You are on a call with two engineers and a product manager. Someone says, “we ran it through Gemini.” The room nods. Then the questions start. Which surface? The app or the API? The fast one or the thinking one? The version you get in a document editor or the one wired into search? Four people, four mental models, one word. By minute three you are not debating results anymore, you are debating vocabulary. Somebody shares a screen just to point at what they mean.
That meeting happens constantly now, and the reporting this week caught up to it. TechCrunch, Yahoo Tech, and Bitcoin World all ran versions of the same story about Google’s Gemini having a branding problem, and about how the rest of the industry shares it. The framing is usually marketing: bad names, confusing tiers, too many rebrands. I read it differently. Naming confusion in AI is not a communications failure downstream of the engineering. It is the engineering, surfacing where users can see it.
Names Are Interfaces
In software, a name is a contract. When I import a library at a pinned version, the name plus the version tells me what behavior to expect, what the inputs look like, and what breaks if I upgrade. That contract is what makes systems composable. You can build on something you can name precisely.
Modern AI products have quietly broken that contract. A single brand name now spans a family of models with different parameter counts, different context windows, different tool access, different safety configurations, and different routing behavior depending on which product surface you entered through. The name refers to a portfolio, not a component. So when someone says they used Gemini, or any of its competitors, they have communicated roughly as much as saying they used a computer.
For agent builders this is not a cosmetic irritation. Agent architectures are chains of delegation. A planner calls a tool, a tool calls a model, a model calls another model. Every link in that chain needs a stable referent, because the whole system’s behavior is the composition of those pieces. If the middle of the chain is a brand that silently re-routes to a different set of weights next month, you have built on sand and written the postmortem in advance.
Why The Naming Keeps Slipping
The names are unstable because the underlying artifacts are unstable, and for structural reasons:
- Models are moving targets. Weights get updated, distilled, quantized, and swapped. The product keeps its name so that customers do not churn.
- Routing is now part of the product. Many systems decide dynamically which model handles a request. The user-facing name describes the front door, not the room they end up in.
- One brand has to cover many price points. Consumer free tiers, paid tiers, and enterprise APIs all get the same family name for marketing reasons, even though they are meaningfully different systems.
- Capabilities arrive as features, not versions. Tool use, memory, and longer context show up as toggles inside an existing name rather than as a new named artifact.
Every one of those decisions is defensible on its own. Together they make the name useless as a technical identifier while keeping it valuable as a marketing one. That gap is where the confusion lives.
The Same Pattern, Elsewhere In Google’s Week
Two other stories from the same news cycle rhyme with this. Google gave publishers a new way to fight traffic losses driven by AI, which is an admission that when a system absorbs the web into an answer, the original source loses its identity along with its clicks. Attribution is a naming problem too. And the Pixel 11 Pro XL review landed on a familiar verdict, that faster cameras cannot hide an iterative upgrade. Hardware reviewers have decades of practice separating a new name from a new thing. AI has almost none.
What I Would Actually Want
I do not need better marketing names. I need identifiers that behave like identifiers. A content-addressable reference to a specific set of weights. A version string that changes when behavior changes. A response payload that reports which model actually served the request, not which family it belonged to. Deprecation windows announced in advance, the way sane APIs handle them.
None of that is technically hard. It is organizationally awkward, because it makes the churn legible and gives customers something concrete to complain about. But agent systems cannot be evaluated, reproduced, or debugged without it. Every serious reliability practice we have, from regression testing to incident review, assumes you can say precisely what ran.
The branding stories are being read as a story about consumer confusion. It is also a story about engineering discipline arriving late to a field that grew faster than its vocabulary. Whoever fixes the naming will not be doing it for the marketers. They will be doing it for the people trying to build something dependable on top.
🕒 Published: