AI Product Discovery Beyond Search:
What Mailbox MCP's Early Growth Shows

On 17 September, Mailbox MCP had 11 accounts. On 26 October it had 4,161. Somewhere in those five and a half weeks, a lot of people found a small, specialist product, and I'd like to tell you exactly how. I can't, and the reasons I can't are the most useful thing I've learned about AI product discovery.

AI product discovery is how people find a piece of software when they choose their tools through AI systems rather than only through a search box. It isn't one channel. It's at least four: the ordinary search results, a mention or recommendation in an AI-generated answer, a listing in an app or connector directory, and the plain fact of the product being technically there for an AI client to use. They have different gatekeepers. They leave different traces. And a record that says where somebody signed up says very little about where they first heard of you.

Mailbox MCP is a good case to test that on, because it was built to be found by AI users, it keeps more records than most young products, and I can read all of them. It is also my own product, so read the next box before anything else.

AI product discovery in one case: the short answer

Here's what the records show, before the detail.

• The step change came on 18 September. 11 accounts were created in the 22 days before it, then 32 on that day, and about 128 a day through October. I date Anthropic's acceptance of the Claude directory listing to late on the 18th; the sign-up records agree on the day and show the rise had already begun during it.
• Three in four sign-ups began inside an AI app. Of 2,805 accounts created from 5 to 26 October, 77% started at the consent screen inside Claude (51%) or ChatGPT (26%). Our website was the first door for 6%.
• Not one came straight from a search engine in that window. Over the whole period Google sent the product site 225 clicks, against 3,859 accounts created.
• The smallest door converted best. 28% of website sign-ups have paid, against 16% of ChatGPT sign-ups and 4% of Claude sign-ups.
• None of that says how anyone first heard of the product. A door is the last step. Nothing Mailbox MCP records today captures the first one, and AI answers in particular leave no trace at all.

So the honest summary is narrow. Directories inside AI apps were the main way in. Whether they were the main way people found out is a separate question, and the data can't answer it.

Four AI product discovery mechanisms, and why they are separate

Most writing about AI visibility collapses everything into "will ChatGPT recommend me?". For a software product that's the wrong unit. These are four different mechanisms, decided by four different parties, and each one can succeed while the others fail.

1. Conventional search results

A search engine crawls a page, may index it, and may then rank it for a query. Each step is conditional. Google's own guide to how Search works puts it bluntly: "Google doesn't guarantee that it will crawl, index, or serve your page", even when the page follows its rules. Crawler access is the start of a chain, not proof of anything at the end of it. What search does give you is a trace: a click arrives with a referrer, and Search Console counts it.

2. A mention or recommendation in an AI answer

Here a model names a product while answering someone. For Google's own AI features the eligibility test is the ordinary one: a page "must be indexed and eligible to be shown in Google Search with a snippet". For ChatGPT, Claude and the rest, there's no published route to being named at all. And when it does happen, it often leaves nothing behind. The person reads the name, closes the chat, and later types it into a browser or a directory search box. Their arrival looks like that last step.

3. A listing in an app or connector directory

Claude has a connectors directory. ChatGPT and Codex share a plugin directory, which most people still call the ChatGPT app directory. For MCP server discovery more widely, the official MCP Registry catalogues remote servers for other directories to build on, and calls its own metadata "deliberately unopinionated". A listing is reviewed, placed and searched by the directory's owner. Its trace is a sign-up that starts inside the AI app.

4. Technical availability for an AI client

This is the layer you control outright: a remote server an AI client can reach, a sign-in that works, a description of each tool the model can read. Anthropic lets anyone on a Free, Pro, Max, Team or Enterprise plan add a remote server by URL as a custom connector, with no listing at all. Availability is necessary for everything above it and sufficient for none of it. A server nobody has heard of is available to nobody in practice.

The relationships matter more than the list. A directory listing usually needs technical availability first, because the reviewer has to be able to connect. Search can send people to a directory rather than to you. An AI answer can send people anywhere, invisibly. And this site's own definition of AI Visibility Checking covers only the inputs you control, the first and fourth rows, precisely because the middle two are decided elsewhere.

