What I've learned (and keep relearning) about B2B attribution

This year, in a bid to move upmarket, we've started to invest heavily in executive hospitality. To track success, I anchored on a single metric: net new first-sale pipeline, where the first touch was the event.

In hindsight, this was a disaster.

To hit the number, our reps had to invite strangers, net new accounts they'd never spoken to. The whole point of executive hospitality is depth. You take your most important relationships somewhere memorable and you let the experience do the work. Instead, our reps were cold-pitching people they'd just met.

We fixed it by changing what we measured. Now we focus on total open pipeline touched at the event, including existing customers, with explicit credit for pipeline acceleration and cross-sell.

B2B attribution is full of traps like this. The model you choose shapes the behavior you get, often in ways you don't see until it's too late.

Be honest about why you want attribution

Before you pick a model, you need to be honest with yourself about what you're actually trying to do. There are at least three reasons people set up attribution, and they call for different tools.

Reason 1: You want to perfectly optimize every dollar of marketing spend. Sorry but this is impossible. There are too many interactions, the data is too messy, and the human brain is too irrational. If this is genuinely your goal, focus on lift studies. Run a holdout, measure the difference, accept that the answer comes with a confidence interval.

Reason 2: You want to justify marketing's existence internally. I get it. For better or worse, this is the easiest goal to cheat. You can pick the metric, pick the model, and shape the narrative however you want, changing it each quarter or even each campaign to serve your needs.

Reason 3: You want to track the relative performance of channels and campaigns to make better decisions. This is the best reason to get into attribution. And even here, it's more art than science.

Attribution models worth understanding

No model: taste and intuition

This is where every company starts and how many decisions still get made. Decisions get made based on what the marketing leader thinks is working, what the CEO read on Twitter last weekend, and what the sales team is complaining about.

It works and should be a part of every model, but it's hard to scale taste and becomes increasingly ineffective as people making purchasing decisions get further away from the customer. For a small team running a handful of channels though, taste and intuition are often more accurate than any model you'd build.

Signs of life

The simplest real attribution: ask people how they heard about you. Put it on the contact sales form and ask in the first call.

You will not capture every conversion. The data will be directional at best. Someone will write "Google" and mean a podcast they heard about on a friend's recommendation that they then Googled. That's fine. You're not trying to build a model, you're trying to find a signal.

Again, this works and should be a part of every model, particularly because it captures sources that are hard to track and ones that you don't own: a podcast mention, a friend's recommendation, a conference hallway conversation.

This model started to break down for us when ElevenLabs became so well known that we'd ask people how they heard about us and they'd say "Is that a serious question? You're ElevenLabs. You're everywhere."

First touch (or last touch)

In a first touch attribution model, you log the first trackable touchpoint for every lead and give 100% credit to that channel. For example, if a prospect navigates to your site for the first time from a non-branded google search ad, even if it takes them 12 months to convert they will always be known as a non-branded google search lead.

This model works well for short sales cycles where there aren't many touchpoints to argue about. One issue is that if you commit to first touch, it's hard to justify spending on anything mid or upper funnel. YouTube, CTV, LinkedIn document ads, and podcasts all influence buyers without being the click that led to a form fill. Your first touch model will tell you these channels don't work.

Data-driven multi-touch

A multi-touch model splits credit across every touchpoint in the buyer journey. A data-driven multi-touch model splits credit across every touchpoint in the buyer journey, weighted by how much each one contributed (as opposed to splitting linearly or U-shaped). DreamData has a good explainer on how this works.

This is the right model for long, complex sales cycles where buyers interact with you many times before converting. It will tell you things that first touch cannot: that the LinkedIn ad someone scrolled past in March mattered, that the webinar in June moved them, that the case study in August closed them.

But multi-touch has its own blind spots. It overweights channels that have clean tracking: anything with a UTM, anything with a direct platform integration. LinkedIn, paid search, your owned email program. It underweights channels that don't click well, like podcasts, where someone hears your CEO on a show and Googles you a week later. The model sees a "direct" or "organic search" touchpoint and gives the credit to the wrong channel.

The truly dangerous thing about a data-driven multi-touch model is that you spend so much time setting up this mystical black box that you will convince yourself it's an all knowing oracle when in fact it's as flawed as any other attribution model.

How attribution starts controlling you

You set up an attribution model so that you could supplement your intuition with data. If you're not careful, the model starts making the decisions for you. And because the model has blind spots, those decisions get worse over time.

If you commit to first touch, you'll defund every top and mid funnel activity, even the ones that work. If you commit to multi-touch, you'll overinvest in the channels with good tracking and underinvest in the ones without. Your strategy will gradually narrow to whatever your attribution platform can see clearly.

There are other things attribution struggles with that are worth naming.

Channel interaction effects. Some channels work in conjunction with each other but not in isolation. The podcast plus the retargeting ad plus the case study work together. Pull any one out and the other two stop performing. Attribution will tell you the case study closed the deal, then suggest you cut the podcast.

New markets and TAM expansion. When you move into a new geography or audience, you have to prime them first. The early spend will look like it isn't working because nothing converts immediately. If you let the attribution model judge the campaign in month two, you'll kill it before it has a chance to compound.

Maturation lag. A new campaign takes time to mature. If your sales cycle requires fifty touchpoints to become a lead and another fifty to become an opportunity, you will not see meaningful pipeline attributed to a campaign for months. Patient capital is required, and attribution will make you impatient.

What we do at ElevenLabs

Multi-touch as the primary system. Signs of life as the supplement, because the contact sales form and gong calls catch things our model can't see. Lift studies for finding and measuring performant channels.

Attribution is a tool. The moment you start treating it as the truth, it starts working against you. Taste and intuition still matter. At least for my sake, I hope they do.