Attribution Models Explained: Which Channel Actually Made the Sale?
A Singapore B2B services firm sat down with three monthly reports. The SEO report claimed 18 enquiries. Google Ads claimed 14. Meta claimed 9. The sales team had logged 23 real enquiries in the CRM. Three suppliers had between them claimed 41 leads out of 23, and every one of them could show you the dashboard.
This is the most common measurement argument in Singapore marketing, and almost nobody wins it, because both sides are arguing about the wrong thing. The disagreement is not usually about honesty. It is about attribution: the rules each system uses to decide which touchpoint deserves credit when a customer saw four things before buying.
This guide explains attribution in plain English — which models still exist in 2026, what changed, why the credit windows matter far more than the model you pick, what Meta’s 2026 redefinition did to your numbers, and what actually tells you the truth. It sits under our performance marketing guide for Singapore, alongside how to set up conversion tracking properly.
What attribution actually is
A Singapore customer buying a S$3,000 service does not see one ad and buy. A realistic path looks like this: sees an Instagram reel, searches the company name three days later, reads two blog posts, clicks a Google ad a week after that, asks a friend, comes back via a bookmark, and enquires on a Tuesday morning.
Seven touchpoints, one sale. Attribution is the rule that divides one sale among them. Give all the credit to the last click and Google Ads looks brilliant while Instagram looks worthless. Give it all to the first and the reverse happens. Neither is true, and there is no mathematically correct answer — only a set of conventions with different biases.
The important consequence: attribution is not a measurement, it is an accounting policy. Changing your model changes your reported numbers without changing a single sale. Anyone who presents attributed figures as facts about causation has misunderstood the tool.
The models that still exist in 2026
The good news is that this got dramatically simpler. Google spent 2023 deleting most of the models everyone used to argue about, and both Google Ads and GA4 now offer a short list.
Google Ads Help is blunt about it: “The first click, linear, time decay, and position-based attribution models are no longer supported by Google.” GA4’s documentation says the same four models “are no longer available as of November 2023.”
| System | Models available in 2026 | Default | Gone since Nov 2023 |
|---|---|---|---|
| Google Ads | Data-driven, last click | Data-driven for most conversion actions | First click, linear, time decay, position-based |
| GA4 | Data-driven, paid and organic last click, Google paid channels last click | Data-driven | The same four |
| Meta Ads Manager | No model choice — only attribution settings (windows) | 7-day click, 1-day engage-through, 1-day view | n/a |
Three things are worth extracting from that table. Meta never offered models at all, only windows, which is why “which attribution model does Meta use” has no answer. GA4’s “Google paid channels last click” is a deliberately narrow view that only credits Google properties — useful for reconciling with Google Ads, misleading for anything else. And the once-popular position-based model, which gave 40% to the first touch and 40% to the last, no longer exists in either tool.
What data-driven attribution actually does
Data-driven attribution (DDA) is the default in both systems now, and it is the least understood. It does not apply a fixed formula. Google describes it as distributing “credit for the conversion based on your past data for this conversion action”, using your account’s own data “to calculate the actual contribution of each interaction across the conversion path.”
In practice it compares paths that converted with paths that did not and assigns fractional credit accordingly. This is genuinely better than a rule-of-thumb split, with two caveats. It needs enough data to model, and Google “filters out any network or campaign without enough data” — so a low-volume Singapore account may see thin or unstable results. And because it is derived from your own history, it cannot tell you what would have happened with no advertising at all. It divides the credit; it does not test whether the credit was deserved.
The windows matter more than the model
Here is the part almost every attribution discussion misses. Switching between last click and data-driven typically shuffles credit between channels by a modest amount. Changing the window — how long a touchpoint stays eligible for credit — changes the totals themselves. And the defaults across platforms are wildly different.
| Setting | Default | Range | Retroactive? |
|---|---|---|---|
| Google Ads click-through window | 30 days | 1 to 30, 60 or 90 days | No — applies going forward |
| Google Ads engaged-view | 3 days | 1 to 30 days | No |
| Google Ads view-through | 1 day | 1 to 30 days, or 1 to 4 weeks | No |
| GA4 key event lookback, acquisition events | 30 days | 7 or 30 days | No — applies going forward |
| GA4 key event lookback, all other key events | 90 days | 30, 60 or 90 days | No |
| GA4 reporting attribution model | Data-driven | Three models | Yes — applies to historical and future data |
| Meta click-through | 7 days | 1 or 7 days | Reporting view only |
| Meta engage-through | 1 day, fixed | 1 day or none | Reporting view only |
Read the last column carefully, because it produces a specific and confusing behaviour. In GA4, changing the reporting attribution model rewrites history — “changing the reporting attribution model applies to historical and future data” — while changing the lookback window does not, because window changes “apply going forward”. So one setting appears to change the past and the other does not. Both are documented; neither is intuitive.
One more GA4 subtlety that has caused a great many pointless arguments: the reporting attribution model does not touch the traffic dimensions most people actually look at. Google states that “user- and session-scoped traffic dimensions, such as Session source or First user medium, are unaffected by changes to the reporting attribution model.” So you can switch models and watch your Sessions by channel table refuse to move.
