Quick answer: how does AI measure advertising results after cookies and iOS tracking limits?
AI advertising measurement uses machine learning to estimate conversions that can no longer be tracked person by person, because of Apple's tracking prompt, browser blocking and consent rules. Platforms model the missing sales, Meta can predict incremental sales, and marketing mix models estimate channel impact. Treat all of these as estimates, then check them against your store's delivered net revenue.
Why Meta, Google and your store report different revenue
Each platform credits itself for any sale it touched, and each one sees a different slice of the customer journey. So the same ₹1 of revenue can be counted by Meta, by Google and by your store at the same time.
Your store only knows that an order happened. Meta and Google only know about the clicks and views on their own ads. Neither platform sees the other one's ads, so both can claim the same sale.
The data gaps make this worse. Some users never get tracked at all, so platforms fill those gaps with modelled conversions. A model is an estimate, and estimates from a party that is grading its own homework tend to run generous.
If you want the basics of how platform AI learns from these signals, our pillar on how artificial intelligence in advertising works covers it. This article is about checking the numbers it reports.
Cookies are not the real cause of signal loss in 2026
The "cookieless future" framing is out of date. Google decided not to remove third-party cookies from Chrome, and it then retired its Privacy Sandbox replacement.
In April 2025, Google confirmed it would neither deprecate third-party cookies in Chrome nor launch a cookie-choice prompt (Usercentrics). In October 2025, it announced the retirement of most remaining Privacy Sandbox APIs, including Attribution Reporting, Topics and Protected Audience. Third-party cookies are still on by default in Chrome, while Safari and Firefox have blocked them for years (Cookiebot).
The real causes of lost signal are:
- Apple's App Tracking Transparency (ATT), the iOS prompt that lets users refuse tracking across apps.
- Safari and Firefox blocking third-party cookies.
- Consent rules, where visitors decline tracking on a cookie banner.
- Ad blockers that stop the browser pixel from firing.
This changes what you should fix. Server-side tracking through Meta's Conversions API (CAPI) and clean consent setup matter far more than any "cookieless" product. Our Meta Pixel and Conversions API setup guide walks through it for Shopify and WooCommerce.
Meta reportedly changed what counts as a conversion twice in 2026
Meta's reported conversions moved in 2026 even where real sales did not. Reset your baseline before you judge any campaign on a before-and-after comparison.
| Date | What changed | What it does to your numbers |
|---|---|---|
| 12 Jan 2026 | 7-day view and 28-day view windows removed from the Ads Insights API (per Meta's developer blog, as widely reported) | Reports built on those windows lose conversions or show nothing |
| 3 Mar 2026 | Click-through narrowed to link clicks only (as widely reported, not yet confirmed from Meta's own posts) | Fewer purchases counted as click-through |
| 3 Mar 2026 | New engage-through category added (as widely reported) | Purchases within 1 day of a like, comment, share, save, profile tap or video view of 5+ seconds get their own column |
Reports on Meta's October 2025 developer update say the windows still available are 1-day click, 7-day click, 28-day click, 1-day engaged view and 1-day view (PPC Land). Several industry reports say the default in Ads Manager for website conversion campaigns is now 7-day click, 1-day engage-through, 1-day view. We could not confirm this from Meta directly.
So a ROAS (return on ad spend) target you set in 2025 was measured differently. If your click-through ROAS dropped in March 2026, some of that drop is reclassification, not lost sales. Compare like with like, and re-set targets on data from after 3 March 2026.
Industry reports say queries that ask for the removed Meta windows return empty data, not an error. A spreadsheet or third-party dashboard can quietly show zero view-through sales, or a wrong total, with no warning. Check every connector you use. Reports also say historical data recorded before 12 January 2026 can still be pulled, within Meta's data-retention limits.
