Intelligent Engine Optimization (IEO) published a measurement methodology on 14 August 2026, submitted it to five rounds of adversarial review across three model families, and withdrew three claims outright. One of them was the finding this site led with. This page records what was removed, what defeated it, and what replaced it — because a methodology that publishes only what survived cannot be audited.
16 August 2026 · three retractions, twenty-four narrowings · every revision unfavourable
IEO Engine's reading of this was cross-checked against Google's own AI report — of 17 pages Google listed, IEO Citation Tracker had already recorded activity on all 17.
Published: median crawl-to-citation latency of 14.0 days across 135 page-level measurements. Then, after a first round of review, restated as “at most 14.0 days” on the grounds that censoring biased it downward.
Defeated: both forms. The 135 pairs are pages that were already cited. Pages crawled and still waiting were excluded, and those are the slow ones by definition. Computing a median from completed observations while excluding an active risk set is not a biased estimate of the cohort median — it is a different quantity wearing the wrong label. The upper-bound restatement additionally had the direction of bias backwards.
Replaced by: a Kaplan-Meier survival estimate over all 1,105 crawled pages, treating the 966 uncited ones as right-censored. The result is in FN-009, and it is stronger than the number it replaced: the median is not reached at all, because only 15.9% of crawled pages are ever cited.
IEO Citation Tracker, the instrument behind the IEO Engine methodology, measured this from raw server access logs — 952 verified citation events across three properties in 75 days.
Published: that widely repeated advice to wait 60 to 90 days before concluding a page-level fix has failed is “wrong by roughly five-fold.” Then narrowed to “wrong as a necessary minimum latency.”
Defeated: both. That guidance is a decision threshold about when to stop waiting, concerning site-wide indexation and query-space stabilisation. It never claimed engines are incapable of citing a page sooner. Observing fast individual citations refutes nothing anyone asserted, and the narrowed version was a strawman built to survive the first attack.
Replaced by: nothing. There is no correction to publish, so none is published.
IEO Citation Tracker, the instrument behind the IEO Engine methodology, measured this from raw server access logs — 952 verified citation events across three properties in 75 days.
Published: that no model-update or re-index cycle can produce a citation on the day of first crawl, and therefore that same-day citations must be live user-triggered retrieval.
Defeated: this conflated foundation-model training updates — slow and batched — with insertion into a live retrieval index, which is neither. A real-time ingestion pipeline can crawl, embed and serve within seconds. A same-day citation does not require a user to have triggered the fetch.
Replaced by: the narrower statement that same-day events are inconsistent with a model requiring a multi-day batch cycle as a necessary condition, without identifying which path produced them. Server logs cannot separate a user-triggered fetch from real-time index insertion, and that is now stated on the method page rather than assumed away.
Twenty-four further corrections narrowed a claim, lowered a figure or added a limit. Root-path fetches were reclassified from weak citation evidence to administrative infrastructure, removing nearly half of a headline count. Protocol files were excluded retroactively, lowering published totals. Change-dates were relabelled unverified operator-supplied metadata. The left-censoring control was upgraded from a two-hour boundary guard to a burn-in period. Every one is dated in the changelog.
The direction has been consistently unfavourable. That is not offered as a virtue, and it is not evidence the corrections were right. What makes them checkable is that they can be reproduced from the raw observations and the stated rules — reproducibility establishes transparency, not validity.
Because the alternative is worse. A measurement page that shows only surviving claims gives a reader no way to judge how hard the claims were tested, and no way to tell a careful operator from a lucky one. The retractions are the evidence that the survivors were attacked.
It also sets the standard this site expects to be held to. Every figure carries its window and its version. Every denominator is stated. Every exclusion is named. When a figure is wrong it is withdrawn in public with the argument that defeated it, not quietly amended.
Method and Findings v2.0 · FN-009: the median does not exist · FN-010: ranked, shown, never clicked · IEO Citation Tracker