The platform that sells the ad also reports how well it worked

Picture the monthly report for a 12-location studio brand. Google Ads shows one number of conversions. Meta shows a different number of leads. The booking system shows fewer new intro classes than either, and the front desk remembers fewer still. Every one of those systems is working as designed, and the two biggest numbers come from the companies that also sold the ads.

When you buy media from Google or Meta, the same company runs the auction, places the ad, counts the result and decides how much of the result to credit to itself. We use both platforms every day for the businesses we work with, and we rely on their reporting. We also check it, because the platforms have told us, in their own help pages and court filings, how their numbers are made.

This part covers four documented examples, and each one is a reason to keep your own count.

A conversion Google predicts

Google Ads reports two kinds of conversions in the same column. Some it observed: a person clicked an ad and Google saw the form fill or purchase. Others it modeled. Google's help page explains: "Modeled conversions use data that doesn't identify individual users to estimate conversions that Google is unable to observe directly."

"When Google surfaces modeled conversions in Google Ads, we're predicting attributed conversions."

The same page says the model "doesn't determine whether or not a conversion happened." It estimates which conversions to attribute to your ads. And "In the 'Conversions' column, Google reports both modeled and observed conversions," at the same level of detail, so you cannot tell them apart in an ordinary report. Google also notes that "Conversion values are subject to retroactive increases for a period of several days as data modeling is finalized." The number you read on Monday can be higher by Friday.

There is a real reason for this. Browsers, phones and privacy rules now hide a lot of what used to be observable, and modeling fills the gap. But the company predicting the credit is the company being credited, which is worth remembering when the report lands on your desk.

Meta changed its windows

An attribution window is the rule for how long after someone sees or clicks an ad a conversion still counts toward it. Change the window and the same campaign reports different results.

On October 16, 2025, Meta's developer blog announced that from January 12, 2026, "7-day view (7d_view) and 28-day view (28d_view) attribution windows will no longer be returned" through the Ads Insights API. One-day click, 7-day click, 28-day click, 1-day engaged view and 1-day view windows "will continue to be available." Meta gave a general reason: "We continuously review the metrics we offer advertisers to ensure that they are accurate, intuitive, and valuable."

This may well be a sensible cleanup. View-through credit, where someone saw an ad, did not click, and converted weeks later, was always the softest number in a report. The point for an owner is narrower. A platform can change how results are counted, on its own schedule, for every advertiser at once. Any report that pulled those windows through the API changed on January 12 even though nothing changed at the front desk.

When the numbers were wrong

Two cases show the stakes. Neither ended in a court ruling against Meta on the merits, and we describe them only as the filings do. This is not legal advice.

Video metrics. Advertisers sued Facebook in 2016 over two video metrics: Average Duration of Video Viewed and Average Percentage of Video Viewed. The lawsuit alleged Facebook overstated viewership "by as much as 900%," as class counsel described it. Facebook agreed in October 2019 to pay $40 million in cash. As reported by ppc.land, Facebook acknowledged the calculation error but denied that it affected what advertisers spent. Judge Jeffrey S. White granted final approval on June 26, 2020. The class covered anyone who bought Facebook video ads between February 12, 2015, and September 23, 2016.

Potential reach. This one is still open. Meta's own quarterly report for the period ending September 30, 2026, describes it: "Beginning on August 15, 2018, multiple putative class actions were filed against us alleging that we inflated our estimates of the potential audience size for advertisements, resulting in artificially increased demand and higher prices." The court certified a class in March 2022. The Ninth Circuit affirmed in part and reversed in part in March 2024, and the Supreme Court denied Meta's petition. Meta then moved to compel arbitration, lost, and appealed that ruling on December 3, 2025. In Meta's words, "The matter is stayed in district court pending resolution of our appeal." It has not been tried, and it has not been settled. The claim is an allegation, and Meta is contesting it.

Source: Gibbs Law Group, Meta 10-Q, Q3 2026, Meta for Developers

Platform numbers are the seller's estimates. The record above shows the counting method changing on the seller's schedule, and two lawsuits over whether the numbers were right.

Keep your own definition and reconcile monthly

We wrote in How do you measure marketing when an agent did the shopping? about tracking AI assistants and reading direct traffic. This is the step before any of that: decide what counts, write it down, and compare it with the platforms once a month.

1. Write down what a lead is

   For example: a form fill or call with a real name and a working phone number, from someone who does not already hold a membership. One sentence, dated, kept where everyone can read it.

2. Write down what a join is

   A paid membership or package, counted in your booking or billing system, on the day the first payment clears. Not a trial, not a free class.

3. Count both in your own system

   Your booking system and your staff record the source of every lead and join. UTM-tagged links help it do that for every ad you run.

4. Pull the platform numbers on a fixed day

   Take Google's conversions and Meta's leads for the same calendar month, a week after month end, so Google's modeling has settled.

5. Put them side by side

   Platform conversions, your leads, booked intros, shows and joins in one sheet you own. Note the gap between each step.

6. Investigate changes in the gap, not the gap itself

   A steady gap is normal, since platforms count differently. A gap that suddenly widens means a method, a window or a tracking tag changed somewhere. Find out which before you move budget.

The definitions matter more than the spreadsheet. If a lead means something different every quarter, no comparison holds. When a platform changes a window or adds modeling, your definition stays put, and the change shows up as a moving gap you can see.

For the arithmetic of reading the numbers once you have them, Reading cost per lead without fooling yourself covers account averages, pooling and small samples.

Earlier in the series, The EU fined Meta over other advertisers' data looked at what a regulator found about data inside one platform. Next, in Part 9, we follow the money: Three companies take most of the ad money.

---

Sources

• Google Ads Help, "About modeled online conversions," accessed October 24, 2026. support.google.com
• Meta for Developers, Chris Cutlip, "Ads Insights API Metric Availability Updates," October 16, 2025. developers.facebook.com
• Meta Platforms, Form 10-Q for the quarter ended September 30, 2026, filed October 22, 2026, Legal Proceedings. sec.gov
• Gibbs Law Group, "Groundbreaking $40 Million Facebook Video Advertising Settlement Receives Final Approval," June 30, 2020. classlawgroup.com
• ppc.land, "Facebook pays $40 million in cash to settle video measurement class action," October 8, 2019. ppc.land