In a third of competitor reports, the biggest number is a rival's

The twelve

In each of these reads, the headline creative count — the biggest number in the report — described a competitor's advertising rather than the brand's own.

Source: our own 35 attributed competitor reads, re-audited 2 October 2026. Each figure is a count of distinct creatives decoded, matched to the advertiser it was pulled from. Categories only; no brand is named.

Why the small ratios are the dangerous ones

The lending case is obvious once stated: a brand with 3 creatives cannot plausibly own a figure of 96, and anyone reading carefully would stop.

The innerwear case is the one that actually gets through. The brand had 385 creatives and the largest figure in the report was 402 — a rival's, and only 4% higher. Nothing about 402 looks wrong next to 385. There is no implausibility to catch it on. Five of the twelve sat within 2× of the brand's own figure, which is well inside the range where a number simply looks like it belongs.

So the defence cannot be a sense of what looks reasonable. It has to be structural: every figure carries the advertiser it came from, or it does not get used.

How it happens

A competitor read produces one bundle per category holding several advertisers — the brand and its rivals — each with its own counts. Summarise that bundle by taking its largest value and you get a figure that is correct in isolation and unattributed. Put it in a sentence beginning with the brand's name and it becomes false.

Nothing is corrupted at any point. The scrape is right, the hashing is right, the number is right. The error is entirely in the sentence written about it, which is why it survives review so well: there is no bad data to find.

What it cost us

Our own corpus page stated that live library sizes ranged up to 2,247 ads, implying that figure described a brand we work with. It was a rival's library in the innerwear category. We removed it, replaced the range with a smaller figure we could attribute, and only put 2,247 back once every advertiser in that bundle had been matched to its own numbers — where it now appears, correctly, as a competitor's.

Two more figures went the same way in the same week: a deduplication rate of 94% that was the video-only rate rather than the library rate of 89.5%, and a cost-per-order improvement of 14% that had no source at all, because no spend or conversion data exists anywhere in this dataset. One wrong attribution, one wrong denominator, one invention.

The three checks that would have caught all of them

• Name the advertiser beside every figure, always. Not in a methodology appendix — in the same row as the number. An unattributed figure is not a fact yet.
• State the denominator. “89.5% of 352 media files” cannot be confused with the video-only rate. “94% of files” can, and was.
• Check prose against its own table. Both of the errors we found in our own writing sat in sentences summarising tables whose every cell was correct. The table is usually fine. The sentence about it is where things break.

A measurement is not more reliable for having been computed carefully. It is reliable when the claim attached to it survives being traced back to the source, and that is a different activity from measuring.