disclose it.
Validation questions
- How do you validate your model against holdout or incrementality tests? This is the hardest and most important question. Attribution models fit observational data; only experiments observe causality. A serious vendor either runs lift tests natively or shows you how customers calibrated the model against geo holdouts. "Our model is the validation" is a red flag.
- When your numbers disagree with Meta or Google, how do I decide who is right? They will disagree, constantly. You want a vendor with a documented reconciliation approach, not one that just says "trust us over them." Nobody, including us, can claim to be automatically right.
- Do you show uncertainty, or only point estimates? A channel ROAS of 3.2 means something different if the plausible range is 2.9 to 3.5 versus 1.5 to 6. Vendors that never express confidence ranges are hiding how noisy attribution is.
- What can your model not measure? Every honest vendor has a real answer: brand halo, long B2B cycles, dark social, offline word of mouth. A vendor with no answer has not thought hard enough or will not tell you.
Data and integration questions
- Do you attribute against backend revenue or pixel conversions? Attribution based only on pixel events inherits every pixel problem, including duplicates and missing refunds. Ask whether orders, refunds, and LTV from your actual backend feed the model.
- Which integrations are native, and how fresh is the data? Get the list in writing and ask what "supported" means: a maintained API connection or a CSV template.
- Who owns the data, and can I export everything if I leave? Raw event data and model outputs should be exportable. If leaving means losing your history, you are renting your own numbers.
Commercial and operational questions
- What does implementation require, and when do numbers become trustworthy? Ask for the honest timeline in two parts: setup time, and the data accumulation period before the model stabilizes. Anyone promising reliable numbers in week one is overselling.
- How does the model handle a new channel or an unusual promotion? Big sale events and channel launches break historical patterns. Ask how long the model takes to adapt and what it reports in the meantime.
- What does pricing scale with, and what happens at renewal? Pricing tied to order volume or ad spend can double while your team stays the same size. Get the scaling formula, not just this year's quote.
- Can I speak with a customer who ran a validation test, and one who churned? Reference calls with happy customers are easy. What you learn from a churned customer or one who tested the model against a holdout is worth more than any case study.
Score the answers, don't just collect them
Fifteen questions produce a lot of talk, and the vendor who talks best is rarely the one who measures best. To compare fairly, convert each answer into a rating rather than trusting your memory of who sounded most confident in the demo. Score every vendor on the four areas on a one-to-five scale, weight validation and data ownership highest, and the shortlist tends to rank itself. The pattern to watch for is simple: strong answers get concrete and admit limits, weak answers get vague and reach for adjectives.
| What you are scoring | A strong answer sounds like | A weak answer sounds like |
|---|---|---|
| Methodology transparency | Names the model, explains the trade-off, shows where it fits your mix | "Proprietary AI that just works" |
| Validation against experiments | Runs or supports holdout and geo lift tests, shows calibration | "The model is accurate, we don't need tests" |
| Data ownership and export | Raw events and outputs are exportable, no lock-in | Evasive on export, history stays with the vendor |
| Total cost at scale | Discloses the scaling formula and renewal behavior up front | Only quotes this year's price |
Before you even book the calls, it pays to run a measurement audit on your own stack, because you cannot judge a vendor's answers about reconciliation or missing data until you know how bad your own gaps are. A vendor that asks to see your audit findings is a good sign; one that waves them away and jumps straight to the dashboard is selling polish. Weigh every answer against the specific decision you need the tool to improve, the same discipline that separates a good pick from a shiny one when you are shortlisting attribution software in the first place.
How to read the answers
You are not looking for the vendor with the most confident answers; you are looking for the one whose confidence is calibrated. Specific numbers, admitted limits, and documented validation beat polish. If a vendor, including Admira, gets defensive at questions 5 and 6, keep shopping.
FAQ
What is the biggest red flag in a vendor answer?
Certainty without validation. If a vendor claims their model is accurate but has never compared its outputs to a holdout or geo test, the accuracy claim is unfalsifiable marketing.
Do I still need incrementality testing if I buy attribution?
Yes. Attribution tells you what likely drove results day to day; experiments confirm causality and calibrate the model. The best setups treat them as one system, not competing products.
How long should a pilot run?
Long enough to cover a full purchase cycle plus one meaningful budget decision, typically 60 to 90 days for ecommerce and longer for B2B. Judge the pilot on decisions improved, not dashboards delivered.
Should I trust vendor case studies?
Treat them as existence proofs, not expected outcomes. Results depend on your data quality, mix, and cycle length. Ask instead for the validation methodology behind any headline number.
Ask Admira the same fifteen
Admira is a unified measurement platform that combines multi-touch attribution, marketing mix modeling, and built-in lift testing in one place, so the model is validated by experiments rather than by faith and every channel is credited once against a single source of truth. Our integration list and how plans scale with your spend are on the table before you sign, not hidden until renewal. Bring us these fifteen questions and pressure-test the answers on your own data: book a demo and we will walk through methodology, validation, and reconciliation against the exact channels you run.
.svg)
%20(2).png)
