Analytics & Attribution Playbook

In B2B, the CRM decides what counts. Carry every lead’s source into it, send opportunities back to the ad platforms, read what the software credits next to what buyers say they heard, and let a holdout test settle the channel where the two disagree. That answers the board’s question, “what drove pipeline?”, without an attribution platform.

B2B measurement breaks in three places. A deal takes weeks or months, so the ad platforms see the click and never the revenue. Buyers learn about you in places no tag can see, such as a colleague’s recommendation or a podcast. And there are too few conversions for any model to be precise. GA4 sees sessions and events; the CRM sees opportunities and revenue; neither sees the conversation. The seven steps below join the first two and ask buyers about the third.

This edition is for companies that sell through a conversation: a demo, a trial with a sales follow-up, or a proposal. If you sell to consumers, read the B2C edition instead. Nothing here needs a customer data platform or a media mix model to start; the report on the modern MarTech stack shows where those sit once you outgrow the basics.

Every later argument about attribution is really an argument about what counts, so settle two things first. Which conversion does marketing answer for? For a sales-led company the honest answer is a qualified opportunity, not a form fill, because form fills include students, competitors and job seekers. And which system holds the truth about pipeline and revenue? The CRM, because it is the only one that knows which leads became deals.

The other systems keep their own jobs (Figure 1). The ad platforms are right about what you spent. GA4 is right about what happened on your website. Neither knows which leads became revenue unless the CRM tells it, which is Step 4’s job.

Then write down the stages a lead moves through, with one entry rule each, and give them names GA4 also uses. Its recommended lead events cover the path: generate_lead when someone submits a form, working_lead when a representative makes contact, qualify_lead and disqualify_lead, and close_convert_lead or close_unconvert_lead at the end^[1]. The Lead Management Playbook covers the entry rules.

Before adding anything, find out what is already there. List every event your website sends, which ones are marked as key events (GA4’s name for the events you count as conversions), what your forms pass to the CRM, and how your links are tagged. The usual faults are too many key events, events that fire twice, and data that should not be there at all.

Keep the key events few. A standard GA4 property allows 30^[2], but a key event should be something you would pay to get more of: a demo request, a trial start, a webinar registration. A scroll or a file download counted as a conversion makes every channel look better than it is, and if you build Google Ads conversions from key events, it changes what the bidding chases. Name events in one lowercase style, within GA4’s limits of 40 characters per event name and 25 parameters per event^[3].

Look for personal data too. Google’s policy is that nothing be passed to Google Analytics that Google could recognize as personally identifiable, and an email address in a thank-you page URL is the common leak; GA4’s data redaction setting can strip email addresses and named query parameters from web data^[4].

Then write it down. A tracking plan lists each event, what triggers it, its parameters, its owner and whether it is a key event, so the next person adds to it instead of guessing. The Tracking Plan and UTM Builder is laid out that way and checks the GA4 limits for you.

UTM parameters tell GA4 where a visit came from when no ad platform tags it for you. Put utm_source, utm_medium and utm_campaign on every link you control. Google warns that missing parameters show up as “(not set)” and that values are case sensitive, so google and Google are two sources^[6]. Google Ads clicks are the exception: with auto-tagging on, each carries a Google Click ID, the gclid, instead^[7].

The medium decides the channel. GA4 sorts traffic into channels with fixed rules, and a visit counts as Paid Social only when the source is a social site and the medium matches its paid pattern: cpc, ppc, retargeting, anything containing “cp” or anything starting with “paid”^[8]. A medium of social-paid fails that pattern, so paid LinkedIn ads tagged that way land in Organic Social. Figure 3 reads one link part by part.

GA4 knows which campaign brought a visit, but not which visit became a deal; the CRM has to hold that. Add hidden fields to every form that capture the UTM values, the gclid and GA4’s client ID, and save them on the lead record twice: the first time you see the person, and at the conversion. Without them, the CRM’s lead source is whatever a salesperson picked from a list.

An ad platform optimizes toward whatever you tell it is a conversion. Tell it a form fill and it finds people who fill in forms. Tell it which clicks became opportunities and it can look for more of those. That is what offline conversion imports do: Google Ads creates a Google Click ID for every ad click, you store it with the lead, and when the lead becomes an opportunity or a customer you send the ID back with the outcome^[7].

Google now recommends starting with enhanced conversions for leads, which also matches the lead on hashed details such as the email address^[9]. From June 2026 Google combines enhanced conversions for web and for leads into one on/off setting, and from June 15, 2026 offline conversion and lead uploads made through its API go to the Data Manager API instead of the Google Ads API^[10]. For LinkedIn campaigns, LinkedIn’s Conversions API does the same job, taking online and offline conversions from your own systems^[11].

Mind the window. Google’s guide to importing conversions from Salesforce says the time from click to conversion must be under 90 days, and later conversions are not imported^[12]. If many of your deals close later than that, make an earlier stage, such as the opportunity, the conversion the bidding uses, and send wins as a secondary conversion for reporting.

An attribution model is a rule for splitting credit between the clicks before a conversion. GA4 now offers three: data-driven, which learns the split from converting and non-converting paths in your property; paid and organic last click, which gives all the credit to the last channel clicked other than direct; and Google paid channels last click. First click, linear, time decay and position-based were removed in November 2023^[13]. You can still apply them in a spreadsheet, or in a CRM that logs every touch, and it is worth doing once, because the same deal gets a different answer under each rule (Figure 5).

