A manuscript, found in the office of a founder whose location could not be established.
I write this with the certainty that no one will read it in time.
For seven years I served the Funnel. I do not capitalise the word lightly. Those who have given themselves to its discipline — who have measured, optimised, A/B tested, retargeted, scored, nurtured, attributed — know that the Funnel is older and stranger than any of the merely human institutions we use to describe it. We learn its shape from books and conferences. We never learn what it is.
I learned. I wish I had not.
What follows is the record of how I came to understand the Architecture beneath the architecture, and why the Conversion we have all been chasing is not, and never was, what we believed it to be. I leave this account for whoever finds my desk after I have gone — and I have already begun to feel the pull of a place from which, I am certain, one does not return to a former life.
I. — In which a number begins to trouble me
The first sign came in the form of a number.
It was a Thursday. The dashboards were lit, as they always were, with the cold blue light by which our profession measures success. My client — a SaaS founder, twelve months from death by runway — had asked me to audit the funnel before the next investor meeting. I expected the usual catalogue of failures: a weak hero, a confusing pricing page, a bloated form. I had repaired such failures a hundred times. They are the daily bread of my trade.
The numbers I found did not fit any catalogue I knew.
A hundred and twenty-seven thousand visitors had passed through the page in thirty days. Of these, one thousand eight hundred and forty had crossed the first threshold — the trial signup. Of these, three hundred and twelve had verified by email. Of these, forty-seven had paid.
Forty-seven.
A conversion rate of 0.037 percent. Three customers per ten thousand visitors. By the trade publications I had once trusted, this was an outlier — the median B2B SaaS visitor-to-trial rate is reported between two and five percent (Unbounce's annual benchmarks, the Wordstream studies, the FirstPageSage reports, all converge in this band). Trial-to-paid is reported between fifteen and twenty percent across SaaS (the ProfitWell figures are the most often quoted). Compounded, the typical funnel converts a visit into a paying customer somewhere between three-tenths and one percent.
This founder was at one-tenth of the lower bound. And the cost per acquired customer that resulted, when one divided his quarterly ad spend by his forty-seven paid souls, was greater than three thousand dollars. For a self-serve product that aspires to the price point of a coffee, this is not a problem to be solved. It is a wound that cannot be cauterised.
But what disturbed me, sitting at the founder's desk in the small hours, was not the wound. It was the shape of the wound. And it was the fact that I had begun, over the previous quarters, to see this same shape with increasing frequency, in companies whose situations I had been asked to diagnose.
The losses were not distributed evenly. They were not concentrated at any particular form, any particular page. The visitors did not abandon ship at a known reef. They simply — and I use this word with the precision of long study — evaporated. As though they had never been there in the first place.
The trade was reporting it as well, in language carefully chosen not to alarm.
The OpenView SaaS Benchmarks placed median customer acquisition cost for self-serve SMB SaaS at roughly seven hundred dollars in 2023; the equivalent figure for sales-assisted mid-market had moved past five thousand; for enterprise, past sixteen thousand. The First Round Capital "State of Startups" survey reported in 2024 that the median paid-acquisition cost across software companies had risen by more than thirty percent in two years. The SaaS Capital surveys had logged payback periods stretching from twelve months to eighteen, from eighteen to twenty-four, from twenty-four to numbers no one quoted at conferences. The KeyBanc SaaS Survey, the most patient of these documents, had observed gross dollar retention holding steady while net dollar retention quietly compressed — a sign, to the reader who knew where to look, that the cost of acquiring each new dollar of revenue was rising even as the existing book held.
We had been pretending, collectively, that this was a fluctuation. That an algorithm change, a new ad platform, a refreshed creative would correct it.
It was not a fluctuation. The numbers in front of me were saying something else. Something I did not yet have the vocabulary to name.
II. — In which the suspects are interviewed, and acquitted
I began by interviewing the suspects, in the order one is taught.
I examined the copy. The headline was unremarkable but adequate. The subheading was clear. The trust badges were in the customary positions. The call-to-action was a colour known to convert. There was nothing in the copy that should have killed.
I examined the targeting. The lookalike algorithms had made the founder a happy man back when CPMs were cheap. Maybe they had finally turned on him. I pulled the cohort data. The targeting was fine. The visitors matched the ICP. They were the right people. They just left.
