Google Ads Attribution Models Compared.

Google killed four attribution models in one go. First-click, linear, time-decay, position-based, all deprecated, all gone from the picker. By 2026 you get exactly two real choices in Google Ads: last-click and data-driven attribution. The official line is that data-driven is smarter, so you do not need the rest.

I will be blunt about what that leaves you with. Last-click, which is dumb but at least predictable. And data-driven attribution, which is a machine-learning model trained on your account's own conversion history.

That second one sounds obviously better. It is better, on one condition that no guide states clearly: only if the conversion data it learns from is clean.

Here is what every "Google Ads attribution models compared" article skips. Data-driven attribution does not invent its credit assignments. It learns them from your historical conversions. So if your conversion events are contaminated by bot clicks and invalid traffic, DDA does not detect that. It learns from it. It treats the contamination as a pattern, builds credit rules around it, and then feeds those rules straight back into Smart Bidding.

This is not a how-attribution-models-work post. There are a hundred of those. This is a post about what happens when the data feeding the model is already poisoned, because that is the part that decides whether DDA helps you or quietly burns your budget.

DataCops is in this conversation because the fix is upstream of the model, in how conversion data is collected and filtered before it ever reaches Google. See the Google Conversion API, fraud traffic validation, and marketing attribution models.

Quick stuff people keep asking

What attribution models are available in Google Ads in 2026? Two that matter: last-click and data-driven attribution. First-click, linear, time-decay, and position-based were deprecated. Google nudges every account toward DDA as the default.

Is data-driven attribution better than last click? Mechanically, yes. It distributes credit across the path instead of dumping it all on the final click. But that advantage only holds if your conversion data is clean.

Train DDA on contaminated conversions and a dumb-but-stable last-click model can actually misdirect you less.

How many conversions do you need for data-driven attribution? Google removed the old hard threshold, but DDA still needs meaningful conversion volume to model anything useful. Thin accounts get a model fitted to noise. Low volume plus contaminated events is the worst case, a confident model built on almost nothing real.

What happened to first-click and time-decay attribution in Google Ads? Deprecated and removed. Google decided DDA subsumes them. If your strategy depended on time-decay, you do not get it back. You get last-click or DDA.

Does attribution model affect Smart Bidding? Yes, directly. The attribution model decides how conversion credit is assigned, and that credit is the signal Smart Bidding optimizes against. Change the model and you change what Target ROAS and Target CPA are chasing.

This is why a bad model does real budget damage.

What is the difference between GA4 attribution and Google Ads attribution? GA4 attributes across all channels for analysis. Google Ads attributes within Google Ads to drive bidding. Different scopes, different numbers, and they will not match.

Use Google Ads attribution for bidding decisions, GA4 for cross-channel context.

Should I use data-driven attribution for small accounts? Cautiously. Small accounts give DDA too little data to model well, and any bot contamination is proportionally larger. A small account with dirty conversions often does better on last-click until volume and data quality both improve.

The gap: data-driven attribution launders bad signals, it does not filter them

Every guide explains the mechanics. Path data in, credit fractions out, Smart Bidding consumes the result. Fine. Here is the part they leave out.

Data-driven attribution is a learning system. It has no concept of a fraudulent conversion. It cannot.

It is handed a set of conversion events and a set of touchpoint paths, and its entire job is to find the pattern that best explains which paths lead to conversions. It does that faithfully. If 24 to 31 percent of the traffic feeding those events is bots, DDA does not flag it.

It models it. The bot pattern becomes part of what DDA thinks a converting path looks like.

That is the difference between a dumb model and a learning model when the input is dirty. Last-click is dumb, so it makes one dumb mistake consistently, all credit to the final click. You can predict it and correct for it.

“

DDA is smart, so it faithfully learns whatever is in the data, including the contamination, and then applies it with confidence across every campaign.

