What analytics tools actually do: recording, modeling and drawing the numbers

Almost everything sold as analytics tools is doing one, two or three separate jobs: recording what happened, deciding what it means, and drawing it. The three have different trade-offs, different lock-in and different ways of going wrong. A vendor strong in one is usually thin in another, and the demo is designed so you never find out which. Google Analytics 4, Mixpanel, Amplitude, Segment, Supermetrics, Looker Studio and Tableau all sit in different places on that ladder, and the label on the box rarely matches the layer the product actually owns.

What to take away

Work out which of the three layers a vendor actually owns. That single fact predicts the price shape and the failure mode.
The number in a report is produced by four settings, not by the data: identity, attribution window, time boundary, and filtering.
Presentation is the layer you look at and the layer that matters least. A charting product will not fix a collection problem.

Layer one: recording what happened

Collection is the part that decides what questions are answerable at all. Data arrives from a tag or SDK you install (Google Analytics 4, Mixpanel), a stream sent from your own servers (Segment), an API on someone else's platform (Supermetrics pulling ad and platform data), or a file somebody uploads. Each route has a different failure mode, and the routes are not interchangeable.

Two properties are worth more than any feature here. The first is what is structurally invisible. Nothing collected on your site can see a competitor's spend, a conversation that happened offline, or a person who declined tracking — a choice recorded under California's CCPA or the EU's GDPR. Cross-device behavior is invisible unless there is a login.

A tool that appears to report on these is modeling them, and modeling belongs to the next layer. The audience and creator figures sold by influencer platforms such as CreatorIQ, Traackr and HypeAuditor are built the same way, which the guide to discovery platforms sets out in detail.

The second is whether the record is replayable. If you can pull the raw events back out and load them elsewhere, a change of vendor costs you effort.

If you cannot, a change of vendor costs you your history. The new tool starts from the day you install it. Whoever owns collection owns the switching cost, which is why collection is where the serious contracts live.

Creator and campaign measurement usually runs through someone else's API — a platform's own figures, or an influencer tool's estimate of them — rather than through a tag you control. That makes the line between what is measured and what is modelled harder to see in influencer work than anywhere else, and it is the first thing to settle with a creator or an agency before anyone agrees on a number.

Layer two: deciding what it means

Between the events and the report sits a set of choices. Four of them move the numbers more than any feature comparison will.

Four settings that move numbers

Identity

What it decides: Same person rule
Typical options: Device, login, probabilistic
Downstream effect: Visitors, conversion rate
Vendor default: Inherited by all

Attribution

What it decides: Which touch credited
Typical options: Model and window
Downstream effect: Channel credit
Vendor default: Vendor's choice

Time boundary

What it decides: Which day it belongs
Typical options: Time zone, week start
Downstream effect: Monday totals
Vendor default: Vendor's choice

Filtering

What it decides: What gets removed
Typical options: Bots, internal, thresholds
Downstream effect: Row suppression
Vendor default: Changes quietly

Identity. What makes two visits the same person: a device, a login, or a probabilistic match. Mixpanel, Amplitude and Google Analytics 4 each have to answer that question, and each answers it in its own way. Everything downstream inherits the rule, including your unique visitor count and your conversion rate.

Attribution. Which touch gets the credit, and over what window. Every model of attribution is a rule somebody chose rather than a fact about the customer. Google Analytics 4, Adobe Analytics and the ad platforms each ship a default, and those defaults do not agree. The default is a decision made by the vendor, not a property of your campaign, and it is worth naming beside the number every time.

Time boundary. Which day a thing belongs to. Time zone, week start, and the exchange rate date for anything in another currency. Two tools with identical events and different boundaries will never agree on a Monday total.

Filtering. Bots, internal traffic, test orders, and thresholds that suppress rows below some size. Filtering rules differ by vendor and change without much announcement, so when a figure looks impossible, find out what is being removed before building a story on what is left.

Two tools reading the same events will disagree if any one of these four differs. That disagreement is the useful part of a comparison, not a defect in either product.

Layer three: drawing it

Dashboards, scheduled reports, client portals, and the export button. This is the layer that fills the demo, takes the largest share of the interface, and is the easiest of the three to replace. It is also the layer where a good product over a weak collection layer produces confident wrong answers faster than a spreadsheet would.

Judge it on one thing: can a person who did not build the report reproduce a number in it. If the only route back to a figure is asking the person who made the chart, the layer has failed at its actual job.

Reading which layers a vendor really owns

Vendors describe themselves generously. Their pricing and their setup requirements describe them honestly. Reading those two signals across a shortlist is the same exercise as any other vendor comparison.

