How to Diagnose an AI Visibility Drop in Promptwatch
Your Promptwatch Visibility Score fell. Rule out noise, prompt changes, model changes, and competitor changes in that order, then find the prompts that really moved.
Most visibility drops are not drops. They are the denominator changing. Somebody added twenty new prompts on Tuesday, or switched on a model where the brand is weak, and the blended average fell the same day. Nothing in any AI answer changed. The fix for that kind of drop is a sentence in the weekly report, not a content sprint.
This guide is the order we check things in when the Promptwatch Visibility Score falls. It follows Promptwatch's own score drop guide and its page on reading performance over time, both read October 15, 2026. The cheap checks come first because they explain most cases.
Know what the score averages
The Visibility Score is a 0 to 100 prominence measure averaged across analyzed responses, and a response that doesn't mention you counts as zero. That one fact drives most false alarms. Anything that adds zero-scoring responses to the pool pulls the number down whether or not a model changed its mind about you.
Check 1: is the sample too small?
A monitor with 5 prompts on 2 models produces 10 responses per run. One changed answer moves the average by 10 percent. Widen the date range. If the dip vanishes in the 30-day view, it was noise.
Promptwatch's rule of thumb is that a window should contain at least four or five scheduled runs, and that three or more consecutive runs moving the same way is a trend. One or two days of movement in a daily-tracked project is normal variance. If your line is jagged every week, the long-term fix is more prompts and more models so single answers stop mattering.
Check 2: did the prompt set change?
Ten new prompts where you don't appear yet add ten zero-scoring responses per model per run. Your average drops the day they start, even if your position on every old prompt is unchanged. Deleting prompts where you scored well does the same thing from the other side.
Compare the drop date against when prompts were added or removed. Then filter a view down to only your long-standing prompts: a tag on a monitor view, or a topic on the project dashboard (the dashboard slices by topic, not tag). Flat there means the baseline reset to include harder prompts. That is a healthy reset, and it should be reported as one.
If you use the API or MCP, the project events endpoint returns prompts added or deleted and action items completed, grouped by UTC day. Joined against the visibility time series, it answers "what did we change that day" without anyone's memory.
Check 3: did the model set change?
Models score very differently and the dashboard blends them. Add a model where you are weak and the blended number falls from its first collection day. Filter by each model in turn. If every model is flat and only the blend moved, the mix changed, not your visibility.
Check 4: did a competitor join the comparison?
Adding a competitor changes Share of Voice, because their mentions join the denominator. It does not change the Visibility Score, which is computed per response regardless of who else you track. So the pattern to look for is Share of Voice down while Visibility Score holds. If both fell, a competitor is actually replacing you in answers, which is check 5.
Check 5: the real drop
Once setup changes are ruled out, something real moved, and it almost never moved everywhere at once. Filter by model to find the platform. Sort prompts by visibility to find where. Then open the responses for those prompts and compare recent answers against older ones. You are looking for two things: who replaced you, and which sources the answer now cites.
The Heatmap shows the same picture per competitor per model in one grid. The Compare previous checkbox on the Brand Visibility chart overlays the preceding period on the Line and Bar views, which is what you want before you put a number in a report. One more cut worth making: if the Visibility Score fell but Share of Voice held, you still appear as often, just less prominently. If both fell, you are losing answers outright.
Turn the finding into work
A slow decline across many prompts on one platform usually means that platform's source pool moved away from you. That is content work. In Promptwatch it lands on the Action Board or in Content Gap, and the rewrite itself can go through Content Refresh with the competitor pages that now win the prompt as references.
This is the part a monitoring-only tool can't finish. A tracker can tell you the line bent. Diagnosing it needs response-level history, per-model filters, the citation sources behind each answer, and a dated log of your own setup changes. Promptwatch keeps all four in one project, which is why we rank it first for teams that have to explain a drop to someone else.
Plans: Explore is free (10 ChatGPT prompts). Essential $95/mo, Professional $245/mo, Business $579/mo. Agency Kick-off $199/mo, Growth $399/mo, Scale $799/mo. Review: Promptwatch. Directory: all tools.
FAQ
My score dropped the day I added prompts. Is that bad?
Usually not. New prompts where you don't appear yet count as zeros. Filter to your long-standing prompts to confirm.
Does adding a competitor lower my Visibility Score?
No. It lowers Share of Voice. The Visibility Score is computed per response independent of the competitor list.
How many days before I call it a trend?
Three or more consecutive scheduled runs in the same direction, inside a window that holds at least four or five runs.