Ethan Smith puts parametric AEO impact at two to five months
Analysis of LLM training disclosures sets a floor on how fast model weights can reflect new content.
Ethan Smith, writing on LinkedIn on 21 October 2026, reports that AEO work aimed at a model's parametric knowledge takes at least two to five months to show up. The figure comes from his team's analysis of public disclosures from LLM companies about training cycles and data cutoffs.
Smith's earlier line was that AI optimisation can move in days. He now qualifies that. Fast movement applies to answers that use search and return citations. Depending on the prompt set, that covers between 10% and 80% of answers.
The rest draw on parametric knowledge, meaning what the model's weights already hold. New content cannot reach the weights until the next model trains on it. Cutoffs run from a few weeks to a few months before release, and new models ship every few months.
How Smith says to measure it
Smith lays out three steps. Watch for discrete jumps between model versions rather than weekly drift. Check each model's training data cutoff before crediting a change to your own work. Run the test through APIs with web search disabled, pin the model version, and collect at least 50 responses per prompt.
The cutoff date decides which of your work could have reached the training set. Anything published after the cutoff cannot affect that model. Smith notes the response distribution stays stable on a pinned model as long as prompts, system instructions and sampling settings are fixed, so the baseline test only needs to run once per version.
What this adds to citation-based AEO
We think the useful move here is to run both measurements in parallel. Track where you are cited now in live answers. Also run the web-disabled API test on a pinned model to see what the weights already say about your brand. The two readings move on different clocks and answer different questions.
Smith calls the second track parametric knowledge optimization, or PKO, and frames it as a complement to citation work, not a replacement. He is asking openly whether anyone has found ways to influence the training set itself. That part is unsettled, and worth treating as such before acting on any single tactic.