The cutoff gap: why your brand exists in ChatGPT but not in Gemini
Major AI models have knowledge cutoffs spanning 12+ months. Your brand can be famous in one and unknown in another. Here's the fix.
The cutoff gap: why your brand exists in ChatGPT but not in Gemini
Pick a brand that launched in March 2025. ChatGPT will tell you about it. Claude will tell you about it. Gemini will look at you blankly and either hallucinate something or punt to web search. Same brand, same web presence, three different realities.
That’s not a bug. That’s the knowledge cutoff gap, and most marketers I talk to either don’t know it exists or know it exists but treat it as somebody else’s problem. It is, in fact, your problem. The gap between when each frontier model stopped training and when each frontier model gets used is now wide enough to swallow entire product launches, rebrands, and pivots.
What the cutoffs actually look like right now
Anthropic publishes the dates in their public model documentation. OpenAI and Google are looser about it, but third-party trackers and release notes confirm the same range. As of late May 2026, the leading models look like this:
| Model | Reliable knowledge cutoff | Training data cutoff |
|---|---|---|
| Claude Opus 4.7 | January 2026 | January 2026 |
| Claude Sonnet 4.6 | August 2025 | January 2026 |
| Claude Haiku 4.5 | February 2025 | July 2025 |
| GPT-5.2 | August 2025 | August 2025 |
| Gemini 3 Pro | January 2025 | January 2025 |
Sources: Anthropic model docs for Claude; Temso AI’s 2026 cutoff tracker for GPT-5.2 and Gemini 3 Pro.
Look at the spread. Gemini 3 Pro stopped learning sixteen months ago. Claude Opus 4.7 was learning four months ago. A brand that announced a product in May 2025 is three things at once: established history in Claude Opus 4.7, recent news in GPT-5.2, and a non-event in Gemini 3 Pro.
And note the gap between Claude’s two columns. Anthropic distinguishes “reliable” from “training data” cutoffs because training scrapes through a date doesn’t mean the model has dense, well-formed knowledge of that period. Coverage thins out as you approach the edge. Even within a single model, your brand’s recall depends on whether you were a frequent mention well before the cutoff or a sparse mention near it.
Why models forget you (and why retrieval doesn’t always save you)
The defense most teams reach for is “but they all have web search now.” True. Mostly. The fix is partial, and the partial part matters.
When a model has trained on you, the brand sits in its parametric memory. The model can produce information about you without retrieving anything. Recommendations, comparisons, descriptions, sentiment, all of it comes from the weights. When a model hasn’t trained on you, every answer about you depends on a real-time web search that the model may or may not run, may or may not run well, and may or may not run consistently across users.
Three things go wrong with retrieval as your fallback:
First, models don’t always retrieve. They retrieve when the prompt looks fresh (“latest,” “today,” “recent”) or specific (“who is X”), and they skip retrieval when the prompt looks general (“best CRM for startups,” “what’s a good project management tool”). General prompts are exactly the queries that drive discovery, and discovery is exactly the query category where parametric knowledge wins.
Second, when models do retrieve, the retrieval surfaces a different brand population than parametric memory does. The web has its own ranking biases. Reddit threads outrank your product page. Forbes outranks your blog. Wikipedia is essentially mandatory. The brand that wins parametric recall and the brand that wins retrieved recall are often not the same brand. That’s the bones of the citation divergence we’ve documented across platforms.
Third, retrieval is slower and costlier, so platforms route around it. ChatGPT’s free tier, mobile interfaces, and integrated assistants all bias toward parametric answers. The pure web-search experience that GEO playbooks assume isn’t what most users get most of the time.
Three brand scenarios
The cutoff gap maps cleanly to three scenarios. Marketers usually think about visibility as “do AI tools know me, yes or no.” The real question is “which models know me, and what did they learn about me when they learned it.”
You’ve been around since before 2024
You’re in everything. Your problem isn’t recall; it’s whether the AI knows your current positioning or some legacy version of it. Sentiment drift, outdated product descriptions, and rebrands that haven’t propagated are the work. We cover the pattern in how AI answers change over time.
You launched between January 2025 and August 2025
You’re in OpenAI and Anthropic’s recent training data and not in Gemini’s. If your buyers use Gemini, AI Mode, or AI Overviews, you depend entirely on Google’s retrieval layer to bridge you in. That’s a different set of optimization moves: ranking on Google’s organic results, getting cited by AI Mode’s preferred source set, and showing up cleanly in the entity graph.
You launched after August 2025
You’re not in any model’s parametric memory. Every mention of your brand is generated through retrieval. You have the most volatile visibility profile of the three, the highest variance from user to user, and the most upside from disciplined off-site presence. This is the group that should be hardest on Wikipedia, on Reddit, on YouTube, on the fifteen domains that disproportionately control AI visibility.
The painful part: most teams don’t know which scenario they’re in because they only test one or two models, and they only test their own brand by name. A real picture requires testing across all three model families, across category queries (where parametric memory dominates), and across direct queries (where retrieval can paper over the gaps).
How to break into a model that doesn’t know you yet
There’s no clean way to “force” a model to add your brand to its parametric knowledge before the next training cycle. The training run is what it is. What you can do is increase the odds that the next training run picks you up densely, and that the current retrieval layer surfaces you reliably until then.
Start by saturating the high-frequency training corpora. Wikipedia first, then high-authority editorial. Models weight these sources heavily because they appear repeatedly across the training set. We covered the Wikipedia connection in detail; the short version is that brands without a Wikipedia presence are giving up the cheapest parametric ranking signal that exists.
Then optimize for retrieval consistency, not retrieval frequency. A brand cited five times in five different model answers is more visible than a brand cited fifteen times in three. The cited-once brands are noise. The cited-everywhere brands are recommendations. So track citation stability across reruns, not just citation counts.
Run named queries against every major model family monthly. Don’t trust “ChatGPT knows me” to mean “Gemini knows me.” They are different models with different training pipelines and different cutoffs. If you only sample one, you’ll mistake a one-model presence for a market-wide one.
Finally, treat your launch date as data. When you ship a new product, log the date and watch for it to appear in each model family. The lag tells you which model is leaning on retrieval versus parametric memory, and that signal informs where to spend your visibility budget. We touched on this asymmetry when looking at the GPT-5 vs GPT-4 citation gap, where version-to-version differences within a single provider can be as large as cross-provider differences.
The teams that win the next eighteen months will treat AI visibility as a per-model problem with per-model fixes. The teams that lose will keep running monthly screenshots of ChatGPT and declaring victory.
The uncomfortable summary
A brand has as many identities as there are models in the market. Each model’s identity for your brand is frozen at a different point in time, partially patched by retrieval, and biased by whichever sources that retrieval prefers. Calling all of that “AI visibility” as if it were one number is the same kind of mistake as treating Google and Bing rankings as one number in 2008. They were different. These are different.
Find out which model has the wrong picture of you, figure out whether the fix is parametric or retrieval, and stop assuming that good visibility in the model you happen to test is good visibility in the model your customers happen to use.
RivalHound tracks your brand’s visibility across ChatGPT, Google AI, Perplexity, and more. Start monitoring to see where you stand.