> ## Documentation Index
> Fetch the complete documentation index at: https://promptwatch.com/docs/llms.txt
> Use this file to discover all available pages before exploring further.
# Why did my score drop?
> How to tell a real visibility drop from statistical noise, and the setup changes, prompt edits, model additions, that move your score without the market moving.
Your Visibility Score fell and you need to know if it means anything. Most drops turn out to be one of five things, and they're worth checking in this order, because the most common causes are also the fastest to rule out.
One fact drives most of them: the score is an average across analyzed responses, and a response that doesn't mention your brand counts as zero. Anything that adds zero-scoring responses to the average pulls the number down, whether or not any AI model changed its mind about you. See [Visibility Score](/docs/academy/visibility-score) for the full computation.
### 1. Noise: your sample is small
With a handful of prompts on one or two models, each day's average rests on very few responses. One answer that skips your brand can move the number visibly, and it will often bounce back the next run without you doing anything.
**How to check**: widen the date range. A dip that disappears in the 30-day view was noise. If your day-to-day line is jagged in general, the durable fix is more prompts and more models, so single answers stop mattering.
### 2. You changed the prompt set
Adding prompts changes the denominator. Ten new organic prompts where you don't appear yet add ten zero-scoring responses per model per run, and your average drops the day they start running, even though your position on every existing prompt is unchanged. Removing prompts where you scored well has the same effect from the other side.
**How to check**: compare the drop date against when prompts were added or removed. To confirm, filter a monitor view to a tag, or the dashboard to a [topic](/docs/academy/organizing-your-setup), covering only your long-standing prompts (the project dashboard slices by topic, not tag): if the score is flat there, the "drop" is your baseline resetting to include the new, harder prompts. That reset is healthy, it means you're now measuring ground you want to win.
### 3. You changed the model set
Models score very differently, and the dashboard blends them. Adding a model where you're weak drops the blended average from the day it starts collecting, with no change on any platform you were already tracking.
**How to check**: filter by each model separately. If every individual model is flat and only the blended number fell, the model mix changed, not your visibility. See [choosing your models](/docs/academy/choosing-your-models).
### 4. A real ranking change on specific prompts or one platform
Once setup changes are ruled out, the drop is real somewhere, and it is almost never everywhere at once. A model update, a competitor's new content, or a shift in which sources a platform cites usually hits specific prompts on specific platforms.
**How to check**: filter by model to find which platform fell, then sort your prompts by visibility to find where. Open the responses for those prompts and compare recent answers against older ones: you're looking for who replaced you and which sources the answer now cites. The [competitor heatmap](/docs/academy/competitors-and-heatmap) shows the same picture per competitor per model in one view.
### 5. A competitor entered the comparison
New competitors change relative metrics. If you added a competitor to your project, their mentions join the Share of Voice denominator and your share falls, correctly, because the market you're measuring got more crowded. Note the direction of cause here: your competitor list changes [Share of Voice](/docs/academy/share-of-voice), not the Visibility Score, which is computed per response independent of who else you track.
**How to check**: if SOV dropped but Visibility Score held, look at the competitor list first. If both dropped, a competitor is actually taking your place in answers, which is case 4.
### When a drop is expected
Scores drift down without any mistake on your side: models update, sources rotate, and competitors publish. A slow decline across many prompts on one platform usually means that platform's source pool moved away from you, which is content work, not troubleshooting. Track the trend weekly rather than the daily wiggle, and when a real gap shows up, the [Action Board](/docs/academy/action-board) and [Content Gap](/docs/academy/content-gap) are where you turn the finding into work.