Mailbox MCP's launch timeline: website, first sale, directory listing

, the product this case study is about

Mailbox MCP is a remote email MCP server: it connects Claude, ChatGPT or another AI client to a mailbox the person already has, Gmail, Microsoft 365 or any IMAP host, so the AI can read, search, draft and file their real email. It sells per mailbox, with a free tier. If you want the long version of what it does, your AI subscription can already run your inbox covers it. For this article, what matters is when each discovery surface appeared, and that the launch was not one event.

These are the dated milestones. Every date comes from a record I can show you, except one, and that one is labelled.

Three things are easy to blur here, so I'll keep them apart. The website went live in late August. Money changed hands from 1 September. And the period I think of as the launch, the one in the chart below, began on 18 September. They are three different milestones, and only the third coincides with growth.

What the chart can't do is isolate the directory. The three free launch listings went in the same week. The official registry entry had been there since 1 September with almost nothing to show for it. And the jump on the 18th started before what I remember as a late acceptance, which tells me either that my memory of the hour is wrong or that something else moved too. I'd rather publish that wrinkle than smooth it over.

Here's what launch day looked like from the inside, because it's a useful warning about the next stage. By 16:40 on 18 September we'd had 17 sign-ups. Twelve of them stopped at the step where you connect a mailbox. One Gmail user tried three times in 43 seconds, was told twice that Google wanted an app password, and left. A person who has found you is not yet a person the product has worked for.

Mailbox MCP adoption on 26 October 2026: what each figure measures

The dashboard on 26 October said this. I've put each figure next to what it actually counts, because they are not stages of one funnel and they don't share a unit.

This is evidence of adoption and of use. More than half the connected mailboxes were used in the last week, and those 131,084 calls came from real AI clients doing real work in real inboxes. It is not evidence that any particular channel produced the growth. I'm going to resist the temptation, which I feel quite strongly, to read the numbers as a scoreboard for the directory.

It is also why I won't draw a rate between the rows. Dividing connected mailboxes by accounts gives a number, but it's a number about two different things, and it would sit in a funnel chart looking like a conversion rate nobody ever measured.

Where Mailbox MCP sign-ups started: Claude, ChatGPT, website or search

Since 5 October, each new account records the door it came through, once, at creation. It stores the referrer's host and never the full address, the page our own links came from, any campaign tag, and whether the sign-up began at an AI client's consent screen. There's no cookie, no IP address and no analytics script. That narrowness was deliberate: the product's privacy position is that it doesn't track visitors, and this was the smallest record that could still answer the question.

An earlier, rougher measurement agrees. Before the field existed, the team matched each sign-up from 24 September to 4 October to its first request in the server's access log: 990 of 1,051 began inside Claude or ChatGPT, 34 on the website, and one came straight from a search engine.

Three readings of the table, in order of how far I'd trust them.

The AI apps are the main door. That's solid. 77% of recorded sign-ups began at a consent screen inside Claude or ChatGPT, which is where a directory listing, or a custom connector somebody added by hand, sends people.

"Direct" is partly the AI apps too. The classifier can only name an AI client once one has been approved. Someone who began inside Claude, arrived with no referrer and gave up before approving lands in Direct. That would explain why only 19% of Direct sign-ups ever added a mailbox. It's an inference from the code, not a measurement.

The website sends fewer people who pay far more often. 28% of website sign-ups have paid, nearly seven times the rate for Claude. I don't think the website causes that. The pages that sent the most sign-ups were the home page (89) and the personal pricing page (21), and somebody arriving from a pricing page has usually decided already.

And search? Zero recorded sign-ups came straight from a search engine. Search Console shows Google sending mailbox-mcp.com 225 clicks between 25 August and 23 October, out of 12,632 impressions. Even if every one of those clicks had become an account, Google's organic results could account for at most 6% of the 3,859 accounts created in that window. That's a bound, and a useful one. It isn't a verdict on search, because a search that ends on the website is counted as a website sign-up, and a search that ends on someone else's page, a directory or a forum, isn't in Search Console at all.

Why a Claude or ChatGPT directory listing is not an AI recommendation

If the directories were the main door, it's tempting to say "get listed and the AI will recommend you". The vendors' own documentation says otherwise, and our listing shows the gap.