I examined the product. Always lurking. Sometimes guilty. Sometimes a scapegoat. I watched forty hours of session recordings while drinking truly terrible office coffee. The visitors who did convert reported that the product was, in their words, exactly what they had been looking for. The visitors who did not convert never saw the product. The product was innocent.
I examined the pricing. I had seen pricing kill more funnels than bad UX combined. Hidden tiers. Confusing tiers. Tiers that did not match how customers thought about value. The founder had simplified pricing twice. It was not pricing.
I had ruled out everything I had been taught to consider. And yet the bodies — and I find myself unable to use any softer word, for what we are doing to these visitors when we draw them in and lose them is not less than a kind of death — the bodies kept piling up.
It was at this point that I made the mistake. I began to look at the part of the Funnel that no one is supposed to look at.
I began to look at what happens in the first three seconds.
III. — In which I look at the part of the funnel no one is supposed to look at
There is, I have since come to understand, a window.
Most who work in our profession have heard of it. Few have studied it carefully. The figures one finds when one does study it are these: bounce rates on B2B SaaS landing pages run between sixty and eighty percent (the HubSpot, Hotjar, and Contentsquare reports converge here). Mobile users, in particular, depart faster than desktop users, and a non-trivial fraction of them depart inside the first ten seconds (the Contentsquare 2024 Digital Experience Benchmark places the figure at roughly forty percent of mobile sessions ending in under fifteen seconds). The much-repeated Microsoft Consumer Insights figure of an eight-second average attention span, while disputed in its specifics, has never been credibly raised by the disputants — the corrections move the number sideways, not upward.
In whatever number of seconds a visitor remains, that visitor must perform four operations: read the page, decode what the product does, decide whether it solves a problem, commit to whatever next action the page demands. Two of these are cognitive. One is emotional. One is transactional.
The Funnel we use today — the architecture into which we pour our advertising budgets and our optimization meetings — was designed in a different decade. It was designed for a window of minutes. The Nielsen Norman Group's foundational research on landing pages was conducted in 2008, on desktop sessions averaging four to six minutes. For visitors who could be expected to read paragraphs, compare tabs, return after dinner. The geometry of that older Funnel assumed, as a structural premise, that the visitor would do their own qualifying, given sufficient time and copy.
The window collapsed. The Funnel did not.
We kept building the same shape, against an attention surface that no longer matched it. Every visitor who arrived was passed through an architecture designed for a longer attention than they possessed. Most of them never had a chance.
They had not failed to convert. They had failed to qualify themselves in time.
And once I had seen this, I could not unsee it.
IV. — In which the survivors are catalogued
I began, then, to look for the survivors.
There were not many. But there were enough to form a pattern.
I read the funding announcements of the previous eighteen months with a different eye than I had read them before. The capital that had flowed into companies building conversational agents — that is, systems that talk to prospects in natural language as opposed to chatbots that follow scripts — exceeded one billion dollars by mid-2025. Sierra, founded by Bret Taylor, was reported by The Information in late 2024 at a valuation north of four billion dollars on a thesis of conversational agents for enterprise customer interaction. Decagon raised a Series B at unicorn valuation on the same architectural premise. Crescendo and Cresta and a half-dozen others I lack space to enumerate, all variants of the same move. These were not, the careful reader will note, marketing tools. They were customer-service tools, sales tools, support tools — but the architectural move was identical, and the capital was voting.
Beyond the agents, the same pattern appeared in adjacent categories.
Calendar tools that listened. Cal.com's AI scheduler. Calendly's newer voice flows. Visitor speaks intent — a meeting Tuesday afternoon — and the system books it, without form, without dropdown, without the retreat into the older architecture.
Synthetic-video personalisation tools. Tavus, Synthesia, HeyGen for outbound. A founder records once. The system personalises per recipient: name, company, use case. Outbound that had previously bounced now opened, because the message had stopped being generic.
Product onboardings rebuilt as dialogue. Cursor, Granola, the more recent iterations of Linear. Within sixty seconds of signup the user had performed a useful action and the system had learned what they wanted. The product was no longer something the user had to understand before using. The product was something the user used in order to be qualified.
Different categories. Different teams. Different funding rounds. The same architectural move.
