Attribution models: who really deserves credit for your landing page's conversions?

End of the month. Google Ads claims a certain number of conversions, Meta's ads manager claims another, GA4 counts a third. And in your CRM, the list of leads you actually received matches none of the three totals. Nothing is broken: you have simply discovered that "a conversion" is not an objective fact but the output of a counting convention. That convention has a name — the attribution model — and on a single landing page fed by several sources, it is what decides who gets the credit. Which means, in practice, where your budget goes next month.

One landing page, five sources, a single form submission

Take the journey to a waitlist signup page. A visitor first finds one of your blog articles through an unpaid Google search: that is discovery, she had never heard your name ten seconds earlier. Three days later, a Meta ad brings her back to the landing page; she reads, she doesn't sign up. The following week, a link in your newsletter brings her back again. Then she types the site address directly one evening, still hesitating. Finally she searches for your brand on Google, clicks your ad, and fills in the form. Five touchpoints, one conversion to distribute — and four of those touches are invisible in the report your tools show by default.

This is well documented academically. In "Mapping the customer journey: Lessons learned from graph-based online attribution modeling" (International Journal of Research in Marketing, 2016), Eva Anderl, Ingo Becker, Florian von Wangenheim and Jan Hendrik Schumann model customer paths as Markov walks, across four large customer-level data sets each covering at least seven distinct online channels. Their conclusion: the sequence of touches carries information, and the usual heuristic rules — last click chief among them — value channels quite differently from a model that accounts for real journeys. Hongshuang (Alice) Li and P. K. Kannan, in "Attributing Conversions in a Multichannel Online Marketing Environment" (Journal of Marketing Research, 2014), estimate the carryover and spillover effects of one channel on another and show that ignoring them distorts each channel's measured contribution.

The six attribution models, plainly

An attribution model is a distribution rule: it takes a conversion and splits it across the touchpoints that preceded it. There is no "true" model — only models more or less suited to the length of your decision cycle, your volume, and the question you are actually asking.

Run those rules over the journey above. Last click gives everything to the brand ad: in your spreadsheet, SEO produced no leads this month. First click flips the verdict exactly. Linear gives a fifth to each of the five touches. Time decay gives the largest share to the brand ad and a residual sliver to SEO. The U-shaped model reserves most of the credit for the first and last touch. Five readings, five different budget decisions — for one and the same lead.

Why last click over-credits retargeting and brand

Last click is the default in most dashboards, and the one that produces the most expensive steering errors. End-of-journey channels capture all the credit regardless of their real role — and they speak to people who are already convinced. Retargeting only addresses visitors who already know your page; ads on your own brand only capture people typing your name; the newsletter only reaches people who already handed over their email. These channels therefore post excellent acquisition costs almost by construction, and trigger the fatal reflex: "let's increase retargeting, that's what converts." But cut discovery and there will soon be nobody left to retarget.

The most-cited experiment on this point is Thomas Blake, Chris Nosko and Steven Tadelis's "Consumer Heterogeneity and Paid Search Effectiveness: A Large-Scale Field Experiment" (Econometrica, 2015). The authors actually switched off eBay's paid search ads in selected markets in order to measure their causal rather than correlational effect. The result: the genuinely causal return on paid search was only a fraction of what non-experimental methods estimated, with the gap widest on brand queries, where users would have arrived through the organic result anyway. The most flattering line in your last-click report is often the one with the least real incremental value.

Why Google Ads, Meta and GA4 will never agree

• Conversion windows differ. Each platform only counts a conversion if it happens within a certain delay after the click, and that delay is neither the same everywhere nor necessarily configured the same way in your accounts.
• View-through attribution is not click-through attribution. Meta may credit a conversion to a mere impression, where an analytics tool reasons in click-driven visits: the same conversion exists on one side and not the other.
• Each ad platform only sees its own perimeter. Google Ads has no idea what Meta did, Meta has no idea about your newsletter, and each claims 100% of what it observes. None of them can deduplicate a lead touched by both.
• Modelling kicks in when data is missing. Consent refused, third-party cookies blocked, restricted tracking on mobile: some conversions are never observed, and each platform estimates a share of them in its own way.
• The reporting date is not the same. Google Ads attributes a conversion to the click date by default, an analytics tool to the date it actually happened: over a rolling month, the totals diverge even with identical visitor behaviour.

The one practical takeaway: never add up the conversions claimed by your different ad platforms. The total will always exceed the number of leads actually sitting in your CRM, because several platforms are claiming the same people. The only source of truth on real volume is your CRM or your billing tool. It is also the only sane basis for calculating your landing page's ROI.

What consent and server-side measurement change

All of this rests on an identification chain that, in Europe, is conditional on consent. If a visitor declines measurement cookies on your banner, the journey fragments: today's visit and next week's are no longer connected, and the model has little left to distribute. That is what Google's consent mode v2 tries to compensate for, by transmitting consent state instead of dropping the signal entirely, which lets Google estimate part of the unobserved conversions. Note what that data is: modelled conversions, not observed ones. Useful for reading a trend, never for justifying a figure down to the unit.

Server-side measurement pursues the same goal differently: instead of letting the browser send the information to the platform, your server does it, using first-party data. That is the principle behind Meta's Conversions API paired with the pixel and, on Google's side, enhanced conversions, which send a hashed identifier to match a lead with an ad click. It hardens measurement against technical blocking, but it does not exempt you from consent: what is off-limits in the browser stays off-limits on the server. And it only works if your inputs are clean — hence the importance of rigorous UTM tagging across every link, emails and social posts included.

The pragmatic stance when you have no data team

1. Pick one model and stick with it. The worst outcome isn't picking the "wrong" model, it's swapping lenses every month: your comparisons become uninterpretable.
2. Designate a single source of truth for volume. Your CRM or billing tool counts the real leads; the ad platforms are there to compare campaigns against each other, not to establish the monthly total.
3. Watch trends, not absolute values. "This campaign's cost per lead has been rising for three weeks" is actionable; "Meta and GA4 don't show the same total" is not, and reconciling them will eat time you don't have.
4. Run a blackout test when the doubt is expensive. Before increasing budget on an end-of-journey channel, pause it for a few weeks and watch total lead volume. If it barely moves, that channel was harvesting conversions you already had.
5. Ask the source directly. A "How did you hear about us?" field is self-reported and imperfect, but it captures what no cookie ever will: word of mouth, a podcast, a recommendation in a meeting. Cross-referenced with your conversions tracked in GA4, it reveals complete blind spots.
6. Check the interface before relying on a specific model. Attribution options in both GA4 and Google Ads have changed several times, with some models removed along the way: look at what your property actually offers today rather than trusting a dated tutorial.

Attribution is not a problem you solve; it is an uncertainty you learn to frame. No model will tell you the truth about a human decision spread over three weeks and five touchpoints. But a consistent model, a single source of truth and a healthy scepticism towards end-of-journey channels are enough to decide considerably better than average. The prerequisite is a clean, properly instrumented page: the saas-waitlist template from LanderKit is built for exactly that, with its waitlist signup form and a separate confirmation page that lets you fire a clean conversion event. Like the nine other templates in the catalogue, it is available at €89 each or in the full pack at €229.