Start with the four switches, then say who you are, then fix whichever of the 23 checks is failing worst, then ask for new words on the one page that sells. Each later request depends on the one before it: there is no point rewriting a title AI cannot yet reach.
Installing the agent is the easy part; knowing what to type into it the first week is where most owners stall. The order below follows the same logic as what to fix first so AI recommends you: access before entity before content before reach, because a blocked fetcher makes every later fix invisible.
Because each layer depends on the one under it. A rewritten title does nothing for an engine that cannot fetch the page at all, and a page an engine can fetch still gets treated as an unclear source if there is no Organization data saying who published it. Ask for the later things first and you are polishing a page AI has not been let near yet.
Type it in your own words — "let ChatGPT, Claude and Perplexity read my site" — and the agent flips the robots.txt access for AI retrieval fetchers itself, in one click. This is one of the four switches it is allowed to apply directly, from inside the site, and it is reversible: turn it off and the site is exactly as it was.
Ask it to "add my company's Organization data," and it writes and switches on the Organization block in the page head itself. This is the second switch. Without it, an engine that can now read the page still has to guess who is behind it, which is the entity-ambiguity problem behind many wrong or missing recommendations.
Turning on the sitemap and building /llms.txt from your own page titles — both one click, both reversible. Together with crawler access and Organization data, these are the four things the agent is built to apply itself, from inside the site, with nothing written into the theme or the content.
| You ask, in your own words | What it comes back as |
|---|---|
| "Let the engines in" | A robots.txt switch, one click |
| "Say who you are" | An Organization-data switch, one click |
| "Turn the sitemap on" | A switch, one click |
| "Fix what the failing check names" | Varies — a switch, words to approve, or a person's job |
| "Rewrite the title of the page that sells" | Exact words, shown before you approve |
| "Answer the buyer's question in the first sentence" | Exact words, shown before you approve |
| "Add the FAQ" | Exact words plus its schema, shown before you approve |
| "Redirect what is broken" | A named redirect, checked against its target, then approved |
Ask it plainly — "what's failing right now?" — and it reads its own 23 structural checks and names the one worth fixing first, rather than you working down a list. Some failing checks come back as a switch it can flip itself; others come back as words for you to approve; a few, like a rebuilt page template, it will say plainly are a developer's job. See inside What to Fix for how it gets from a losing answer to a concrete change.
The title and search description of the one page that actually sells, then an opening paragraph that answers the buyer's question in its first sentence, then an FAQ. Each comes back as exact text, shown to you the way it will appear, before anything is sent. One worked example, a title rewrite end to end, walks through exactly this request.
Name the old URL and ask for a redirect — the agent checks that the target actually answers before proposing it, then applies the 301 on your click. This is the same request behind fixing an old URL that still gets traffic: an address in an old campaign or someone else's link that now answers with an error loses both the buyer and the engine at the same point.
"Why did AI stop opening the blog?" and "get /pricing read by AI" are the two shapes worth typing once the switches are on. The first asks it to diagnose a drop; it comes back naming which check changed and what it saw. The second asks it to act on one page; it comes back with a plan where every line is marked — a switch it applies itself, words it writes for your approval, or a step only a person can do.
None of this needs a developer, and each step takes minutes:
Start with the four switches: let AI crawlers in, add your Organization data, turn on the sitemap, and build /llms.txt. Each is one click and fully reversible, and nothing else matters until AI can reach and recognize the site.
Yes. A rewritten title or a new FAQ cannot be read by an engine that is blocked at robots.txt, and an engine that can read the page still won't trust an unclear entity. Access, then identity, then content, then reach, in that order.
Ask it directly, in plain words: "what's failing right now?" The agent reads its own 23 structural checks and names the one worth fixing first, rather than you guessing from a list.
Yes. Beyond the title, description, opening, FAQ, canonical, robots meta and redirect it writes by name, you can ask for a stylesheet change, a snippet in the head, a replaced passage, a new page or a response header, and it writes the exact change for approval before anything happens.
Nothing is applied before you see it exactly as it will appear, and every change keeps its own undo, so asking for the wrong thing costs you one click to reverse, not a rebuild.
The free check reads your site the way ChatGPT, Claude and Perplexity read it, and names the check worth fixing first. 3 minutes, evidence included.
The product: Reflexa Agent · how it works · på svenska