Look at the label. It's a Community connector. Anthropic's review criteria say a submitted server is scanned automatically and, by default, listed as Community, and its verification page is careful to say that even the higher Verified label "isn't a security audit or a guarantee of how the connector will perform". As for where a listing appears, Anthropic decides that during review, and there's no application for it.

Then there's what Claude does in conversation. Anthropic's help centre describes when Claude brings an app into a chat, and the passage that matters for anyone selling software is this one:

"Claude doesn't take payment to recommend any connected app. There are no sponsored rankings, no paid placements, and no advertising."

Anthropic, in How Claude suggests connected apps, Claude Help Centre, 23 April 2026 (verify quote at source)

I was relieved when I first read that, and then a bit deflated, because of one word in the article's title: connected. The whole page is about apps a person has already connected, and when two of them fit, Claude shows both and lets the person choose. That's an even playing field, and I'm glad it can't be bought. But it's a playing field you only stand on after someone has installed you. Nothing on the page describes Claude pointing a stranger towards a connector they've never added. I'd assumed, without ever checking, that a directory listing put us in front of people mid-conversation. It puts us in a catalogue they have to open.

OpenAI's documentation says the same thing more plainly. Once a plugin is published, people find it in the directory by a direct link or by searching for its name. Placement and proactive suggestions are what its plugin guidelines call "enhanced distribution", which plugins with strong real-world utility may become eligible for. You can't ask for it.

We found out how much a directory's own search matters by accident. Before the listing's text was rewritten on 5 October, Mailbox MCP didn't appear at all when we searched Claude's directory for "hotmail" or "yahoo", and it sat 162nd of 636 results for "email". The rewrite led with the mail providers people actually type. The next day it was second of three for "hotmail", third of ten for "yahoo", second of 13 for "icloud" and third of 11 for "imap". Same product, same review, same label. Different words in the listing.

That supports one narrow claim: a listing's own text decides whether the directory's search can find it. It doesn't show the rewrite changed sign-ups, because nothing else was held still that week.

There's one more thing a listing isn't: a web page an AI answer engine can learn from. Fetched without signing in, neither our Claude listing nor our ChatGPT listing returns the product's name anywhere in its HTML, and claude.ai's robots.txt tells GPTBot, OAI-SearchBot, ChatGPT-User and Google-Extended to stay out of the whole site. Whatever an AI says about a product, it isn't reading it from that product's directory page.

AI answers: the discovery channel this case study cannot measure

I haven't measured whether ChatGPT, Claude, Gemini or Perplexity mention Mailbox MCP when someone asks for an email connector, and I'm not going to pretend a handful of prompts would tell us. The best public evidence on this says why.

"These tools are probability engines: they're designed to generate unique answers every time."

Rand Fishkin, co-founder and CEO of SparkToro, in NEW Research: AIs are highly inconsistent when recommending brands or products, SparkToro, 27 January 2026 (verify quote at source)

Fishkin's study ran the same prompts through ChatGPT, Claude and Google's AI 60 to 100 times each and found that the chance of getting the same list of brands twice was under one in a hundred. What I like about that sentence is that it isn't a complaint. It's a description of the machine. And once you accept it, a screenshot of an AI naming your product stops being evidence of anything except that it happened once. We wrote a whole article about why paid AI visibility trackers measure the wrong thing for exactly this reason, and I'd be a hypocrite to run three prompts now and print the results as a finding.

There's a worse problem than inconsistency, which is that an AI answer usually leaves no trace in your own records. We've seen the same shape on this site's search data: conversational queries in Google's AI Mode producing thousands of impressions and no clicks at all, written up in how to appear in AI search results. If someone met Mailbox MCP in an AI answer and then added it from Claude's directory, our records file them under "Inside Claude". The AI answer is invisible, and the directory gets the credit.

That's not a hypothetical worry about small numbers, either. G2's survey of 1,076 business software buyers, run in March 2026, found that 51% now start their research with an AI chatbot more often than with Google. If that's anywhere near true of our customers, a large part of the story happens where no referrer reaches.