Walk the layers. Analytics and conversion scripts get blocked for 25 to 35 percent of real users, so a chunk of genuine conversions never make it into the dataset at all. Of the traffic that does get measured, 24 to 31 percent is bots.

So DDA is trained on a dataset that is missing real humans and stuffed with fake ones. It is not modeling your customers. It is modeling a distorted shadow of them.

Here is the proof moment. An AI startup called PillarlabAI ran a signup honeypot expecting a bit of noise. They got 3,000 signups, 77 percent fraudulent, and 650 accounts traced to one device fingerprint.

One machine, 650 identities. Picture those 650 fake signups firing as conversion events into a Google Ads account. DDA receives 650 conversions.

It does not see fraud. It sees 650 data points and asks which ad paths preceded them. It builds credit rules to explain them.

Then Smart Bidding, told those paths convert, goes and buys more traffic that looks exactly like the traffic the bots came through.

That is Layer 5, and it is the part that should worry you most. The error does not stay inside the attribution report. DDA hands its corrupted credit to Smart Bidding.

Smart Bidding feeds the bidding decision back into Google's algorithm as training signal. The algorithm learns to find more traffic resembling the bots. ROAS degrades.

You respond by trusting the model more, because it is the smart one. Garbage in, garbage optimized, garbage out, on a loop, with each cycle teaching the system to be more wrong.

Clean conversions in, accurate DDA. Dirty conversions in, DDA becomes a machine for laundering bot noise into budget decisions and dressing it up as data science.

The root cause is not Google's model. DDA is a reasonable model. The root cause is that conversion events are collected by third-party scripts that pour human and bot traffic into the same pipe, with no isolation and no filtering, before the data ever leaves your infrastructure.

Google receives whatever those scripts send. It cannot un-mix it.

The fix is architectural and it sits upstream of the model. Collect conversion data first-party, on your own subdomain. Filter bots at ingestion, before events are counted, against an IP intelligence database of 361.8 billion-plus addresses that separates residential traffic from datacenter, VPN, proxy, and Tor.

That is what DataCops does, and then it sends the cleaned conversion signal onward to Google and Meta through CAPI. Feed DDA filtered conversions and the smart model finally has something real to be smart about. DataCops is a newer brand than the analytics incumbents and its SOC 2 Type II is still in progress, so regulated buyers should factor that in.

“

But on the thing that actually decides whether DDA helps you, the cleanliness of the conversion input, the architecture is the answer.

Decision guide

Running a high-volume ecommerce account with clean tracking? Data-driven attribution, and let Smart Bidding use it. This is the case DDA was built for.

Running a small or low-volume account? Last-click until you have both volume and verified-clean conversions. DDA on thin, dirty data is a confident wrong answer.

Lead-gen with form fills as conversions? Audit those form fills for bot submissions before trusting DDA. Form-fill conversions are a favorite bot target, and DDA will happily learn the fraud.

Just switched models and ROAS dropped? Do not assume the model is wrong. Check whether your conversion events are contaminated. A model change only exposed a data problem that was already there.

Seeing DDA credit a campaign your gut says is junk? Trust the gut, then verify. Pull the traffic quality on that campaign. DDA may be faithfully modeling a bot-heavy placement.

Not sure if your conversions are clean? Then you are not ready to trust any attribution model. Measure your bot contamination rate first. Everything downstream depends on that number.

You are tuning the model when the data is the problem

The whole "which Google Ads attribution model" debate assumes the conversion data is sound and you just need the right math to slice it. That assumption is the mistake. Last-click versus data-driven is a real choice, but it is a second-order choice.

The first-order question is whether the conversion events feeding either model were ever filtered for bots and invalid traffic.

If they were not, then picking data-driven attribution does not make you smarter. It makes you confidently wrong, because you have handed a learning system a contaminated dataset and told it to teach your bidding algorithm.

So before you touch the model picker again, answer this. How many of the conversions in your account last month were real humans, and how do you know? If that number is a shrug, your attribution model is not your problem. Your data is.