When ChatGPT names or cites your brand less, the instinct is to start editing pages. Before you do, find out whether the cause is on your side at all. Mid-2026 research by Peec AI documented AI engines silently changing how they expand a prompt into sub-queries and which sources they prefer, which means a drop can happen with no change to your content. This is a five-step diagnostic to tell a real, your-side regression from a model shift you should simply ride out.
It is the practical companion to why your AI visibility changes even when your content hasn't. That piece is the evidence; this is the procedure.
A drop that hits your whole category is the model changing its mind. A drop that hits only you is a problem on your side. Diagnose which before you touch a single page.
Step 1: Establish a baseline before you diagnose
You cannot diagnose a drop you can't measure. AI answers are probabilistic and shift between observations, so a single reading tells you nothing. Before anything else, make sure you have a baseline built from many prompts, sampled repeatedly, across every engine, read as a trend line rather than one number.
If you do not already have this, you cannot yet tell a drop from ordinary answer volatility. Start sampling now and treat the first sustained pattern, not the first dip, as your signal. Everything below assumes you are reading a trend, not a snapshot.
Step 2: Split the drop into three signals
"Visibility went down" is too coarse to act on. Separate it into three measurable signals, because each points at a different cause:
| Signal | What it measures | A drop here suggests |
|---|---|---|
| Mentions | How often your brand is named in the answer text | A naming/entity problem, or a ghost citation pattern (cited but not named) |
| Retrieved sources | How often your domain is pulled in during retrieval, including silent, uncited uses | A retrieval/reachability problem, or the engine reordering which domains it trusts |
| Fan-out terms | The sub-queries the prompt actually triggers | The model changed how it decomposes the question |
Watching all three at once is the whole point: a mention drop with steady retrievals is a very different problem from retrievals falling off a cliff. A single citation-rate figure blurs them together.
Step 3: Check breadth. Did it hit only you, or everyone?
This is the fastest way to assign blame. Compare your movement against your competitors and the broader category over the same window.
- Category-wide shift (rivals and comparable sites moved too): almost certainly the model. When an engine changes its query fan-out or source preferences, it reorders the whole field, not just you. Peec AI's Reddit and arXiv examples moved entire source categories at once.
- You-specific shift (the category held steady, you fell): look for a your-side cause. Something changed what the engine can fetch, parse, or trust about your pages specifically.
Without a competitive baseline you will misread a category-wide model change as a personal failure, and burn a sprint fixing pages that were never the problem.
Step 4: Line the drop up against model releases
Timing is evidence. Note when the drop began and check it against known model updates. ChatGPT 5.6 (the Sol/Terra/Luna tier family) shipped on 9 July 2026, and Peec AI reports its Luna tier fans out very differently from 5.5, leaning heavily on the site: operator, the word "official", and year modifiers, with more multi-step retrieval.
If your drop clusters around a release date and coincides with a category-wide move, you have your answer: the retrieval behaviour changed, not your content. If the timing is unrelated and the drop is yours alone, keep looking on your side.
Step 5: Decide. Fix a regression, or ride out a shift
Now act on the diagnosis, not the panic.
If it is a your-side regression, work the reachability-to-quotability chain in order:
- Reachability: confirm AI crawlers still get a 200, not a block. A CDN or robots change is the most common silent killer. See how to see which AI bots crawl your site.
- Render: confirm your facts are in server-rendered HTML, not injected by JavaScript a crawler skips.
- Structure: confirm exact facts (specs, pricing, attributes) are in clean lists and tables a precise, targeted search can match, not buried in prose. This is where structured data earns its keep.
- Freshness: confirm the page isn't stale past the citation cliff; refresh with substantive updates, not a bumped date.
- Entity: confirm your naming and entity strength are intact and corroborated across the web.
If it is a model shift, do not thrash. You cannot control an engine's fan-out or source weighting. Ride it out, keep sampling, and lean into the durable levers above, which pay off across whichever retrieval behaviour ships next. As Luna's precision turn shows, exact structured facts and reachable first-party pages are the safe bet regardless of the update.
The discipline in one line: measure a trend, split it into mentions, retrievals, and fan-outs, check whether the whole category moved, and only spend effort where the cause is actually yours. Doing that continuously across ChatGPT, Gemini, Claude, and Perplexity, so you can tell a model shift from your own regression, is exactly what Buffy Intel is built to do.