Which AI measurement method fits your data
Pick the method based on how much history you have and what decision you are making. No single method is the truth, so most brands at ₹10 lakh+ a month need two of them.
| Method | What it estimates | Best for | Watch out for |
|---|---|---|---|
| Platform attribution (7-day click) | Sales near an ad click | Daily optimisation | Over-credits, counts the same sale twice across platforms |
| Meta incremental attribution | Sales Meta predicts your ads caused | Retargeting and warm audiences | Model-based, needs volume |
| Geo holdout test | Extra sales in regions with ads vs without | Proving a channel works | Needs weeks and clean regions |
| Marketing mix model (MMM) | Each channel's share of total sales | Annual budget splits | Commonly said to need 2-3 years of data |
When to use Meta's incremental attribution setting
Analytics vendor AdBeacon describes incremental attribution as an ad-set setting that uses machine learning to predict which conversions your ads caused, then optimises and reports on that basis instead of fixed windows (AdBeacon).
Our rule of thumb: try it on retargeting, where many buyers would have purchased anyway. Keep 7-day click on ad sets with thin conversion data, because a model needs volume. Never switch the setting in the middle of a test. Reports suggest the attribution setting cannot be edited once incremental is selected, so decide before launch.
If you have less than 2 years of data
Common industry guidance is that an MMM needs 2-3 years of history, which many D2C brands do not have. Start with geo holdout tests and platform lift tests instead. Pause or cut spend in a set of regions, keep it running in similar regions, and compare sales.
Once you have the history, Google's free, open-source Meridian MMM is a sensible starting point. Its Scenario Planner, announced in February 2026, lets non-technical teams test budget splits. Search Engine Land reports that it accepts Meta and TV data but that its deepest integrations favour Google, so check its Meta conclusions against your holdout results.
Reconcile platform revenue against delivered net revenue
Platform AI optimises toward whatever value you send it. If you send order value at checkout, it learns to find orders, not revenue.
This matters most in cash-on-delivery (COD) markets such as India, the Gulf and parts of South-East Asia. A "purchase" event fires when the order is placed. Later, some COD parcels are refused at the door and come back as RTO (return to origin), and some prepaid orders are returned. The platform never hears about it.
Tax adds a second gap. If your purchase value includes GST in India, VAT in the UK, EU and Gulf, or sales tax in the US, plus shipping, your ROAS looks higher than the money you keep.
How to find your true ROAS from reported ROAS
Use your store's own order data to work out what share of order value turns into delivered net revenue. Then apply that share to the platform's number.
Worked example: a brand spending ₹40 lakh a month on Meta
Example, round numbers only: Meta reports ₹1.2 crore revenue, a ROAS of 3.0. The store shows ₹1.5 crore of gross order value across all channels.
Step 1: Remove tax and shipping
Say tax and shipping make up ₹20 lakh of the store's gross order value in this example.
Net order value
₹1.5 crore - ₹20 lakh = ₹1.3 croreStep 2: Remove undelivered and returned orders
Say 20% of that value is cancelled, refused (RTO) or returned in this example. Use your own figure, it varies by category and region.
Delivered net revenue
₹1.3 crore × 0.8 = ₹1.04 crore, which is 69% of gross order valueStep 3: Apply the same share to Meta's number
Meta's ₹1.2 crore was counted at order placement, with tax and shipping inside.
Net ROAS
₹1.2 crore × 0.69 = about ₹83 lakh ÷ ₹40 lakh spend = 2.1If this brand's break-even ROAS on net revenue is 2.5, a "healthy" 3.0 is losing money.
Send the platforms better values
- Send purchase values without tax and shipping, so bidding uses real revenue.
- Send a later "delivered" or "paid" event, through CAPI or offline conversion uploads, once COD orders are collected.
- Recalculate your delivered share every month, and again around Diwali, Ramadan and Black Friday, when order mix shifts.
ScaleWin judges ads against break-even ROAS after product cost, shipping, tax and returns or RTO, which is this same calculation run daily.