The evidence table: what we observed, what it supports, what it cannot establish

This is the core of the case study. Each row is something we actually saw, dated, with the most it can bear and the thing it can't.

The right-hand column is longer than the middle one, and I think that's the honest shape of AI product discovery evidence right now. Anyone who shows you a tidy attribution of AI-era growth is showing you their assumptions.

What Mailbox MCP publishes, and which evaluation question each page answers

If discovery is mostly invisible, what can a product actually control? Its answers. When someone meets Mailbox MCP, through any of the four mechanisms, they arrive with a question, and the product's public footprint is a set of pages that each answer one. On 26 October the product site carried 74 pages in its sitemaps and its AI Discovery Files at their standard addresses: nine answering directly, and llm.txt redirecting to llms.txt.

Two things about that table. First, the rows map onto stages, not channels. "What is it?" matters at discovery; "how do I set it up?" matters at activation; "what does it cost?" at payment. The pages that sent the most website sign-ups were the home page and the pricing page, which is what you'd expect if people arrive knowing roughly what they want.

Second, I can't tell you which of these pages contributed to a single sign-up that began inside Claude. Somebody may have read the threat model on Tuesday and installed from the directory on Thursday. The machine-readable files are the same: I believe consistent, explicit facts, published in an llms.txt file and a structured identity.json, are a precondition for being described accurately by an AI. I can't show you that they caused anyone to be told about us, and I'm not going to claim it.

How to measure AI product discovery: a plan for software businesses

Here's what I'd tell any software business starting now, built from what we got right late and what we still haven't done. The principle is simple: record each stage separately, in its own unit, and never let one stand in for another.

1. Discovery: ask, in their own words. One optional free-text field at sign-up: "How did you first hear about us?" Don't make it a drop-down, because "ChatGPT told me" and "a colleague mentioned it in Slack" won't be in your list. Mailbox MCP doesn't ask this yet. It's the next thing I'd add, because it's the only record that can ever see an AI answer.
2. Arrival: record the door, once. At account creation, store the referrer's host, the landing page, any campaign tags on your own links, and whether the sign-up began at an AI client's consent screen. Write it once and never overwrite it. Links from your own site should carry a marker, such as ?from=site&lp=/pricing/, because many browsers strip the path from the referrer.
3. Sign-up: count accounts as accounts. Nothing more needs saying, except not to call them customers.
4. Activation: define the moment the product worked. For Mailbox MCP that's a mailbox connected and a first call made. Yours will differ. On launch day 12 of our first 17 sign-ups never got there, and the fix was a screen at Google's end, not a marketing change.
5. Payment and continued use: in their own units. We sell per mailbox, so paid is counted in mailboxes. Recent use is mailboxes with a call in the last seven days.
6. Read search and directory data alongside, as bounds. Search Console clicks put a ceiling on what organic search could have done. Directory search positions, read on a date, show whether a listing can be found for the words people use. Neither is attribution.

The stored record can be tiny. This is the shape of one account's first door, with nothing in it that identifies a person:

{
  "entry": "authorize",
  "ref": "claude.ai"
}

And this is enough to turn those records into the table above. It's a simplified version of the classifier Mailbox MCP runs, for a PostgreSQL users table with a signup_source JSON column:

-- Sign-ups since recording began, by the first door they came through.
SELECT
  CASE
    WHEN signup_source ->> 'entry' = 'authorize' THEN 'began inside an AI app'
    WHEN signup_source ->> 'from' = 'site'       THEN 'our website'
    WHEN signup_source ->> 'ref' ~ '(^|[.])(google|bing|duckduckgo|yahoo|ecosia|brave)[.]'
                                                 THEN 'a search engine'
    WHEN signup_source ->> 'ref' IS NOT NULL     THEN 'another referring site'
    WHEN signup_source IS NULL                   THEN 'unknown'
    ELSE 'direct'
  END AS first_door,
  count(*) AS signups
FROM users
WHERE created_at >= DATE '2026-10-05'
GROUP BY 1
ORDER BY 2 DESC;

The self-reported answer is the piece most people skip, and it's the one I'd fight for. Kevin Indig went looking for AI-driven visits hidden in "direct" traffic across a dozen sites and came back with this:

"I could not yet detect 'hidden' AI traffic without self-reported attribution..."

