Search or social: how to read attribution without cutting the wrong channel

The reports are not lying: they answer different questions

An attribution model is the rule that splits credit for a deal among touchpoints: contacts a tracking system recorded, such as an ad view or click, a search visit, an email or a direct visit. The sequence of touchpoints before a deal is a path.

Someone sees your ad on Instagram, finds you on Google a week later and enquires. First touch credits social, last touch credits search. Nothing is broken; these are two questions.

The second source of disagreement is who counts. An ad platform mostly sees its own clicks, analytics misses a friend's recommendation, and your CRM knows who bought but often not how. Before arguing, pin down the system, the model and the attribution window: how many days before the deal a touchpoint still counts.

Which model answers which question

First touch credits the channel that started the path; last touch credits the final visit, direct included. Last non-direct touch skips direct and internal visits. Data-driven attribution splits credit with the platform's own calculation, trained on its own data.

The menu has changed. On April 6, 2023 Google announced that first click, linear, time decay and position-based models were leaving Google Ads and GA4; GA4 dropped them in November 2023. GA4 now offers data-driven and two last-click models, none of which credits direct visits unless the path is entirely direct. First-visit source survives as First user dimensions.

Yandex Metrica, an analytics tool common in Russian-speaking markets, trimmed its list on May 20, 2026. It ends a visit history after a 90-day gap that cannot be changed.

The same 10 deals: paid social gets 6, or zero

A teaching example with made-up data. A home renovation contractor uses paid social, paid search, organic search and email, and gets direct visits. We rebuilt the paths of 10 closed deals, assuming each deal counts once, every touchpoint falls inside the window and every path is known.

The paths: 1) paid social → paid search → direct; 2) paid social → organic search → paid search; 3) paid social → direct; 4) paid social → email → direct; 5) paid search; 6) paid search → direct; 7) organic search → paid search; 8) paid social → paid search; 9) paid social → organic search → direct; 10) email → paid search → direct. Nine of ten deals had more than one touchpoint, 2.4 on average.

Under first touch, paid social brings 6 of the 10 deals. Under last touch it gets zero, and direct leads with 6: the person already knew the business. Under last non-direct touch, paid search gets 7 and paid social 1. Split each deal equally across its meaningful touchpoints (the linear model Google retired in 2023) and paid social gets 3.3, paid search 4.3.

Data-driven attribution cannot be rebuilt by hand: the platform trains it on all paths, converted or not. Every model answers its own question. None of them tells you how many deals would disappear if you switched the channel off.

Paid social looks expensive, yet it opens the paths

An attribution table shows how credit is split, not why people bought.

Add spend: $12,000 on paid social and the same on paid search. Spend per credited deal moves with the model more than with the channel.

Under last non-direct touch, paid social costs $12,000 per deal against $1,714 for search, seven times as much. Yet it was first on all six of its paths and the last meaningful touch on only one. A channel that introduces the business and creates interest is called demand-building. It is rarely last, so last-touch models undercredit it. Some search visits may be people looking up a name they first saw on social.

The opposite mistake is just as easy. Under first touch, paid social costs $2,000 per deal, but that does not prove those six customers would not have come anyway. An attribution table shows how credit is split, not why people bought. That is our reading, not a platform rule. Our article on lead cost, customer cost and ROAS turns this into acquisition cost.

The platforms claimed 13 deals, you closed 10

Each platform counts its own touchpoints. Data-driven attribution in Google Ads splits credit among Google ad interactions, comparing paths of customers who convert with those who do not; social ads are not part of it. Meta's standard attribution counts an event within 1 or 7 days of a link click, 1 day of an impression, and 1 day of another engagement, such as a 5-second video play.

Paid social sat on six paths and paid search on seven. If every touchpoint fell inside each window, Meta could claim up to 6 deals and Google Ads up to 7: 13 against 10 real deals. This illustrates the assumption, not what dashboards will show; windows, models and view counting differ.

Meta warns that ad sets with different attribution models cannot be compared in the campaign overview table. Compare channels in one system, model and window.

Data-driven attribution will not settle the argument for you

In GA4, data-driven attribution applies machine learning to converting and non-converting paths, weighing timing, device, the number and order of ad interactions and creative type. Credit follows how much each touchpoint changes the estimated probability of the action; for Google ad exposures it trains on randomized controlled trials. Credit can be reattributed for up to 7 days.

Google Ads recommends at least 200 conversions and 2,000 ad interactions within 30 days, though the model works with less. Meta's incremental attribution optimizes delivery for conversions its models predict an ad caused. Metrica's automatic attribution uses cross-device matching and anonymized household signals, with data from April 8, 2023.

The model sees only recorded touchpoints, cannot be checked by hand, and each platform trains its own. It neither replaces a view of CRM deals nor proves you would sell less without the channel. That is our assessment.

