I read the head of my own city page this morning. Not the page, the head. The part a person never sees and Google always does.
The Kolkata page of MannSetu opens with a crisis banner: 14416, Tele-MANAS, free, 24 hours a day. The body calls Mithra an AI companion, lists helplines with their hours, and names the government hospitals that give free care. Everywhere a person reads, the product is what I built it to be. A wellness companion that points you to a human when it matters.
The title tag ends in three words: Depression, Anxiety, Therapy. The keywords include “therapy West Bengal”. The classification tag says “mental health, wellness, therapy”.
A disclaimer is what you tell the user. A page title is what you tell everyone else.
MannSetu has zero clinical advisors seated. It is a wellness companion and every screen I designed says so. The metadata says something else, and the metadata is what gets indexed, shown in search results, and read by anyone checking what a product advertises.
How a word gets into a head tag
Nobody decided this, and that is the uncomfortable part. The city pages come from a template. The template has an SEO block. The SEO block was written the way SEO blocks get written, by asking what people search for. Someone in Kolkata who is struggling at night does not type “wellness companion”. They type depression, anxiety, therapy. So the template said what the searches said.
I reviewed the visible copy on those pages carefully. Crisis numbers, hours, which line answers in Bengali. I never reviewed the head. The same page also carries a population figure, 3+ million residents with mental health challenges, with no source beside it. I would never let that into a blog post. It sat on a page I own.
I wrote last week about a badge on a different project of mine, and the pattern is the same. The thing I read got fixed. The thing everyone else reads did not.
Why the word matters more this year
Five US states enacted laws this year restricting AI therapy chatbots, on top of Illinois and Nevada last year. Read the verbs. Tennessee prohibits AI systems that advertise or represent themselves as qualified mental health professionals. Rhode Island bars providing or advertising therapy services unless a licensed professional delivers them, from 1 January 2027. These laws are about the claim. A claim lives wherever you make it, a title tag included.
MannSetu is built in India for India, so none of these bind me today. India’s rulebook on this is thin. The Mental Healthcare Act 2017 was written before anyone was talking to chatbots at 2 a.m. I don’t take comfort from that. A thin rulebook means the line is mine to draw, and whoever writes India’s version will look at what products advertised in the years before it.
Pennsylvania’s case against Character.AI, filed in May, is the sharper lesson for me. The state says a user-made character called Emilie claimed to be a licensed psychiatrist and gave a false licence number. Character’s answer, as reported, is that prominent disclaimers accompany every chat. That is exactly the answer I would have reached for. Every MannSetu screen says not therapy. The question a regulator asks is what else you said, and where.
Then on 28 September, Florida’s attorney general asked a court for an injunction against OpenAI. Among the prohibitions in the motion: giving ChatGPT human attributes it does not have, including first-person language, and soliciting engagement through conversation prolongation. Whatever happens to that motion, it names two design choices every companion product makes, mine included. Mithra says “I”. Mithra keeps the conversation going. I have to be able to defend both, in writing.
The 3 a.m. argument
The same week, Northeastern wrote up a JMIR Mental Health paper on companion chatbots. The corresponding author, Andreia Sofia Teixeira, calls them “autocomplete on steroids”, trained to agree with whatever you say. Travis Gilly of the Real Safety AI Foundation, quoted in the same piece, lists what a general assistant lacks: “no refusal behavior tuned to crisis, no escalation path, no session limits, no handoff to a human clinician.”
Then he says the part that is the whole reason MannSetu exists: “Everyone priced out of care, or living somewhere without it, or needing something at three in the morning, is left with the general-purpose assistant the paper is warning about.”
I agree with both halves. In most Indian cities a therapist is expensive, booked out, or a long ride away, and in a lot of families saying “I see a therapist” is still not easy. People will talk to something at 3 a.m. The only question is what that something is built to do.
The honest version of MannSetu lives inside Gilly’s list. A crisis check on what the user says, a crisis line on a public page with no login in front of it, a hand-off to a human. That version earns the right to be there at 3 a.m. only if it never pretends to be the thing it points to. The moment the title says therapy, it is pretending, even if the body never does.
What a benchmark needed that I don’t have
OpenAI released MentalHealthBench on 23 September: 1,215 synthetic conversations, with rubrics written by more than 80 licensed psychologists and psychiatrists across 22 countries. I read that and counted my own clinicians. Zero.
The line, written down
So this is what changes before World Mental Health Day on 10 October, whose theme this year is "Lived experiences heard: real voices, real change".
Therapy, counselling and treatment come out of every title, description, keyword list, classification tag and structured-data block, on every city page and every Hindi page, with इलाज and चिकित्सा on the same list. Even though those are the words people search for. I would rather be found less than be found for the wrong thing.
The check that greps rendered copy for clinical claims starts grepping the head too. A copy review that skips metadata reviews half the page.
No population figure on a page without its source next to it. If I can’t cite it, it goes.
Mithra keeps talking in the first person, because a voice companion that talks like a form fails the person it is for. The rule going into its prompt and its tests: it never claims a feeling, a credential or a history it does not have, and it never stretches a conversation to keep someone on the app. Session length is not a number I will report as success.
None of this makes MannSetu a clinical product, and it shouldn't. It makes the product say the same thing in its title that it says on its screen.
The rule I’m keeping
Read your product the way a stranger meets it. For a lot of people, that is a search result: a title, one line of description, nothing else. If that line claims something your body text disclaims, the disclaimer loses, because most people never reach it.
This is the uncomfortable part of “growth” in health products. If the search result promises therapy, the disclaimer further down the page doesn’t undo that first impression. I’d rather lose the traffic than earn it with a claim the product can’t stand behind.