Consent rules decide how much signal you get in each market
Consent requirements follow the visitor, not your company's home country. An Indian or Gulf brand with UK or EU shoppers is affected.
Industry reports say Consent Mode v2 became required for advertisers targeting users in the European Economic Area and the UK from 6 March 2024. We could not confirm the date from Google's own help pages. Those reports say it added two signals, ad_user_data and ad_personalization. When a visitor declines, Google receives less data to match conversions, so more of your reported results become modelled.
Practical steps:
- If you sell to the UK or EU, use a consent banner that passes both v2 signals to Google tags.
- Expect a higher share of modelled conversions from those markets, and judge them on store data.
- For other markets, check local privacy rules, which vary and keep changing.
For the wider regulatory picture, see AI advertising benefits, risks and regulation in 2026.
What most guides get wrong about AI attribution
Most ranking guides repeat three errors. Each one can push you into the wrong budget call.
- "Cookies are disappearing." Google has said Chrome will keep them, and Privacy Sandbox was retired in 2025. Fix CAPI and consent instead.
- Unsourced accuracy scores. Some guides give exact accuracy percentages for modelling methods with no source. Accuracy varies by account, data volume and market. Test it with your own holdout.
- Treating an AI tool as the source of truth. Every attribution tool, paid or platform-built, produces an estimate. Your bank and your delivered orders are the only ground truth.
Guides also tend to ignore COD. A model trained on placed orders in a market with high refusals will chase buyers who never pay.
Our order of checks before you move budget
Run these checks in this order before shifting spend between campaigns or channels. Skipping ahead is how brands cut campaigns that were working.
- Dates: is the comparison crossing 12 January or 3 March 2026? If yes, compare on the same attribution setting only.
- Connectors: does any dashboard use a removed window and show blanks or zeros?
- Values: are purchase values sent without tax and shipping?
- Delivered share: what share of order value became delivered net revenue this month?
- Page drop-off: do landing-page views roughly match link clicks? A big gap is a site or tracking problem, not an ad problem.
- Net ROAS vs break-even: only now judge the campaign.
- Size of the move: our rule of thumb is to shift no more than 20% of a channel's budget in a week on attribution data alone. Bigger moves need a holdout test.
Divide total delivered net revenue by total ad spend across all channels. If this blended figure moves less than 10% after a change, treat the platform's claimed win or loss as noise until more data arrives.
Our guide to AI reporting and analysis for Meta Ads covers how to automate these checks. ScaleWin's assistant pulls the last 7-10 days, checks learning-phase status and compares landing-page views with link clicks before it suggests a change.
Read next in this series
Key takeaways
- Google has said Chrome will keep third-party cookies, so "cookieless" is the wrong frame. Signal loss comes from iOS, Safari, Firefox, consent rules and ad blockers.
- Meta reportedly changed what it counts as a conversion twice in 2026. Reset your baseline ROAS before you cut any campaign.
- In COD markets, a platform "purchase" fires before refused deliveries and returns. Reconcile to delivered net revenue, not orders.
- Under 2 years of data? Run geo holdout tests before you attempt a marketing mix model.
- Send purchase values without tax and shipping, so platform AI bids on real revenue.
Artificial intelligence advertising: frequently asked questions
Why did my Meta ROAS drop in March 2026 when sales did not?
Should I use Meta's incremental attribution on every ad set?
Can I run a marketing mix model with one year of data?
Should purchase values sent to Meta and Google include GST or VAT?
Are third-party cookies going away in Chrome?
Sources
- Usercentrics: Google third-party cookies decision
- Cookiebot: Google third-party cookies status
- PPC Land: Meta restricts attribution windows in Ads Insights API
- AdBeacon: Meta incremental attribution vs 7-day click
- Google: Meridian marketing mix model open to everyone
Let ScaleWin read your ad data every day
ScaleWin checks your Meta and Google Ads daily, judges every ad on profit and tells you exactly what to change.
Start free, no card



