"Something's off with Gemini lately." I've thought this more than once. A Gem that was producing sharp, useful output started feeling generic. Deep Research results that used to feel thorough started missing obvious angles. The quality felt like it had slipped.
After investigating, I found that "Gemini got worse" was the right explanation far less often than "something on my end changed." Here's how to figure out which situation you're actually in.
Start With One Test
Before doing anything else: take the same prompt and try it in a completely fresh conversation.
If quality returns in a new conversation, the issue is conversation context — Gemini's responses degrade when conversations accumulate ambiguous instructions, contradictions, or vague references. This isn't a Gemini flaw exactly; it's a limitation of how language models handle long conversation histories. The fix is simple: start a new conversation.
One more habit worth building: keep a few "benchmark prompts" on hand. Save a prompt that once produced great output, together with the response it produced at the time. Whenever something feels off, run the same prompt and compare against the saved response. You stop judging by feel and start judging by diff — and a surprising number of investigations end with "it was in my head."
If the problem persists in a new conversation, keep reading.
Cause 1: Your Prompts Have Drifted
When you first set up a workflow with Gemini, your prompts were probably specific and clear. Over time, they tend to erode: "do the usual thing," "continue from before," "same style as last time."
Gemini is very context-sensitive. Vague inputs reliably produce mediocre outputs. The model hasn't gotten worse — your instructions have gotten less precise.
Test: find your original prompt from when Gemini was working well and run it verbatim. In most cases, quality returns immediately. This tells you the drift is in the prompt, not the model. To prevent the erosion entirely, store your frequently used instructions in full form — in a notes app or a text snippet manager — and paste from there every time instead of retyping abbreviations.
Cause 2: The Default Model's Underlying Version Changed
Gemini updates its model lineup frequently, and the default model for a given tier can change without much fanfare.
In June 2026, for example, Gemini 3.5 Flash reached general availability and became the model behind the "latest Flash" reference. Around a transition like that, identical prompts can produce responses with different lengths, tone, and clarifying behavior. Speed and quality may actually be better — but "different from usual" alone is enough to register as a downgrade. I've caught myself doing exactly that more than once.
Check your current model in settings. If you're on a paid plan, explicitly select the model you want — Gemini 3.1 Pro, for instance — rather than relying on the default. Pinning your choice keeps the day-to-day feel stable.
If you use Gemini through the API, you can go one step further. Alias references like gemini-flash-latest change behavior the moment the underlying model is swapped. If stability matters more than novelty, specify a pinned version name and upgrade on your own schedule instead of Google's.
Cause 3: Your Gem's Custom Instructions Are No Longer Working as Expected
If you're using a custom Gem, model updates can change how the instructions are interpreted. Rules that were followed precisely before might be treated more loosely after an update — especially abstract style instructions ("write like a thoughtful advisor") or negative constraints ("never use bullet points").
The fix is to revisit your Gem's instructions and make them more concrete. Instead of "use a thoughtful advisor style," describe specific behaviors: "Respond in 2-4 paragraphs, no bullet points, and always acknowledge the tradeoffs before making a recommendation." Concrete behavioral instructions survive model updates far better than tone descriptions — that's the pattern I've seen across several rounds of rewriting my own Gems.
Cause 4: Browser or App Cache
This sounds trivial but accounts for a surprising number of reported issues. Garbled output, responses stopping mid-generation, old conversation content appearing in new conversations — these often trace back to stale cache.
Try Gemini in a private/incognito window before anything else. If quality improves, clear your browser cache. For the Android or iOS app, clear the app cache from settings.
When It Actually Is Gemini's Side
If none of the above resolves it, there may be a real service issue.
Check the Google Workspace Status page (workspace.google.com/status) for active incidents. Search "Gemini not working" on X/Twitter — if there's a widespread issue, you'll see reports within minutes of the problem starting.
Also: in the few days immediately following a major model update, response consistency tends to vary more than usual. This typically stabilizes within a few days without any action on your part.
My Routine: Re-running the Same Samples After Every Update
In my own work as an indie developer, one step of my app operations hands image-asset sorting to the Gemini API. What that pipeline taught me is that every model update shifts the tendencies of the output slightly. Overall accuracy tends to improve, but the judgment calls on borderline cases move around. It's not degradation — it's change — and if you ignore it, it quietly affects everything downstream.
So I built a small habit: after each update, I re-run a fixed set of samples whose correct answers I already know, and compare against the previous run. For any items that shifted, I tighten the prompt conditions until they hold. The same idea applies to chat usage — the benchmark prompts from earlier serve exactly this role. Having one mechanism that detects change turns "I think it got worse" into "here is what changed," which is something you can actually act on.
What I Do When Quality Seems Off
When I'm not satisfied with a Gemini response, the most reliable thing I do is ask it to explain its reasoning: "Why did you answer this way?" Gemini walking through its logic often reveals the assumption or misunderstanding that led to a mediocre response — and from there, a single targeted follow-up usually gets back on track.
I also try rephrasing — sometimes in English if I originally asked in Japanese. For technical topics especially, there can still be performance differences depending on language. It's not a workaround I love, but it works.
The frustration of "this used to work better" is usually a sign you've been using Gemini enough to develop real expectations. That's worth something. Start with the fresh-conversation test, and work through the diagnosis from there — it almost always leads somewhere useful.