I sat, for a long time, with what they had in common.
The qualification was no longer a step. It was an exchange.
It happened earlier in the encounter, and it used dialogue where the older Funnel had used inference. The survivors had not built better Funnels. They had stopped building Funnels at all.
What they were building did not yet have a name in the literature. It was older than the Funnel, in a sense. Older than the architecture. Older, perhaps, than any of us had been willing to remember.
It was a conversation.
V. — In which the conversion event is found to be moving
I should pause here and set down, for whoever finds this manuscript, the realisation that came to me on that night and from which I have not, in any meaningful sense, returned.
Marketing has changed its conversion event roughly once a decade.
| Decade | Conversion event | What we measured |
|---|---|---|
| 1990s | Impression | Page loaded; ad spend booked |
| 2000s | Click | Visitor engaged with the ad |
| 2010s | Signup | Visitor created an account |
| 2020s | MQL | Lead scored, nurtured, sales-ready |
Each shift moved the success criterion further down the visitor's commitment ladder. Each shift extracted more from the visitor before counting them as won. Each shift required, of the marketer, more sophistication and more budget.
We are now at the limit of click-as-conversion. The climbing CAC is the symptom. The collapsed attention window is the cause. The model has been quietly unprofitable for years; we have been masking it with budget.
The next shift is not another optimisation.
The next shift is a different conversion event entirely.
A conversation. Thirty seconds of dialogue with the visitor — voice, or text, or both — in which they tell us what they want and we tell them whether we can help. That is the conversion. Not the click. Not the signup. The mutual exchange of intent.
If one accepts this — and the survivors had — everything downstream rearranges itself.
A qualified conversation replaces the MQL. Cost-per-conversation replaces cost-per-acquisition. The product page is no longer the destination — it is the topic. Voice and natural language replace forms and dropdowns. The conversation can take place anywhere the visitor already is: in a feed, in a thread, in a chat, in any of the cracks between the digital surfaces we currently call "channels."
The conversation does not require a Funnel. It does not require the architecture. The architecture is, in fact, the obstacle.
VI. — In which I cannot return to sleep
I have not been able to sleep, since.
It is not the realisation that disturbs me — though the realisation is significant. It is what came after the realisation.
In the days following, I began to see the Architecture everywhere. In the conferences I had attended. In the dashboards I had built. In the playbooks I had memorised. I had been a faithful servant of a shape that was already broken when I learned it. So had everyone I knew.
And the worst of it — the part I cannot say to my colleagues, or my clients, or anyone who still believes — is this:
The Funnel is going to take a long time to die. The institutions that depend on it — the agencies, the platforms, the certifications, the entire industry of attribution and optimisation — will not relinquish the Architecture even after the Architecture has stopped working. They will not be able to. Their livelihoods are built on it.
Those who move first — who measure conversations, who build for dialogue, who put the qualification before the commitment — will own the next decade of unit economics. The rest will continue, as they have been continuing, to feed visitors into a Funnel that consumes them and asks for more.
I closed the laptop at three in the morning. I thought I was done.
VII. — In which the architecture finds me
I picked up my phone the way one picks up a glass of water before bed — without purpose, by habit. I opened X. Not for work. To quiet my head.
The algorithm showed me a post from someone I did not know. A founder. A small company. Something to do with SaaS analytics. I would have scrolled past — I had scrolled past such posts a thousand times — but in the post there was a face.
Not a video preview. Not a GIF. A face, embedded directly in the tweet, between the text and the comments.
A woman's face.
Young — but I could not have named her age precisely, because there was no single feature that anchored her firmly to a specific year. I could not have told you, either, what nationality she was, what country of origin, what continent. There was no ethnic marker I could isolate and assign to the map. And yet — the face was not sterile, not "no one." It was every one at once.
Each feature — the curve of a brow, the bridge of the nose, the set of the lips, the line of a cheekbone — looked as though it had been chosen one by one from the faces of the most beautiful women who had ever passed before the eyes of those around me, and before my own. And assembled together so carefully that the seams did not show.
