Copy Fission · Campaign matrix
Expand one product claim across format × angle × channel × language: 3 formats, 6 content angles, 9 social channels, 8 languages. Every line ships with a naturalness score and specific AI-tell tags, and any single line can be rewritten with a one-sentence note.
Angles are multi-select: the same claim gets one version per angle with everything else held constant — a naturally aligned A/B set.
Two starting points
No reference copy needed. The system writes each line from the product facts you've confirmed — a claim that isn't in the fact pool doesn't make it into the output.
Hand it a line that already performs. It keeps the meaning and the claim, then rewrites for the target language and platform register — localization, not word-for-word translation.
What gets written here is short copy: social posts, landing-page blocks, ad lines. Full-length SEO articles — keyword plan first, then drafted piece by piece — run on a separate track. Go to the content engine ›
Naturalness QA
A 0–100 naturalness score with a public anchor: copy humans write off the cuff usually lands between 70 and 90. If a sentence reads machine-written, the review names it — this is the system inspecting its own output, not a black box that only hands you a number.
Each line is reviewed from a senior copy editor's viewpoint. The anchor is public: human copy usually scores 70–90; obvious template voice drops below 55.
More than a number. Every line carries concrete review notes — where it reads like a template ad, where the ending oversells — so you can see how the score was earned.
Unhappy with one line? Add a note like “more casual” and only that line is redone. The rest of the batch stays put.
People recommending something don't repeat the brand name. This is a generation constraint, not an afterthought check.
Software never gets described as something you hold in your hand. Usage scenes stay true to how the product is actually used.
Real samples
Three lines we generated with LinkBloom, for LinkBloom. Naturalness scores and AI-tell tags are copied over verbatim — including the flaws the review called out. Output is in Chinese, its original language.
我前两天赶着提案，先把官网链接丢进去了。 出来的不是那种空泛总结，像真的先帮我看了下：哪里能讲，哪里该补。 我最喜欢的是，它不是直接给我一张“看起来很忙”的报告图，而是顺手把下一步也写出来了。省得我再对着空白页发呆。 LinkBloom 这次终于不像在猜。
今天我本来只想随手看一下我们产品页的问题，结果越看越像在做体检。 我把链接丢进 LinkBloom 之后，出来的不是那种空泛的“建议优化一下文案”，而是能直接往下走的结论。那一刻我挺意外的，因为我最怕的就是这种东西最后变成一份漂亮 PPT，看完还是不知道下一步干嘛。 最有用的是，它会先把产品信息读明白，再往后推创意。不是靠我瞎想，也不是把卖点随手改一遍就交差。至少我看完以后，脑子里第一次很清楚：这个页面该从哪里改起。
SaaS 增长团队给新功能做投放，常见的卡点是：文案写了好几轮，AI 也出了几版图，结果总有一句话说得不准。功能边界被夸大，原本没有的效果被顺手补上，最后还得回到产品页逐句核对。 这时候，单纯换提示词没什么用。创意得先有依据，不能靠碰运气。LinkBloom 的做法是先把 SaaS 产品读成事实图谱，再从这些事实出发推演推广创意。
The tags are the review model's own words about this output — it doesn't go easy on the house product either.
Details that help
Pick the copy language per target market (8 available); your in-app notification language is a separate setting.
Every line comes split into headline, body, and CTA — no manual re-splitting before it goes into an ads manager.
2 credits per line, total cost shown before the job starts, failed generations refunded automatically.
2 credits per line. Start free, no credit card — 100 one-time credits at sign-up.