Kevin Indig, author of Growth Memo, in Does AI drive more direct traffic?, Growth Memo, 19 October 2026 (verify quote at source)

That landed with me, because it's the conclusion our own records reached by a different road. Our "Direct" row is 469 people we can't place, and our "Inside Claude" row is 1,436 people whose door we know and whose reason we don't. Neither gets clearer with a cleverer query. They get clearer when somebody types "my accountant uses it" into a box. It's humbling, slightly, to have built a careful classifier and to conclude that the most useful field is a free-text one, but I think that's right. Self-reported answers have their own faults: people forget, and they credit the last thing they remember. That's why the plan keeps them in a separate column from the door rather than letting either overwrite the other.

AI product discovery hypotheses worth testing next

Everything in this section is a hypothesis. None of it is a finding, and some of it will be wrong.

• ChatGPT sign-ups activate more because they arrive with a particular mailbox in mind. Test: compare activation by mail provider and by door, once a few more weeks of recorded sources exist.
• A good share of directory sign-ups learned of the product somewhere else first. Test: the free-text question above, read after a month, against the recorded door.
• Listing text matters more than listing presence. Test: change one directory listing's first line, hold everything else, and compare that directory's sign-ups for two weeks either side.
• Website sign-ups pay more because they're further along when they arrive. Test: compare payment by landing page, pricing against everything else.
• Organic search will matter more as the product ages. Test: the same Search Console bound, repeated each quarter.

What this AI product discovery case study cannot tell you

Plenty, so here it is in one place.

• It isn't independent. I founded the product and the publisher. I've tried to make every claim checkable, but you're reading the owner's account of his own growth.
• It's one product, in one category, over about two months. An email connector for people who already use Claude or ChatGPT is a narrow, high-intent niche. A product people don't already go looking for inside an AI app may see something quite different.
• The door was only recorded from 5 October. Everything before that is either the cohort count or the one-off log join, and the launch period itself has no recorded source.
• Paid is to date. The newest accounts have had the least time to pay, so the payment rates will move.
• The directory positions were read once. On one day, by a parser that counts card titles, so they are approximate.
• Search Console sees Google only. Bing, DuckDuckGo and the rest aren't in the 225.
• AI answers weren't measured at all, for the reasons above.

Practical lessons for software products that want to be found by AI users

This is what AI visibility for software products comes down to, as far as one case can show it. None of these will get you ranked, cited or listed. They'll help you understand what's happening, and make sure that when someone does find you, through any of the four mechanisms, the answer they get is right.

1. Treat the four mechanisms as four jobs. Pages for search, listings for directories, a working server for availability, and consistent public facts for whatever an AI might say. Success in one tells you nothing about the others.
2. Get technically available first. A remote server that a reviewer can connect to, with clear tool descriptions, is the price of entry to every directory.
3. Write the listing for the directory's search box. Lead with the words your buyers type. Ours went from invisible to near the top for "hotmail" with the same product and a better first line.
4. Don't confuse a listing with a recommendation. Both Anthropic and OpenAI describe directories you're found in by looking, and suggestions that apply to apps already connected or to distribution you can't request.
5. Answer the evaluation questions in public. What it is, whether it works with my setup, how to install it, whether it's safe, what it costs, who's behind it. Each is a separate page, and each serves a different stage.
6. Record the door and ask about the discovery. Two fields, kept apart. The first you can automate; the second only a person can fill in.
7. Measure activation as hard as acquisition. Our biggest early losses weren't in discovery at all. They were at a password screen on Google's side.
8. Write down what your evidence can't establish. It's the column that stops you investing in the wrong channel because of a coincidence of dates.

If you run a website rather than a connector, the inputs side of this is what the free AI Visibility Checker tests: whether your site publishes consistent, machine-readable facts that AI systems can find and read. It won't tell you who recommends you. Nothing reliable can. The validation guide sets out what can be tested and what can't, and if your site already publishes AI Discovery Files you can submit it to the directory for free verification. 365i hosts the site you're reading, and the 365i web design team builds sites with these files in from the start.

Frequently asked questions