A general-purpose AI can evaluate a string. It cannot see your product. Every localization error that depends on where the text lands is invisible to it, and those are the expensive ones.
Updated October 2026
Credit where it is due. Pasting text into a capable LLM will reliably catch:
For a small batch of strings and a translator who wants a second opinion, that is real value.
It cannot see rendering. Text overflow, truncation, clipped buttons, broken layouts, horizontal scroll, none of these exist in the text. They exist in the interface. An AI reading a string list has no access to the thing that is broken.
It has no context for ambiguity. Open is a verb or a state. Free is a price or a description. Order is a noun or a command. The model has to guess, and it guesses from the string, which is exactly the information that does not resolve the ambiguity. Your product resolves it, and your product is not in the chat window.
It cannot check consistency across screens. A term translated one way on the settings page and another way in onboarding is a defect. Detecting it requires knowing both screens exist and that they refer to the same concept.
It produces no severity model. You get prose. What you need for triage is typed, scored, prioritized findings you can route to an owner.
It does not track anything. No coverage record, no per-release comparison, no evidence of what was checked. A conversation is not a QA process, there is nothing to show an auditor or a client.
It does not scale. The workflow is manual copy-paste. That is fine for 50 strings and unusable for a product.
New to rendered-product QA? Read the guide: Localization QA for live websites and web apps.
| General AI chat | Loqalit | |
|---|---|---|
| Sees the rendered product | No | Yes, DOM-level, in browser |
| Detects layout/overflow errors | No | Yes |
| Resolves context-dependent ambiguity | Guesses from the string | Evaluates against actual usage |
| Cross-screen consistency | No | Yes |
| Scored, typed findings | No | MQM typology and severity |
| Coverage record per release | No | Yes |
| Runs continuously | No | Yes |
This is not AI bad, our AI good. Loqalit uses AI too. The difference is what the AI is given: a decontextualized string list, or the running product with the text in place.
For a small batch of strings, a general-purpose AI will reliably catch grammar errors, spelling mistakes, and unambiguous mistranslations. What it cannot do is see your product. Every error that depends on where the text lands, overflow, truncation, layout breaks, and context-dependent ambiguity, is invisible to a model reading a decontextualized string list.
It cannot detect rendering failures such as text overflow, truncation, clipping, or broken layouts, because those exist in the interface rather than in the text. It cannot resolve context-dependent ambiguity, such as whether Open labels an action or a state. It cannot check terminology consistency across screens, produce typed and scored findings, or keep a coverage record of what was checked.
Yes. The difference is not the model, it is what the model is given. A chat window receives a list of strings with no context. Loqalit gives its analysis the running product, with the text in place, rendered in the browser at real viewport sizes.