A worked Singapore example
Take a Singapore renovation company with a genuine six-week consideration cycle. A homeowner sees an Instagram ad on 1 March, clicks a Google search ad on 20 March, and signs a S$45,000 contract on 12 April.
| System | What it sees | Credit given |
|---|---|---|
| Meta, 7-day click default | Instagram click was 42 days before the sale | Nothing. Outside the window entirely. |
| Google Ads, 30-day click default | Search click was 23 days before | Full credit — one conversion, S$45,000 |
| GA4, 90-day lookback, data-driven | Both touchpoints | Split fractionally between paid social and paid search |
| The CRM | One signed contract | One sale |
Nobody has malfunctioned. Meta’s default window is simply shorter than this company’s sales cycle, so Meta cannot see its own contribution. If the marketing manager reallocates budget away from Instagram on the strength of the Meta dashboard, they will be cutting the channel that started the sale — and the Google Ads number will fall a quarter later, mysteriously.
Five structural reasons the numbers never add up
- Different windows. As above. Two platforms looking at the same purchase from 7 and 30 days out will disagree permanently.
- Different scope. Each ad platform only counts conversions it can tie to itself. Neither can see the other. Only GA4 and your CRM see everything.
- Self-reported credit. Every platform is simultaneously the participant and the referee. Nobody is lying, but nobody is disinterested either.
- View-through and engagement credit. Meta and Google both claim conversions where the ad was merely seen or engaged with. Whether that credit is deserved is a genuinely open question.
- Date conventions. Google Ads credits a conversion to the date of the click, not the sale, so last month’s total keeps rising for weeks. Meta reports on the impression date. Your invoice is dated when you invoiced.
The practical implication is a rule you can hold onto: never add platform-reported conversions across channels. The sum always overstates reality, and the overstatement grows with how much retargeting you run.
What Meta’s 2026 changes did to your numbers
If your reported Meta conversions dropped in early 2026, this is almost certainly why. Two changes landed within two months.
In January 2026 Meta removed the longer view-through windows, leaving one day as the longest available. Then on 3 March 2026 it redefined click-through attribution to require an actual link click — one that sends someone to a website, app or lead form. Likes, shares, saves, comments, profile taps and short video views, all of which previously counted as clicks, moved into a new engage-through category with a fixed one-day window that cannot be extended to seven. The engaged-view threshold also fell from ten seconds of video to five.
The default attribution setting is now 7-day click, 1-day engage-through, 1-day view. Reported click-through conversions fell as a definitional artefact, not a performance decline — Meta confirmed the change applies to how conversions are classified in reporting, and billing was unaffected. If you compare a March 2026 report to a January one without noting this, you will conclude your campaigns broke. They did not; the categories moved. These dates come from practitioner and vendor accounts rather than a fetchable Meta support document, so treat them as well-corroborated rather than officially quoted. Our guide to Meta versus Google Ads for Singapore businesses covers what this means for channel comparisons.
New in GA4: the AI Assistant channel
One change from 2026 that Singapore businesses should know about, because it is genuinely new and genuinely under-discussed. On 13 May 2026 Google added an AI Assistant channel to GA4’s Default Channel Group. It identifies visits from “chatbots like ChatGPT, Gemini, and Claude”; when traffic matches a recognised AI assistant referrer, the medium is set to ai-assistant and the campaign to (ai-assistant).
Before this, that traffic mostly landed in Referral or Direct, which is why so many Singapore businesses had a growing unexplained Direct segment. It is now separable, and worth watching monthly: it is the only first-party evidence most SMEs have about whether AI search surfaces are sending them anyone. Note it measures clicks through from an assistant, not mentions inside an answer — someone who reads your figures inside a chatbot and never clicks is still invisible.
Google also added a Source Group dimension on 11 June 2026, which consolidates the messy variants of a single platform — “facebook”, “fb”, the various Instagram strings — into clean categories. If your channel reporting has been split across four spellings of the same source, that is the fix. A new hostname data filter arrived at the same time, letting you exclude events by hostname.
What actually tells you the truth: incrementality
Attribution answers “which touchpoint should get the credit”. The question a business owner actually has is different: “if I switched this off, what would I lose?” That is incrementality, and attribution cannot answer it. Only an experiment can — a holdout group, or a geographic test where some regions get the campaign and others do not.
Two pieces of published evidence are worth holding in mind together, because they point in opposite directions and both are informative.
Haus analysed 640 Meta incrementality experiments and found an average lift of roughly 19% to brands’ primary KPI, with Meta under-reporting by around 15% on a 7-day-click direct-to-consumer basis, and roughly 32% of the incremental impact landing outside the online store for omnichannel brands. Seer Interactive, testing about US$1.05m of spend across six accounts, found Meta claimed 87% of its conversions were incremental but only 67% held up when cross-referenced against GA4 — with broad and tight-retargeting audiences showing the weakest incrementality and mid-funnel the clearest.
