Why did my brand's AI visibility suddenly drop, and how can I tell if an AI model update caused it?
How we diagnose a sudden drop in a brand's AI visibility and tell whether a model update caused it: the checks, the order, and what we tell clients.
A sudden drop gets a panicked message, usually with "did ChatGPT change something?" in the subject line. Sometimes yes. But we have a rule: rule out our own side before blaming an engine. An engine update is outside anyone's control, whereas a blocked crawler is fixable today.
What we do in the first hour
Check the measurement. Has the prompt set changed? Were brand aliases or the Brand Book edited? Did a competitor get added or removed? A configuration change looks exactly like a visibility drop. We then compare a full week against the previous week rather than reacting to one day. Promptwatch's A/B period comparison makes that quick.
Check the pattern across engines. Promptwatch runs prompt panels daily per model, so we can see whether one engine dropped or all did. One engine falling while the others hold usually means that engine changed. A drop everywhere points toward us.
Check the category. We open the competitor heatmap and share of voice. If everyone fell, the engine probably changed. If only the client fell, a competitor gained.
The checks on our side
Crawler access. We open Agent Analytics, which shows IP-verified crawler logs, and look at GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot and the rest for the days around the drop. Did errors appear? Did visits stop? Deploys, CDN rules, firewall updates and bot-protection settings are common culprits, and nobody tells the marketing team. Crawler logs start on the Professional plan ($245 a month), so on Essential this check is not available inside the tool.
Page changes. We compare citations for the top prompts before and after. Page Tracker follows specific URLs. If a once-cited page now redirects, was rewritten or disappeared from the sitemap, that is the story. The weekly site audit reports what changed since the last run.
Source changes. Engines cite forums, review sites, news and video. If a source dropped out, our citation share falls with nothing wrong on our site. Reddit, YouTube and LinkedIn citation reports show this at source level. Promptwatch's own data was cited by Search Engine Journal on the August 2026 drop in Reddit citations, which shows these swings are real and visible to anyone watching.
When it is probably a model update
The signs: one engine, broad category movement, no change on our site, no crawler errors, no page changes. Then we tell the client plainly that an engine-level change is the likeliest explanation, that we cannot prove it because engines rarely publish what changed, and what we are doing in response: adapt content to the new behavior, keep tracking and re-check in two weeks.
When it is probably us
The signs: all engines fall, crawler errors appear, a page change coincides with the drop. Then we fix it, and the fix is usually small: a robots rule, a redirect, a template bug.
What we tell clients
- We will not claim a model update caused it without evidence from the checks above.
- We will not promise recovery on a date.
- We will report what we ruled out.
That last point matters. Even when we cannot name the cause, a clear list of what it is not has real value.
Keeping a change log
The cheapest prevention is a log of site changes, robots edits and publishing dates kept next to the data. On agency plans, which start at $199 a month with unlimited projects and ten seats, we keep that log per client.
If you want to run these checks yourself, start with a free Promptwatch Explore account for the baseline, and move to a paid plan when you need the cross-engine and crawler checks.