Vendor signals and what they mean

Observed signal

Setup: One script only
Access: Wants ad accounts
Export: Event rows blocked
Price growth: Tracked events or volume
Price growth: Seats or dashboards
Attribution: Named and switchable
Historic figures: Change after settings edit

What it tells you

Setup: Collection, defaults modeling
Access: Platform reporting, aggregated
Export: Presentation over others' data
Price growth: Collection is the product
Price growth: Presentation is the product
Attribution: Modeling is really bought
Historic figures: Modeling at read time

What you observe	What it usually tells you
Setup is one script and nothing else	Collection, with modeling left at defaults
It wants access to your ad accounts, not your site	Platform reporting, aggregated from elsewhere
Event level rows cannot be exported	A presentation layer over somebody else's data
Price grows with tracked events or data volume	They are storing your raw record, so collection is the product
Price grows with seats or dashboards	Presentation is the product
Attribution models are named and switchable	Modeling is a real part of what you are buying
Historic figures change after a settings edit	Modeling happens at read time, so old reports are not fixed

That last row deserves a test of its own. Change one setting, then reopen a report you exported last week. Whether the past moves tells you something no feature list will.

Which layer each named tool owns

Google Analytics 4 is mostly collection, with a modeling layer bolted on and a thin presentation layer. Mixpanel and Amplitude are modeling-first: they can collect events through their SDKs, but their product is built around identity, funnels, retention and cohorts, with dashboards as presentation.

Segment owns collection and routing only: it takes a stream and sends it to destinations, where modeling and presentation happen. Supermetrics owns collection from platform APIs, pulling ad and marketing data into a warehouse, sheet or Looker Studio; it does not own modeling or presentation.

Looker Studio is presentation only and owns neither of the other two. Tableau is presentation, with calculated fields that touch modeling but not the identity, attribution window or time-boundary decisions; it does not collect.

Adobe Analytics spans collection and modeling with its own presentation, and its defaults live inside it. CreatorIQ, Traackr and HypeAuditor are modeling and presentation over borrowed API collection: they estimate and display creator and audience numbers, but they do not own the collection layer you control. Run the same three questions against whatever you are being sold.

The agency problem

Buying for one brand is a different exercise from buying for a portfolio. Three things break first.

Access is borrowed and it ends. Your setup lives inside a property the client owns, and one day it does not. Find out what survives that: the tracking plan, the exported history, the report definitions, or nothing. If the client owns the Google Analytics 4 property, you need edit access and an export routine; if they use an influencer platform such as CreatorIQ or Traackr, you need a seat that survives the contract. Ask for the data export and termination assistance terms in the contract before you start.

Templates have to travel. If a report has to be rebuilt by hand for each new account, the tool has quietly added a recurring cost to every win. Make the vendor apply a template to a fresh account during the evaluation.

Separation has to be real. A client should see their own data and nothing else, including the fact that other clients exist. Test this with two actual accounts.

Commercial shape is the last one, and it is where spiky agency volume bites. Work out which line grows when a campaign succeeds, then model the busiest month you have actually had rather than the average. The campaign side of the stack usually carries the same shape, so settle both in the same week.

What analytics cannot settle

Every honest measurement setup has a list of questions it cannot answer. Whether a campaign caused the sales, or the sales were coming anyway, which is the oldest version of the point that correlation does not imply causation. What happened offline. What a competitor did. Anything before collection started. Anything a person declined to share.

Naming those limits in the report is what makes the rest of it credible. It also stops the slow drift where a plausible figure from an audience estimate gets treated as a measurement three slides later.

The same discipline applies when the numbers come from a partner rather than your own stack, which is a standing issue for anyone reporting on borrowed access. It applies twice over when the figures arrive as a screenshot from a creator's manager, since talent representation sits between you and the account the numbers came from.

Common questions

Do we need all three layers from one vendor?

No, and the buyers with the most reliable reporting usually do not. One tool for collection with real export, and whatever presentation suits the client, is a common and durable arrangement.

Two tools show different totals. Which one is right?

Probably both, about different questions. Compare the four settings above before comparing the numbers. If the settings match and the totals still differ, you have a collection problem worth chasing.

Is raw export worth paying for if nobody would use it?

Yes. It is the only way to check a tool against anything outside itself, and it is what keeps the decision reversible.

Where should a small team start?

With a written list of the questions clients actually ask at month end. Fit is measured against that list. Everything in this guide is downstream of it.

How do we handle a client who wants one number for everything?

Give them one number and the sentence that defines it, printed beside it every month. The problem is never the single number; it is a single number whose definition drifts while the label stays the same. Write the definition once, date it, and treat a change to it as a change to the report rather than as a correction.

What is the first thing to fix on an inherited setup?

Find out what is being filtered out and by whom. Internal traffic, bot rules and suppression thresholds are set once by somebody who has left, and they quietly decide what every later report says. That takes an afternoon and it explains more anomalies than any other single check.