Engineering3 min read

Building trust into AI-powered recommendations

Evidence behind every recommendation, a person approving every change, an independent check of the live site, and measurement that never claims more than it saw.

StatusIn development

A recommendation is only useful if you can tell when to believe it. Most of the engineering in Avara so far has gone into making that possible — not by asking for trust, but by keeping each step checkable and keeping different kinds of claim apart.

Every recommendation shows its evidence

Avara starts from what it can observe on your site: the pages it crawled, their titles and status codes, how they link to each other, and — when you connect it — what Google Search Console reports. A finding is recorded together with that evidence, and an opportunity points back to the findings that support it.

When Avara decides what to pursue, it records the decision and says plainly what it chose not to pursue and why. Corrections you make — "that is not what we sell", "this page is not important" — are kept and later analyses do not overwrite them.

A person approves every change

Avara prepares work; it does not do it. Each recommended improvement states what should change, what must not change, and how Avara will check it. It stays a proposal until someone on your team approves it, and approving is a deliberate step: the screen says exactly what is being approved before anything is recorded.

Autonomy settings let Avara keep watching your site and re-assess its plan when something material changes. They never let it approve, execute or buy anything on its own.

Your coding agent does the work

Approved work is handed to the coding agent you already use, through a scoped token and a small protocol: claim the work, read the full instructions, report what was done. The agent works in your repository, on your machine. Avara never sees your code.

An agent's report is treated as exactly what it is — a claim. It is shown as "reported", never as "done".

Avara checks the live site itself

When work is reported, Avara fetches the affected pages from production on its own and checks each criterion that can be observed: that the title exists, that the page still returns the right status, that it is still indexable, that the canonical did not move. Constraints become guards, so a fix that breaks something else does not pass.

Deployments take time to propagate, so a change that is not visible yet is checked again after a few minutes before Avara settles on a result: verified, partially verified, failed, inconclusive, or an error in Avara's own check. Anything only a person can judge is listed for a person and never counted as passing.

Measurement says what it saw

Only a verified change starts measurement. Avara compares search performance after the change with the weeks before it, at 7, 14 and 28 days, using complete days of Search Console data. It reports what moved. It does not claim the change caused it, and when there is no data it says so instead of showing zeros.

Where this stands

These pieces are built and tested, and we have run the whole loop — review, approval, a coding agent's report, a failed check, a fix, and a verified change — on a test website. Measurement needs real weeks of real data, which no beta customer has yet. Avara is in closed beta, so we describe this as in development.