How to Diagnose Over-Attribution in Meta and Google Ads

Over-attribution is diagnosed with orders as the denominator. Sum every platform's claimed conversions for a period without deduplication and divide by the orders the store shipped: that is the claim ratio, and anything above 1 is the amount by which the platforms collectively claimed more sales than exist. Then pull each platform under its default window and under a one-day click window, and the gap is the share of its claim that rests on touches days before the purchase. Four checks, one afternoon, no vendor.

Why it happens, without anyone lying

Each platform reports conversions it can associate with exposure to its own inventory, inside a window it defines, under a rule it sets. Meta's default is 7-day click and 1-day view. Google Ads defaults to data-driven attribution with a click-through window set per conversion action, 30 days unless changed. Each counts an order it touched, and nothing subtracts an order another platform also counted. The Price of Being Found describes the result as an incentive structure that does its work without anybody being wicked: the seller of the media also measures whether the media worked, and there is no audit. Ad platform numbers don't match is the short version.

Check 1: the claim ratio

Orders from the store for a period, say last month. Meta purchases, Google Ads conversions, TikTok, Pinterest, affiliates, and the email platform's attributed orders for the same dates, each under the window the account actually used. Sum without deduplicating; divide by orders. The book's guidance: track the level once and the movement always, because a jump usually means a definition changed. Meta narrowed its click definition in March 2026, per the book's account, and any account comparing across that date is comparing different quantities. The one-hour claim ratio audit has the six steps.

Check 2: the window pull

For each platform, pull the same period under the default window and under a 1-day click window. The difference is the share of the platform's claim that comes from touches more than a day before the purchase. It is not proof of over-attribution on its own; some of those earlier touches caused the sale. It is the size of the population the platform is claiming on the weakest evidence, and comparing the gap across platforms tells you whose claim leans hardest on it.

Check 3: the last-touch cluster

Look at which channels take credit closest to the purchase: branded search, retargeting, email. These collect the most credit and, in the experiments the book quotes, tend to cause the least. At eBay, brand-keyword ads scored as the best channel until they were switched off and 99.5% of the forgone paid clicks came back through natural search. A platform whose claimed conversions are concentrated in lower-funnel placements is the first candidate for a holdout, because the gap between credit and cause is largest where the purchase was already on its way.

Check 4: coverage on the other side

Over-attribution has a mirror image. GA4 or the store's analytics may attribute far fewer conversions than the orders shipped, because a share of sessions arrives with no resolvable source. Coverage is attributed conversions divided by orders, and the book's own census found 48.6% of sessions at one company with no resolvable source. A low coverage rate inflates every cost-per-order figure computed from analytics, in the opposite direction from the platforms' inflation, and both distortions land on the same reallocation decision. The Black Friday number nobody checks covers the arithmetic.

What the four checks cannot tell you

They size the inflation. They do not say which platform holds the double counts, and they do not say what any channel caused. A claim ratio of 1.8 means credit is inflated by about 80% of orders across the account; it does not distribute the excess. That needs a design that could have found nothing: a holdout on the most suspect channel, qualified first against the smallest lift your revenue noise allows, or a causal read on the GA4 export that estimates each channel's incremental contribution.

What to do this week

• If you own the budget: run check 1 on last month.
• If you have to defend the number: run checks 1 and 2 for the last three months and present them as a series, with the March 2026 definition change marked.

As of 9 September 2026. Platform default windows as documented on that date; the claim ratio, the coverage census and the eBay experiment are from The Price of Being Found (Edition 2.10), Chapters 8 and 9, with the book's caveats.