AI Marketing Revenue Attribution: Which Method to Trust

Why attribution figures disagree

Each advertising platform reports the conversions it can link to its own ads, using its own rules. A customer who saw a Meta ad, clicked a Google ad and then bought can be counted by both. Analytics tools then apply their own model to share credit. The CRM knows which deals closed but often not which ads were involved. None of these is lying; each answers a different question.

The decision is not which method is true, but which method is fit for which decision.

Worked example: the double count

Illustrative example with hypothetical round figures. An online shop records 50 orders in a month.

• • Meta Ads Manager reports 40 purchases.
• • Google Ads reports 35 conversions.
• • Together the platforms claim 40 + 35 = 75, which is 25 more than actually happened, or 150% of real orders.

If the business sums platform-reported revenue to judge its marketing, it overstates the return by half. That is before counting orders that would have happened with no advertising at all.

The six methods compared

Attribution and measurement methods

Method	Question it answers	Data needed	Main bias	Effort	Suits
Last click (analytics)	Which channel delivered the final click before conversion?	Analytics tracking	Ignores everything before the last click; favours search and brand	Low	Quick operational reads on landing pages and campaigns
Platform-reported conversions	What does each platform credit to its own ads?	Platform pixels or tags	Each platform counts overlapping conversions; totals exceed reality	Low	Optimising within one platform
GA4 or Google Ads data-driven attribution	How should credit be shared across touchpoints the tool can see?	Enough conversion volume in the account	Only sees tracked digital touchpoints; still correlation, not causation	Low to moderate	Comparing channels inside one analytics view
CRM-based attribution with offline imports	Which campaigns preceded real opportunities and revenue?	CRM with source capture; offline conversion import or conversions API	Depends on clean source capture; misses unrecorded influence	Moderate	B2B and any business where revenue is confirmed after the click
Lift or holdout tests	How many conversions would not have happened without this activity?	Enough audience to split into test and control	Measures one channel or campaign at a time; needs volume	Moderate	Deciding whether a channel is truly incremental
Media mix modelling	How much does each channel contribute, and what happens if budgets change?	Months or years of spend and outcomes by channel	Model assumptions; weak with short or flat history	High	Larger multi-channel budgets and annual planning

What each method is good for

Last click and platform reports

Both are cheap and fast, and both are biased in predictable directions. Use them to run campaigns day to day: which ad, which landing page, which audience. Do not sum them to judge total return.

Data-driven attribution in Google's tools

Google Ads now uses data-driven attribution as the default for most conversion actions, and the older rules-based models (first click, linear, time decay, position based) are no longer supported; last click remains available. In GA4 you choose the reporting attribution model in the property's attribution settings, and changing it applies to historical and future data. The default lookback window is 30 days for acquisition key events and 90 days for other key events. Data-driven models share credit more sensibly than last click, but they only see tracked touchpoints and they describe patterns, not causes.

CRM-based attribution

For businesses where revenue is confirmed later, such as B2B deals or phone sales, the CRM is the only place real revenue lives. Capture the original source on every lead, and send qualified and won stages back to the ad platforms. Google Ads supports importing offline conversions from a CRM, with enhanced conversions for leads as its recommended upgrade, and LinkedIn's Conversions API does the same for LinkedIn campaigns. This lets bidding optimise for revenue-bearing outcomes rather than form fills. AI for B2B Marketing ROI covers how to value the pipeline this produces.

Lift tests

A lift test compares a group exposed to ads with a holdout that was not, so it estimates the conversions caused by the activity rather than those merely associated with it. Meta's Conversion Lift works this way. Use lift tests to answer "is this channel worth keeping?" for the channels where you spend most.

Media mix modelling

Media mix models estimate each channel's contribution from historical spend and outcomes, including offline activity. Google's Meridian is an open-source example using Bayesian methods, and produces response curves and budget scenarios. It is a planning tool for businesses with enough history and channels; for smaller budgets it is usually more than is needed.

A layered set-up for most businesses

1. • Daily and weekly: platform reports and last click, for running campaigns.
2. • Monthly: CRM or order data as the revenue truth, with platform claims reconciled against it. The gap between the two is your double counting.
3. • Quarterly: at least one lift test or budget step test on your largest channel, to check it is incremental.
4. • Annually, at scale: a media mix model to guide the budget split.

Where AI fits

AI-assisted analysis is useful for the laborious parts: matching CRM records to campaigns, reconciling platform claims against orders each month, and drafting the commentary. Platforms' own data-driven models are themselves machine learning. What AI cannot do is make an attribution model causal; only experiments or careful modelling get closer to that. For turning attributed and incremental figures into budget moves, see Optimising Marketing Spend With AI.

FAQ

Which attribution model should a small business use?

Use your order or CRM data as the revenue truth, platform reports for day to day optimisation, and the default data-driven model in Google's tools. Add a lift test when one channel dominates spend.

Why does changing the GA4 attribution model change past reports?

Google applies the reporting attribution model to historical and future data, so earlier periods are recalculated. Note the date and model whenever you save a report.

Can I trust view-through conversions?

Treat them with caution. They credit ads people saw but did not click, which may or may not have influenced them. A lift test is the way to find out.

Is attribution still possible with privacy restrictions and consent choices?

Less precisely at user level, which is one reason aggregate methods such as lift tests and media mix models have become more important alongside consented first-party data.