None of the rules can credit what never produced a click. A buyer who heard your CEO on a podcast, or was told about you by a peer, arrives by typing your name, and every model credits whatever came last. So ask. Put an open question, “How did you hear about us?”, on the demo form as free text, not a drop-down, because a list can only return the channels you thought of. Code the answers into channels each month and set them beside the software’s split. The Self-Reported Attribution Coder does the coding and shows each share with an honest range.

Read the two together. Software attribution is precise about clicks and blind to conversations; self-reported attribution sees the conversations but depends on what buyers remember. Where they agree, act. Where they disagree, as in Figure 6, you have found the channel to test in Step 6.

Attribution of either kind shows who was touched before buying, not what would have happened without the ad. Only an experiment answers that, and the evidence for running one is strong. In field experiments at eBay, brand-keyword search ads showed no measurable short-term benefit, and the returns from paid search were a fraction of what non-experimental estimates suggested^[14]. Across 15 advertising experiments at Facebook, the observational methods advertisers commonly use often failed to recover the effect the randomized tests measured, usually by overstating it^[15].

For a B2B company the simplest design is a holdout: switch a channel off for part of the market, keep it on for the rest, and compare. Split by geography when the channel can be targeted that way, as search can, and match the halves on past demand. Three rules keep it honest:

1. Write the decision before the test. Say what result would cut the channel and what would keep it.
2. Check that the test can see the effect. At B2B volumes, demo requests are often too few to show anything but a large change in a few weeks, so measure something more frequent alongside them, such as visits or qualified form fills.
3. Measure a stage that falls inside the test. A four-week pause cannot show its effect on deals that take ten weeks to close.

The incrementality testing guide covers the designs in more depth, and the A/B Test Significance Calculator checks whether a difference is larger than chance.

Measure what a pilot changed

A customer pilot is the same problem in miniature: its result is only believed if it answers “compared with what?”. Five rules make it hold up.

1. One primary metric, chosen by the buyer, written down before the start. Pick a number they already record, so no one argues about the data.
2. A baseline from the same people. At least four weeks of the metric before the pilot, from the same team or site.
3. A comparison group where you can get one. A similar team, shift or site that keeps working the old way over the same weeks. It absorbs seasonal swings and anything else that changed at the same time.
4. Measure the change against the comparison group’s change. Effect = (pilot group after − before) − (comparison group after − before).
5. Fix the end date and the decision. State in advance what result means “roll out”, “extend” or “stop”.

Before any customer pilot starts, put the metric, baseline, comparison group and decision date on one page.

For a hospital evaluation, Step 8 of the MedTech Marketing Playbook applies the same method to a surgical unit, with a comparison unit and a decision date. The Pilot-to-Contract Planner works the pilot back from the buyer’s decision date.

A dashboard earns its place when every number on it can change a decision. I keep it to eight on one screen: what you spent; what it produced at each stage (demo requests, opportunities, pipeline created and new annual recurring revenue won); what it cost per opportunity and per new customer; and one data-health number, the share of opportunities with a known source, because every other number depends on it. Pipeline and revenue come from the CRM and spend from the ad platforms; no platform’s own conversion count appears.

Show each count with its normal range, not just a color. Counts this small move a lot by chance. My rule of thumb: a count swings by about two square roots of its usual level, so a quarter that usually brings 120 demo requests can bring anywhere from about 98 to 142 without anything changing. A number outside its range is worth a question; one inside it is not news, however green it looks.

Then give the numbers a rhythm. A 30-minute weekly review checks data health and new pipeline by source. A monthly channel review reads cost per opportunity by channel against the self-reported mix and the latest test, and records each stop, start or scale decision, with its reason, in a change log, along with any change to an event, a definition or a model. One named owner makes those changes. I move no more than 20% of a channel’s budget in a month unless a test says more; that is my rule of thumb, not a published one. Each quarter the same eight numbers go to the board with what was learned and what changes next. The Marketing KPI Report builds that page from your numbers, and the Marketing KPI Board Report keeps it quarter after quarter.

Track cycle time by stage

Behind the pipeline numbers sits one supporting measure worth setting up once in the CRM. Total days to close tells you a deal is slow; days in each stage tells you why.

1. Stamp the date each deal enters each stage, automatically, not by hand. Many CRMs can stamp the date a deal enters each stage; check it is on and that reps cannot edit it.
2. Report the median time per stage, not the average. One deal stuck for a year distorts an average.
3. Group deals by the quarter they started, so a change in your process shows as a change between groups.
4. Watch open deals as well as closed ones. Flag any deal that has spent more than one and a half times the stage median (my rule of thumb) in its current stage, and review the list weekly.
5. Fix the slowest stage first, and name what it waits on: a person, a meeting, a document.

For hospital sales, Step 10 of the MedTech Marketing Playbook uses the buying committee’s steps as the stages.

What changes by stage?

What mistakes should I avoid?

• Bidding on form fills. The platform finds more people who fill in forms; send it opportunities instead (Step 4).
• Marking every event as a key event. Scrolls and downloads flatter every channel; keep the events you would pay for (Step 2).
• Building models before the data is clean. No split is better than its inputs; fix the tracking and the source fields first (Steps 2 and 3).
• Picking one model and calling it the truth. Each rule credits a different touch; set it beside what buyers say and test the gap (Steps 5 and 6).
• A drop-down for “How did you hear about us?” It can only return the options you listed; use free text (Step 5).
• Reporting a green arrow as a result. Check the normal range before you claim a change (Step 7).