Why your AI share of voice dropped 40% overnight (and what to do about it)
When a model gets retrained your brand can disappear in a week. Here is the playbook we use with enterprise customers to recover, and how to prevent it next time.
It happens roughly every six weeks. A frontier model retrains, the embedding space shifts, and a brand that was cited in 62% of relevant prompts on Monday is cited in 19% on Friday. No announcement, no changelog entry, just a quiet redistribution of attention.
We have walked thirty-some enterprise customers through this recovery. Here is the playbook.
Hour 1: triage
Pull share of voice for every tracked topic cluster, segmented by engine. The drop is almost never uniform. You are looking for the cluster where you lost the most and the engine where the drop is sharpest. That intersection is your investigation surface.
Day 1: find who replaced you
Run the same prompts and capture the new citation list. There is always a new entrant or an old competitor who suddenly got promoted. Read their cited pages with the same eye a model uses: first paragraph, structured data, named author, original numbers, recency.
In 80% of cases the replacement published a single piece of original research in the last 60 days that we did not match.
Week 1: ship a counter-source
Do not try to recapture every prompt. Pick the three highest-value clusters and ship one strong, quotable, original page per cluster. The shape that works is consistent: a clear question in the title, a one-sentence answer in the lede, a dataset or first-party number in the body, a named author with a credential.
Promote it where the model will see it: your own site with clean schema, your changelog or research index, a syndicated mention on a domain the model already trusts.
Week 2-4: monitor, iterate, harden
Re-pull share of voice daily. You will usually see partial recovery inside 10 days as crawlers re-index and the next minor model update absorbs the new source.
While that happens, audit the pages you lost from. The ones that fell out of citations usually share a flaw: outdated stats, weak ledes, anonymous bylines, or buried claims. Fix them with the same discipline you applied to the counter-source.
Prevention is mostly hygiene
The customers who weather retraining best do four things continuously: they publish original numbers monthly, they revisit their top 20 cited pages every quarter, they keep author bylines and credentials current, and they track share of voice per engine per cluster so they see drops on day one instead of day fourteen.
A 40% drop is recoverable. A 40% drop you did not notice for a month usually is not.
