Analytics tell you what happened on your site. They rarely tell you why. Henkan Vision is our answer to that gap: it records how visitors use your website and turns each visit into something you can read, then hands the same evidence to your AI agents. This post describes what Vision does today, in beta, without the roadmap.
What Vision records
Setup is one script in the head of your site, or a Google Tag Manager tag. The script records page structure, clicks, scrolls and navigation with rrweb. Form values are masked in the browser, and password or payment fields are never captured. Recording follows the consent setting the site owner chooses.
Not every visit is worth keeping. A visit becomes a session only when it lasts at least 30 seconds and the visitor actually does something. Idle tabs, bots and browsers you exclude (for example your team's) are left out, and none of them count toward your quota.
Each session then gets a replay that takes the shape of the visitor's device, a step-by-step timeline (pages, clicks, typing without the text, long pauses, errors, purchases) and, for sessions with friction, a short AI summary: what the visitor was trying to do, how the visit ended and how sure the model is.
From sessions to patterns
A few minutes after a session ends, Vision tags its friction. The tags are rules, not guesses, so the same behaviour always gets the same tag:
- Rage click: four or more clicks within one second on the same spot.
- Dead click: a click on a button or link that changes nothing.
- U-turn, form abandonment, long hesitation on price or delivery pages, checkout exit and JS errors.
One session with a dead click is an anecdote. When the same friction on the same page shows up in at least 25 sessions over 30 days, Vision groups them into a pattern, ranked by how many sessions it affects, with an AI-written likely cause and a test idea. The likely cause is labelled as a hypothesis: the replays are there to check it.
Two recent additions build on the same recordings. Funnels count how many sessions reach each step of a journey, in order, and let you open the sessions that left at any step to watch why. Visitors follow one anonymous browser across its visits, so you can see how many visits it took to convert.
Built to be read by agents
Vision does not stop at the video: it stores each session as data, so an AI assistant can read it. It includes an MCP server for Claude, Claude Code, Cursor or any MCP client, and a read-only REST API for your own tools. An agent can search sessions, read timelines and summaries, list patterns, and read funnels and visitor journeys.
Agents get the same read access as a member of your organisation, and nothing else: they cannot change anything. That makes it possible to ask "Why do mobile visitors leave the checkout?" and get an answer built on real sessions instead of a guess.
Vision also keeps visits by AI assistants that browse for a person apart from human ones. Those agent sessions get their own friction tags (loops, walls, failed forms) and never mix with human figures or count toward your quota.
Where Vision stands
Vision is in beta and available to the teams of Henkan's clients. Access is opened by Henkan, and every organisation works with a Henkan consultant who checks what each pattern means and turns findings into tests. The agent records and groups; the expert decides.
If you already have access, the documentation walks through installation, replays, patterns and the MCP connection. If not, talk to Henkan and we will set you up.
