Reflexa works in two steps: a page check that reads your site from outside, and an agent that goes on the site and does the work. The check is free in a free account. The agent applies the fixes you approve, writes the words for a page and publishes them the moment you say yes, counts every AI visit page by page, and answers what you ask about the site — and says plainly when a thing is not its to do.
Enter a domain. Reflexa fetches your public pages the way Google and the AI engines fetch them and runs 23 checks in four groups — crawler access, page content, brand identity, external sources — with no model and no AI engine involved. Minutes later you have a verdict: what failed, what passed, what an outside read cannot judge. Free, in a free account, every finding with the exact thing it saw.
Try it now: run the free check, or read what each of the 23 checks looks at.
One file in your middleware, CDN worker or server on a coded site; a plugin on WordPress. The agent is bound to that one site by its token, and Reflexa confirms it by fetching a proof phrase from your real domain — a token copied elsewhere is refused. Installing changes nothing on the site: the agent is in place, and waits for you.
Where it goes: the guide for your platform — Vercel, Next.js, Cloudflare, Netlify, Node, WordPress, and what works on hosts that run no code of yours.
Four fixes the agent applies itself, one click each: AI crawlers allowed in robots.txt, Organization structured data in every page, a sitemap, an /llms.txt. It applies them from inside the site on its next check-in and reports back. Switch one off and the site is exactly as it was — nothing is written into your theme or your content.
Which checks these answer: robots.txt, JSON-LD, sitemap.
For the things that need writing — a page's title, its search description, an answer-first opening, a FAQ block — the agent writes the exact text and shows it to you. Approve, and it publishes it as a revision you can undo. The FAQ schema is generated from the same words, so the page and its structured data can never disagree. Ask it to fix a URL that answers 404, and it proposes the redirect to the page that replaces it, checks that page answers, and serves the 301 once you approve.
What it writes: answer-first openings, titles and headings, FAQ blocks with their schema.
Once on the site, the agent sees every request an AI engine makes — the pages an engine opened to answer someone right then, the pages it crawled to learn, the people who clicked through from an AI answer — per page, per engine, per day. It is the traffic analytics never records, because crawlers run no tags.
Who comes: the crawlers behind ChatGPT, Claude, Perplexity and Google, verified by user-agent and IP.
Ask in your own words — attach a screenshot if that is easier. The agent answers from what it actually knows: the checks, the visits, Google Search Console, and the record of who did what and when. It comes back with a plan where every item is marked: this one it switches on itself, this one it writes and you approve, this one only a person can do. It runs a weekly page check by itself and answers recurring questions on a schedule.
What it never does: the limits, enforced not promised — no other plugins, no theme files, no users or DNS, no noindex on its own, no checkout or legal pages, no emails to your customers.
What happens the moment the agent arrives: it runs a full page check, the dashboard fills in — pages, AI traffic, checks — with what is new marked, you get an email, and it is ready to fix. What the agent does, and never does →
Each engine sends its own bots — one to train on, one to build its search index, one to read a page a person just asked about. Block one by accident, through a CDN default or a 2023-era robots.txt line, and that engine goes dark on you. The check reads your robots.txt for each; the agent counts each one's visits. Every crawler, one page each →
| Crawler | Sent by | What it feeds | If blocked |
|---|---|---|---|
| GPTBot | OpenAI | training data for OpenAI models | your pages stay out of future training; ChatGPT still reads you live via ChatGPT-User |
| OAI-SearchBot | OpenAI | the index behind ChatGPT search | you cannot appear in ChatGPT's searched answers |
| ChatGPT-User | OpenAI | a page a person asked ChatGPT about, right now | ChatGPT cannot open your page when someone asks about you |
| ClaudeBot | Anthropic | training data for Claude | your pages stay out of Anthropic's training crawl |
| Claude-User | Anthropic | a page a person asked Claude about | Claude describes you from other people's pages |
| PerplexityBot | Perplexity | Perplexity's index and citations | you disappear from Perplexity's answers and sources |
| Googlebot | Google Search, AI Overviews and AI Mode | you are out of Google and its AI answers alike | |
| Google-Extended | a robots token: Gemini training and grounding | content stays out of Gemini training; Search is unaffected |
Minutes. Reflexa fetches your public pages from outside and runs the 23 checks — no model, no AI engine, nothing to pay. The result lands in the dashboard and by email with the exact finding behind every line.
No. Installing the agent changes nothing. Each of the four switches is your click, every word it writes is shown to you exactly as it will appear before it is published, and anything it cannot reach it says so. Everything is reversible.
Mostly no. Most AI fetchers read the raw HTML the server sends and do not execute scripts, so content that only appears after rendering is invisible to them. The text-without-JavaScript check measures exactly how much of a page that is.
The crawlers of OpenAI (GPTBot, OAI-SearchBot, ChatGPT-User), Anthropic (ClaudeBot, Claude-User), Perplexity (PerplexityBot, Perplexity-User) and Google (Googlebot), verified by user-agent and, where the vendor publishes them, IP ranges — so an impostor borrowing a crawler's name is not counted as one.
The page check is free. Starter is $49 a month and Growth $149 a month; they buy the agent's work on the site, measured by usage, with the dashboard showing how much of the month is used and when it comes back. Free keeps one site; paid plans do not count sites.
It says so and names who can. A plan item is marked as a switch it flips, text it writes for your approval, or human-powered — press coverage, a Wikipedia article, a change in your CDN. It never pretends a thing is done that is not.
No score. A check an outside read cannot judge is marked as such, never guessed. Every change the agent makes is your click, shown to you first, and reversible. What it cannot do, it says — and names who can.
Reflexa Agent, the product built on this: reflexaagent.com · how it works · reflexaagent.se