A sequential give-back from a peak is not a margin step-down

What happened

Two consecutive AppLovin reports (2026-08-05 and 2026-09-10) carried the claim:

"adj EBITDA margin 85% → 83%, and Q3 is guided at ~83% — the step-down is guided as permanent."

It was listed as a headline red flag and it fed a break trigger (EBITDA margin <82% for 2 quarters). The Q2 2026 8-K says otherwise:

	Q2 2025	Q2 2026	Q3 2026 guide
Adj EBITDA margin	81%	84%	83%

The margin expanded 300bps year over year. The "85%" was a sequential high-water mark; the guide was ~100bps below the most recent quarter and ~200bps above the year-ago comparable. The GAAP operating line — which needs no management definition at all — was flat-to-up across the entire window (76.1 → 76.8 → 76.9 → 78.2 → 77.7%).

The error survived a full re-analysis because /analyze-from-before re-tested the claim as written rather than re-deriving it from the filing.

Why the error is easy to make

1. A sequential peak is the most available anchor — it is the number in the last report.
2. "Guided flat-to-down" pattern-matches to impairment on any name that has just missed.
3. Adjusted metrics have no fixed definition, so a vendor's reconstruction (Yahoo's Normalized EBITDA) and the company's own Adjusted EBITDA differ by hundreds of bps. Neither matches the other, and mixing them across quarters invents trends.
4. A decelerating top line makes a margin story feel confirmed. APP's sequential revenue growth genuinely halved (+11.1% → +4.4%). That is real — but it is a growth problem, and it got laundered into a margin problem because both appeared in the same quarter.

The rule

Before calling any margin change structural, run three tests in this order:

1. Year over year, same quarter. Seasonality and sequential noise vanish. If YoY is expanding, it is not impairment — full stop.
2. The GAAP operating line. Management cannot redefine it. If GAAP op margin is flat while the adjusted metric "fell," the change lives in the adjustment, not the business.
3. Same definition on both ends. Never compare a company-adjusted figure to a vendor's normalized one, or a guide (company definition) to a vendor's history.

A give-back to the second-best quarter on record is not a step-down. Say "off the Q1 peak, still +300bps YoY" — never "the step-down is permanent."

Generalisation

This is the margin form of a wider error: anchoring a trend claim on the most recent extreme rather than on a like-for-like comparable. It shows up identically in growth rates (a sequential peak quarter becoming the baseline), share gains, and take rates. See [[pattern-self-defined-segment-migration-flatters-growth]] for the mirror case, where the definition moves instead of the anchor, and [[principle-primary-source-beats-vendor]] — the 8-K settled this in one fetch, after two reports had argued it from vendor data.

Cost of this one

The claim was load-bearing: it was cited as the reason conviction sat at 6.0-6.5 rather than higher, and it framed the August −21% drawdown as justified by fundamentals. It was wrong, and it was checkable in a single primary-source fetch both times.