Both can be true. A platform can understate its total effect on the business while overstating how much of a specific conversion set it caused. That is not a contradiction; it is a warning against trusting any single number, in either direction.
A practical attribution setup by budget band
You do not need a data team. You need a policy proportionate to your spend.
| Monthly media spend | What to do | What to skip |
|---|---|---|
| Under S$3,000 | Platform defaults. Track total leads and total spend in one spreadsheet. Compare blended cost per lead month to month. | Attribution debates entirely. There is not enough volume to model. |
| S$3,000 to S$15,000 | Set windows to match your real sales cycle. Use GA4 as referee. Report blended cost per lead alongside platform figures and explain the gap. | Third-party attribution tools. Multi-touch modelling. |
| Above S$15,000 | All of the above, plus one geo or holdout test per quarter on your largest channel. Measure incrementality on retargeting first. | Believing any platform’s self-reported ROAS without a test behind it. |
And whatever the band, keep the one number that cannot be gamed: total qualified leads or total revenue from your own records, divided by total media spend. It has no attribution theory in it, it cannot double-count, and it is the number your accountant recognises. Our guide to ROAS versus ROI shows how to turn it into a profitability figure, and how to read your monthly marketing report shows what an honest report looks like.
What consent does to your attribution
Attribution depends on being able to recognise the same person twice, which is exactly what privacy protections are designed to prevent. In Singapore the position is narrower than Europe’s but real: the PDPC’s Advisory Guidelines on Selected Topics (revised May 2024) require consent for cookies used to target advertising, browser inaction does not count as consent, and the Legitimate Interests Exception does not cover direct marketing.
Add browser behaviour on top. Safari’s tracking protection deletes JavaScript-written storage after about seven days, which means long attribution windows are partly fictional on iOS regardless of your settings. The honest framing for any Singapore client is that attributed figures are directional, the blended number is reliable, and only a test is evidence. None of this is legal advice.
The conclusion nobody wants to hear
Attribution will never reconcile, because it was never designed to. The B2B firm at the top of this article did not need a better model; it needed to stop treating three self-reported dashboards as three independent witnesses. Its fix took an afternoon: one CRM number as the source of truth for lead volume, windows set to match a real six-week cycle, platform reports used to compare each channel against its own history rather than against each other, and a quarterly holdout test on the largest line item.
The reported numbers went down. The arguments stopped, because everyone was finally looking at the same thing.
If you would rather have someone else own that, see our performance marketing services and the results we have produced for Singapore businesses. If the plumbing is the problem rather than the policy, start with setting up GA4 properly, and for organic channels see measuring social media ROI.
Frequently asked questions
Which attribution model should I use?
For most Singapore businesses, leave both Google Ads and GA4 on data-driven, which is now the default in both. Use GA4’s “paid and organic last click” as a sanity check when you need a simple, explainable view, and “Google paid channels last click” only when reconciling against Google Ads. The model matters far less than setting your credit windows to match your real sales cycle.
Why does Meta report fewer conversions than Google Ads for the same month?
Mostly the windows. Meta’s default click window is 7 days; Google Ads’ is 30. If your sales cycle is longer than a week, Meta structurally cannot see much of its own contribution. On top of that, Meta’s March 2026 redefinition moved likes, comments, shares and short video views out of click-through attribution into a 1-day engage-through bucket, which reduced reported click-through conversions without any change in performance.
Do first click, linear and time decay models still exist?
No. Google Ads Help states that “the first click, linear, time decay, and position-based attribution models are no longer supported by Google”, and GA4’s documentation says those four “are no longer available as of November 2023”. Google Ads now offers data-driven and last click; GA4 offers data-driven, paid and organic last click, and Google paid channels last click.
Will changing my attribution model change my past reports?
In GA4, yes for the model and no for the window. Google states that “changing the reporting attribution model applies to historical and future data”, while lookback window changes “apply going forward”. Google Ads conversion window changes are also forward-only. And note that GA4’s user- and session-scoped traffic dimensions, such as Session source, are unaffected by the reporting attribution model at all.
How do I know which channel really caused a sale?
Run an experiment, not a report. Switch a channel off in some regions and not others, or hold back a portion of the audience, and compare total business outcomes. Haus found an average incremental lift of roughly 19% across 640 Meta experiments, while Seer Interactive found Meta’s claim that 87% of its conversions were incremental fell to 67% when cross-referenced against GA4. Attribution divides credit; only a test measures cause.
What is the AI Assistant channel in GA4?
A channel Google added on 13 May 2026 to the Default Channel Group, which identifies traffic from chatbots including ChatGPT, Gemini and Claude by setting the medium to ai-assistant. Previously that traffic landed mostly in Referral or Direct. It measures clicks through from an assistant, not mentions inside an AI answer, so it is a floor on your AI visibility rather than a full picture.
Related measurement guides
- UTM tags done properly — attribution can only credit a channel it can identify.
- GA4 reports every business owner should check — where these models actually surface in your reporting.
