Yes, on the one site it is installed on, and only what you have approved. Software that writes a recommendation is ordinary by now. Software that carries the recommendation out, on your site, is the part worth checking before you believe a demo.
This is the question behind every demo of what the Reflexa agent does on your site: not whether it can describe a fix, but whether it can apply one. The honest answer is yes, with limits that are enforced in code rather than promised in a policy page, and this article is the full inventory of both sides.
Three different things, each with its own level of trust: a switch the agent flips itself, words it writes that you approve before they publish, and, since October 2026, almost anything else you ask for on the page. Each level carries a different amount of risk, so the agent treats them differently rather than lumping every change into one "automation" toggle.
Why it matters: a vendor who says "it fixes your site automatically" without naming which of these three levels they mean is asking you to trust a black box. Reflexa's split is the opposite of that: every item in a plan is marked as one of the three before anything happens, from the agent's own description of itself.
Crawler access for AI retrieval fetchers in robots.txt,
Organization data in the page head, the sitemap, and an
/llms.txt file built from your own page titles.
These four run from inside the site, each needs one click from you, and
each is fully reversible: switch it off and the site is exactly as it
was, nothing left behind in your theme or your content.
| Switch | What it changes |
|---|---|
| AI crawler access | Allows AI retrieval fetchers in robots.txt, the file that otherwise blocks them silently |
| Organization data | Adds the entity block to the page head so engines can tell who you are |
| Sitemap | Turns the sitemap on so pages can be found at all |
/llms.txt | Builds the file from your existing page titles, no new writing required |
A page's title, its search description, an answer-first opening paragraph, and an FAQ block with the matching schema. The agent writes the exact text, you read it in the dashboard exactly as it will appear on the page, and nothing is sent until you press apply. On WordPress the change lands as a revision with the previous value kept, so it can be restored exactly if you change your mind.
Two more approval-only changes sit next to that core set. A broken or redirected URL: the agent names the exact redirect, from which address to which page, checks that the destination actually answers, and the redirect only goes live once you approve it. And noindex: the agent will tell you a page is accidentally hidden from search, but it never flips that setting on its own, by design, because removing a page from an index is the kind of change that deserves a human's decision, not a default.
A way to ask for anything else on the page, not only the fields the agent already knows by name. A stylesheet, a snippet in the head of every page, a passage replaced with better words, a whole new page, a response header, or a line of markup the structural checks flagged, such as a wrong canonical tag. You ask in plain words, the agent writes the exact change, shows it to you before anything happens, and applies it on your click, each one kept with its own undo.
On a connected repository, through a GitHub App install or a GitLab project token, the same request comes back as a commit you read before it lands on your default branch. Undo is another commit. What the agent has no hand for, it says so plainly instead of pretending it can reach further than it can.
Do this: for the full picture of where the agent's hands are a plugin, an edge file, a repository or a platform connector, see Inside Reflexa: what the agent does on your site.
Every item in a plan is shown exactly as it will appear, not as a summary, before you click anything. A switch shows what it turns on. A piece of writing shows the full text in place on the page. A code change shows the commit's diff. Nothing is applied on the basis of "trust us", and the agent re-reads the live page afterward to confirm the change actually landed, rather than taking its own report of "applied" on trust.
Eight things, enforced in code, not written in a policy the agent could ignore. It never touches theme, plugin or PHP files. It never creates or changes users, passwords, DNS, hosting or caching. It never flips noindex on its own. It never edits checkout, cart, account or legal pages, or any page you mark locked. It never deletes content or changes prices, and it publishes a new page only after you have read and approved that exact page. It never emails customers or posts to social accounts. It never acts on a site other than the one its own token belongs to. And it never spends money without a priced request you started.
Bottom line: the limits are the product's own refusal logic, so asking the agent to do one of these is a dead end by design, whoever asks.
None of this requires buying anything, and each step takes a few minutes:
Yes, on the one site it is installed on and paired to by that site's own token, and only what you have approved. Four switches it applies itself, page text it writes and you approve, and since October 2026 a general change hand for anything else you ask for, each shown in full before it happens.
Crawler access for AI retrieval fetchers in robots.txt, Organization data in the page head, the sitemap, and an /llms.txt built from your own page titles. Each needs one click, and each is reversible: switch it off and the site is exactly as it was.
A way to ask for anything on the page, not only the fields the agent already knows by name: a stylesheet, a snippet in the head, a passage replaced with better words, a new page, a response header, or, where a repository is connected, the same request as a commit you read before it merges.
Yes, all of it. A switch turns off and the site is exactly as before. A page edit lands as a WordPress revision with the previous value kept, and a commit undoes with another commit. Nothing is ever written into your theme or your content directly.
It never touches theme or plugin files, never creates or changes users, passwords or DNS, never flips noindex on its own, never edits checkout, cart, account or locked pages, never deletes content or changes prices, and never emails customers or posts to social accounts. These are enforced in code, not written in a policy.
The free check reads your site the way ChatGPT, Claude and Perplexity read it, and returns a plan marked switch, words for approval, or a person's job. 3 minutes, evidence included.
The product: Reflexa Agent · how it works · på svenska