On this page
You check your dashboard on a Tuesday and a number that had been flat for six weeks has moved eight points overnight. You didn’t touch the site. Nobody published anything. Somewhere in the last few days, one of the engines you track pushed a new model, and now you’re staring at a chart trying to work out whether you did something wrong. You didn’t. But that doesn’t mean the swing is meaningless, either.
This happens often enough that it deserves a real answer instead of a shrug. Model updates genuinely change what gets cited, and they also produce a lot of noise that looks identical to a real change on a dashboard. Telling the two apart is the actual skill, and most of the advice about AI visibility skips straight past it.
Why this keeps happening
Classic search ranking updates are announced, dated, and rare enough that an SEO team could build a whole calendar around them. Model updates are the opposite. They ship on no fixed schedule, across separate products from separate companies, usually without a changelog entry that mentions retrieval or citations at all. OpenAI’s GPT-5.4 release in March 2026 is a good example of the pattern: OpenAI’s own announcement talked about a 33% cut in factual errors and new computer-use ability, not about which sources the model would start preferring in a search-grounded answer. The retrieval and citation behavior around a model shifts as a side effect of a release that was never framed as a search update in the first place.
Google did the same thing from the other direction. On January 27, 2026, it swapped a new model, Gemini 3, into AI Overviews and AI Mode with no public announcement. The visible effect was immediate: the number of cited sources per answer jumped by roughly a third, and entity-rich pages started gaining share at the expense of thinner ones, while ChatGPT’s citations didn’t move at all, because it runs on a completely separate pipeline. One engine can reset overnight while the others sit still, which is exactly why a single-engine chart looks so alarming and is so easy to misread.
The three things that actually move
Not every part of “the model” affects visibility. Three specific mechanisms do.
Which sources it’s willing to cite at all. Each engine keeps its own shortlist of trusted domains, and the shortlists barely overlap. An analysis of 118,000 AI-generated answers across ChatGPT, Perplexity, Google AI Mode and Claude found that only 11% of cited domains appeared across more than one platform. A model update can reshuffle that shortlist for one engine without touching the others, which is why our own engine-by-engine tracking guide treats each surface as its own measurement, not a single “AI visibility” number.
How it decides which sub-queries to run before it answers. ChatGPT Search’s August 2026 shift is the clearest recent case. It began scoping far more of its background searches to individual domains rather than the open web, and Reddit’s share of ChatGPT citations fell from an average of 3.83% to 0.52%, an 86% relative drop, in the space of a week. Nothing about Reddit changed. The routing logic that decided when to consult it did.
What counts as a signal of authority. Freshness, entity clarity, structured data, whether a page reads as written by a real business rather than scraped content, all of it can be reweighted in a single update, which is what happened with the Gemini 3 swap above. This is the mechanism our guide on the sources AI engines trust covers in more depth, and it’s the one most worth checking your own pages against after a swing, because it’s the one you can actually act on.
What looks like a swing and isn’t
Before you diagnose a real change, rule out a fake one. The most instructive case in recent memory is Google quietly removing the &num=100 search parameter in September 2025. Rank-tracking tools had relied on it to pull 100 results per request, and once it disappeared, dashboards across the industry showed impressions collapsing by roughly two-thirds overnight. It read exactly like a mass de-ranking event. It wasn’t. As Botify put it afterward, the drop reflected the removal of synthetic, bot-driven activity from the data, not any real change in how people found a brand; actual clicks and click-through rates barely moved. The lesson carries straight over to AI visibility tracking: a vendor changing how it samples, a shortened prompt list, a new engine added to the mix mid-month, all of these move a chart without a single real answer having changed.
The same August 2026 Reddit story that showed a real routing change also shows how easy it is to over-attribute one. Search Engine Journal pushed back on the tidy explanation, pointing out that the timing didn’t quite line up (the routing change landed on August 8, the citation collapse on August 14) and that the tracking vendor itself couldn’t rule out a data-collection issue on its own end. The same outlet noted an almost identical Reddit citation collapse in September 2025, which turned out to trace back to the &num=100 removal rather than any decision about Reddit at all. Two collapses, two different root causes, both initially blamed on the wrong thing.
How long to actually wait
Treat a single day’s reading as an anomaly until it isn’t. Tools built specifically to watch this, like Surfer’s AI volatility tracker, score day-to-day citation turnover on a scale from “flat” to “storm,” and are explicit that only a sustained storm-level reading held across several engines at once points to a genuine model update rather than routine churn. One noisy day is not that.
Three checks before you act on a swing:
- Check your own domain’s trend, not a vendor-wide average. An industry-wide dip in a metric doesn’t mean your brand moved; average behavior and your behavior are different questions, and conflating them is how the Reddit story above got over-read in both directions.
- Compare across engines before you compare across time. If the number moved on one engine and held steady on the rest, you’re looking at that engine’s pipeline, not a change in your own signals. If it moved everywhere at once, the odds of a real, shared cause go up.
- Wait for two or three sampling cycles before concluding anything. A cadence tight enough to catch this without chasing every blip is worth designing deliberately; our guide to tracking AI visibility covers how to set one up so a single bad day doesn’t get mistaken for a trend.
If it turns out to be real
Once a swing survives those three checks, on your own numbers, isolated to a specific engine or shared across all of them, held for more than a day, it’s worth treating as real and working through properly rather than reacting to the headline number. Pull the specific prompts and engines that moved, look at what actually got cited in the answers you lost, and run it through a proper AI visibility audit rather than guessing from the topline score. Most of the time the fix isn’t a content overhaul. It’s a handful of missing sources or a signal the new model started weighting that your pages don’t carry yet.
Model updates are not going to slow down, and none of the engines are going to start announcing them. The realistic goal isn’t to predict the next one. It’s to have a measurement setup steady enough that when one lands, you can tell within a few days whether it changed anything about you, or just changed what the chart looks like. If you want a read on where you stand before the next update lands, our free visibility check runs the same prompt set across engines that this kind of comparison depends on.