Where paths break: cookies, devices and offline sales

The example assumes every path is known. Real paths break, and the first touch may fall out of the report.

• Cookies and consent. GA4 identifies users by a user ID after login, a device ID, or modeling for those who decline identifiers such as cookies. If someone clears their browser or declines cookies, the path can break and the next touchpoint will look like the first.
• Devices. An ad seen on a phone, a form sent from a laptop: without a login, two visitors. GA4 and Metrica stitch some journeys; none promises a complete match.
• Offline. Calls, chats, meetings, referrals and invoice payments are invisible to website analytics until you send them in.
• Windows. GA4 defaults to 90 days for most key events, with 30 and 60 as options; Metrica ends a history after 90 days without visits; Meta's standard attribution counts clicks for up to 7 days. Since January 12, 2026 Meta's Ads Insights API no longer returns 7-day and 28-day view windows.

Only a test shows what happens without the channel

Incrementality is the part of the result that would not exist without the channel. No attribution model shows it; an experiment does, by comparing people or regions that do not see the ads with the rest. Google calls Conversion Lift a way to measure the causal impact of ads; Meta's lift tests create test and holdout groups automatically and divide incremental conversions by the estimated conversions without the ads.

A holdout test switches the channel off for a set period and compares CRM deals with the previous period and the same period last year, allowing for seasons and competitors. A geo test changes ads in some cities. GeoLift, Meta's open-source tool, builds the control as a weighted mix of other locations from pre-launch history and advises covering a full purchase cycle.

A platform test means Meta Conversion Lift or Google Ads Conversion Lift, which not every account can use; its geography-based form supports offline data. With few deals the difference can drown in noise, and Meta may show spend minimums. Write the plan down first.

1. The question: how many deals we lose if we halve paid social, not which channel is better.
2. A CRM metric: deals or payments, not platform conversions.
3. The method and a duration no shorter than your sales cycle.
4. Which result leads to a cut, a hold or an increase.
5. What stays fixed: prices, offer, sales process.

Brief for your team: one table and one test

You can do this in-house. This journal is published by WEBNOSTRA CAPITAL, which is connected to WEBNOSTRA, a company that sells advertising and analytics setup services. The steps below need no agency and no paid tools.

If the models broadly agree, a budget decision carries less risk. If they diverge sharply, you likely have a demand-building channel, and cutting it untested is a gamble.

1. Export closed CRM deals and rebuild each path from analytics, link tags and the answer to how did you hear about us.
2. Put the deals through first touch, last touch and last non-direct touch in one table.
3. Flag channels that open paths but are rarely last.
4. Label every number with its system, model and window.
5. For the channel you want to cut, draft a holdout or geo test on CRM deals.
6. After the test, decide: cut, hold or reallocate, and write down why.

Key decisions

1. look at any channel decision through at least two models: first touch and last non-direct touch
2. take the deal count from your CRM and never add up platform reports: each is the platform's view of itself
3. cut a channel that often opens paths only after a holdout test or a geo test

What to discuss with your team

• Every number in the report names its system, model and attribution window
• The deal count comes from the CRM, and platform reports are not added together
• Deals are viewed under at least first touch and last non-direct touch
• A channel that opens paths is not cut on a last-click report alone
• Hypothesis, duration, metric and the decision for each outcome are written down before the test

Sources

1. Google Analytics: Get started with attributionsupport.google.com · checked 31/10/2026
2. Google Analytics: Select attribution settingssupport.google.com · checked 31/10/2026
3. Google Analytics: Scopes of traffic-source dimensionssupport.google.com · checked 31/10/2026
4. Google Analytics: Reporting identitysupport.google.com · checked 31/10/2026
5. Google Ads: First click, linear, time decay, and position-based attribution models are going away (April 6, 2023)support.google.com · checked 31/10/2026
6. Google Ads: About attribution modelssupport.google.com · checked 31/10/2026
7. Google Ads: About data-driven attributionsupport.google.com · checked 31/10/2026
8. Google Ads: About Conversion Liftsupport.google.com · checked 31/10/2026
9. Yandex Metrica: attribution models and the May 20, 2026 changes (in Russian)yandex.ru · checked 31/10/2026
10. Yandex Direct: attribution models, significant sources and windows (in Russian)yandex.ru · checked 31/10/2026
11. Yandex Metrica: uploading offline data (in Russian)yandex.ru · checked 31/10/2026
12. Meta: About attribution models and attribution settingsfacebook.com · checked 31/10/2026
13. Meta for Developers: Ads Insights API metric availability updates (October 16, 2025)developers.facebook.com · checked 31/10/2026
14. Meta: Best practices to get started with Experimentsfacebook.com · checked 31/10/2026
15. Meta: Conversion lift percentfacebook.com · checked 31/10/2026
16. Meta GeoLift: Walkthroughfacebookincubator.github.io · checked 31/10/2026