You create a key with each AI provider you want to monitor, OpenAI, Google, Perplexity and so on, and paste it into AI Watch. When a prompt runs, the request goes to that provider on your account, and they bill you directly for it.

There is no reselling and no markup. We never sit between you and the provider on billing, so what you pay is what the run cost.

Because bundling it would mean either charging you a subscription that covers the heaviest user, or quietly capping what you can run. Both make the price a guess.

It also keeps you in control: the models, the spend limit, and the ability to revoke access are all on your provider account, not ours.

Each key is sealed with AES-256-GCM under its own data key, and that data key is sealed again by a master key. Recovering a key needs the database row and the master key, a database dump on its own is inert.

Where the master key lives is configurable. On a self-hosted install it is a root-owned file outside the web root; with AWS KMS the key material never leaves AWS, so taking the server is not enough. An attacker also has to be able to call Decrypt as that IAM principal, and every attempt is logged in CloudTrail.

After you save it, the key is never displayed again, not to you, not to an operator. The interface shows only the last four characters so you can tell which key is in place. The admin panel has no action that reveals one, by design.

Decryption is refused outside a queue worker or console command, so nothing in a web request can unseal a key even if it tried.

Three things. Before any batch runs, the screen tells you how many prompts and engines it covers, roughly what it will cost, and how much of your weekly allowance it uses, so “all brands, all engines” is never a surprise.

Second, you can set a monthly spend cap. Runs are estimated before dispatch and refused once the cap would be breached, and you are emailed as you approach it.

Third, the allowances below are hard limits rather than fair-use suggestions.

After repeated rejections the key is paused automatically, so failed runs stop consuming your allowance and stop hitting the provider. You get an email, and it is recorded in the activity log with the reason.

This matters more than it sounds: a key that dies quietly is the usual explanation for a week of flat data, and without telling you it looks like the product broke.

Any of the supported providers that offer web-grounded answers, currently OpenAI, Google Gemini, Perplexity, Anthropic and xAI Grok. One is enough to start.

Adding engines multiplies cost and reading, not your run allowance: the quota counts prompts, so three engines on one prompt is one run against quota and three answers to read.

By default an account holds 75 stored prompts and can make 225 runs a week, tracking two brands with up to three people on the account.

Prompt slots are shared across brands. Adding a second brand splits the 75 rather than doubling it. The allowance belongs to the account, not the brand.

They can be raised, but that is a conversation rather than a setting, which is what keeps the platform free.

Each prompt is locked for three days after it runs. Answers do not change meaningfully hour to hour, so re-running one is usually spending money to watch noise.

Free because the expensive part, the model calls, is already on your account, so there is little left to charge for. Invite-only because the queue and the support load are the real constraints, and growing slowly keeps both honest.

Full answer text is kept for about six months, and the measurements taken from it, mentions, citations, scores, for around fourteen. Daily rollups behind the trend charts are kept longer, because they are small.

Once an answer passes its window the text is removed, but the row recording that the run happened stays, so your history does not develop holes.

Its results go with it, permanently. There is no archive and no undo. Deleting a prompt means the answers, mentions and citations attached to it are gone.

That is deliberate: keeping a shadow copy of data someone asked to remove is worse. It is also why the export below exists, take a copy first.

Yes. One click gives you a ZIP with a CSV per table, brands, prompts, every run, mentions, citations, rollups and the audit trail, and optionally the full text of every answer still inside its retention window.

Keys are never included. An archive gets emailed around and left in downloads folders, which would outlive every protection around them.

No. Answers are parsed to count mentions and citations, and that is the extent of it. Whether the provider you connected trains on your prompts is governed by your account with them, which is worth reading.

No. Grounded answers are non-deterministic and the underlying web results move, so small differences between runs are normal. That is exactly why the product reports movement between runs rather than a single reading.

Usually one of three things: the answer genuinely changed, a competitor started or stopped appearing, or the provider changed the model behind the name you selected.

The last one is marked on the chart when it happens mid-window, because a move there is the model, not you.

No, and that is on purpose. You declare the names you want counted. Automatic entity resolution that gets it wrong quietly corrupts every number downstream, and a wrong share-of-voice figure is worse than none.

Different spellings can be folded into one rival, so “Acme” and “Acme Inc” count once.

Treat it as a signal, not a measurement. It is read from the sentence around each mention using a fixed vocabulary, so it catches the obvious cases and misses irony and hedging.

It deliberately is not a headline number anywhere in the product, and the sentence it judged is always shown so you can disagree with it.

A weekly digest on Monday covering what moved, and alerts when something needs acting on, a paused key, or a spend cap reached.

The digest can be switched off in your account or from any of the emails. Turning it off does not silence key and spend alerts, because finding out from a flat chart is worse.