A curated jury of personas from your audience reads your copy before it goes live. It runs on Claude; we don't have a language model of our own. Below is what you get back, each with an invented example.
A buying committee reads copy differently from a family comparing two plans. Every audience is one of three kinds: business buyers, consumers or politics. The jury reads in the role that fits.
For political copy it judges whether a message lands. It does not judge the position itself, and it takes no side.
Not every persona counts the same. You decide who your offer is made for, the core, and who only reads along, the edge. The verdict follows the core; the edge is shown next to it.
If a version only wins with people at the edge, you get a warning instead of a win. Otherwise you would ship copy that convinces the wrong people.
Verdict follows the core: new version wins 4 of 4.
At the edge: 1 of 2.
Some copy puts nobody off and is nobody's first choice. The scores look fine, yet the copy pulls no one in. In that case the result carries a note, so you don't mistake a pale text for a good one.
Note: everyone likes it a little, nobody's first choice.
Add your brand promise in one sentence. Each persona says whether the copy keeps it or drifts away, with a short reason. This does not go into the score. It is a second question next to the first: copy can win and still leave the brand behind.
Brand promise: 5 of 6 say "keeps it".
Add your page and the copy of two or three competitors. For each of your core claims, 99takes shows whether a competitor promises the same (taken), something similar (contested) or none of it (open). Every rating carries a verbatim quote with its source.
It also shows where an open claim sits on your page: in the hero or further down. A claim three competitors make the same way decides no shortlist, however well it is worded. It compares what the texts say, not how the market perceives the vendors.
Open: "fixed price per environment". Sits in section 8 of 11, belongs further up.
Taken: "publish without developers" (competitors A, B).
99takes also writes: versions that compete with your current copy. From the objections of a run, from the customer voices and research behind your audience or, without any draft, as a complete page from your positioning and the story your copy should tell (old way, new way). On request with a specific method, such as AIDA, benefit-led or headlines only.
Figures, customer names or results that are not backed anywhere, 99takes does not write: where proof is missing, the copy carries a placeholder. The versions are then tested like any other copy, and the jury does not know where they come from. If a version was written from the objections of this very jury, the result says a win there is expected and only the market decides.
Version 2 · Proof first: "Fixed price per location, term [PROOF: minimum term]. No setup project."
With five drafts, not every one needs a full run. A smaller jury sorts out which one is worth it. This is a shortlist, not a measurement, and the result says so.
A CFO reads copy differently when choosing a vendor than when justifying a renewal. On request, every persona reads the copy in every situation the set knows. You see where the copy carries least.
Instead of testing a whole page at once, the build goes step by step: the hero first, then the offer and the other sections. Each step is tested so the steps stay comparable. At the end, the finished page has to beat the old one.
Every run has a detail page with each persona's scores for both versions. A study becomes a report with its own link that you send to colleagues or your client; in Slack or LinkedIn the link shows a preview with the result. You can revoke a link at any time.
Report which version won in the market. That builds a hit rate per jury, always with the number of results and a range. It is published only from ten decidable results on; until then it reads "not measured yet".
The rate comes only from reported market results and belongs to exactly the jury that judged.
You can stage a persona round with any language model. What is missing is what lets you rely on a result:
claude mcp add --transport http 99takes \
https://srcblvgafgkdeaaowhfu.supabase.co/functions/v1/mcp \
--header "x-api-key: YOUR_KEY"