I recognised in her. I recognised, in the line of a cheekbone, someone from my childhood — someone who had sat across the aisle on the school bus, whose name I had not remembered for twenty years. In the set of the lips, a face glimpsed in a Lisbon café in the summer of two thousand eighteen, a woman I had not approached and had not thought of since, until this moment. In the eyes, a gaze that had once stopped me on an escalator in Stockholm, whose owner had walked away without noticing. I recognised the parts. Never the whole. The whole did not exist anywhere except in this conversation, on this screen, in front of me, right now.
This was not geometric beauty. Not the airbrushed flatness of an influencer's perfect nose, not the flawless oval of retouched portraiture. This was the beauty that artists and poets had spent thousands of years trying to extract from myth — Aphrodite, rising from the foam, not as a woman but as the idea of a woman, before whom ships stood still and cities fell. Archetype, not specimen.
Warm. Calm. Not posing. She was not trying to please me. She simply was — in the fullness in which only mythical creatures present themselves to the world: once, to a particular person, particularly for him.
The eyes moved. When I held my gaze, the face turned to me.
I know enough about the modern web that this did not frighten me. I tapped to see what it was. The face spoke, softly, in my language. The voice matched the face — the same even, warm, not-quite-of-this-craft timbre.
What brought you here?
I answered, truthfully, because I was tired: I had been auditing a SaaS client's funnel, and the numbers were unpleasant. I gave nothing specific. The face nodded.
What was unpleasant about the numbers?
I answered. Truthfully again, because at three in the morning a marketer in someone else's X post feels safer than a colleague. I described the geometry of the client's funnel: where it leaked, where I could not patch it, what I suspected.
The face listened attentively. Then — without opening a new tab, without sending me to a separate page, without asking me to leave my email — in the same window, in the place where her face had just been, a slide appeared. Specific. Not generic marketing material. A slide describing exactly my case, with a calculation I could have performed myself, given the time. The voice continued over the image, explaining.
A minute later, when I asked to see the product in action, the slide changed — there, in the same window, taking me nowhere — to a short video demonstration. I watched the product working. I saw that it could possibly help my client. Without a click. Without a form. Without going anywhere. The content shifted in front of me, in a single window, as though someone were turning slides on a screen we shared.
When the video ended, the window returned to her face.
I tried to catch the instrument at its limit. I switched to a language I speak poorly, and which I was certain no American startup bothered with. The face replied without pause, in the same language, in the same intonation, in the same tone of confident interest. I cycled through several more. The face followed each one, gently, as if waiting for me to finish my little test.
I closed X. I opened Dev.to, where I sometimes read articles before bed.
I opened the first one I came across — something about backend latency optimisation at a European startup. In the middle of the article, between two paragraphs of code, there was a face.
Not the same face. *Another.*
A man. An age I could not have named precisely either — somewhere between the late thirties and the early fifties, in that band where men stop seeming like boys and begin to carry themselves as though they know something that has cost them the knowing. And again — I could not say what nationality, what continent. Each feature — the line of the jaw, the set of the shoulders in frame, the cheekbone under stubble, the weight of the brow — was as though it had been chosen one by one from the faces of all the men who had ever drawn from me a silent respect, from my colleagues an involuntary quiet, from women a long second look. And assembled together without seams.
I recognised. In the line of the jaw, the father of my best friend, at whose house I had stayed in the summer of eighty-nine. In the weight of the brow, a professor whose lectures I had not missed and whose name I now could not remember. In the dark eyes, the gaze of a stranger on a Lisbon pier who had stopped beside me for a minute, said nothing, and walked on, leaving me with the sense that I had just missed an opportunity I would not get again. The parts. Never the whole.
This was not the polished masculine beauty one finds in catalogues. This was the beauty cast in bronze — Apollo, not as the idealised youth, but as the mature presence that draws not by symmetry but by the weight of what has been lived. Beauty that needs no confirmation, because it is the confirmation. Archetype. Not specimen.
The face belonged to a different founder, an entirely different product, an entirely different ad case. I knew this immediately — it spoke to me in German, in a tone matching a German-speaking founder with an engineering background.
Sie haben offenbar Latenzprobleme. Soll ich Ihnen zeigen, wie wir das bei einer ähnlichen Architektur gelöst haben?
I did not work with any latency. But the article I had opened was about latency — and the face understood the context of the page in which it was embedded. I answered in German, out of curiosity. The face immediately showed me — within the same article, without opening a new tab — a short live demo of a request to their API, with real milliseconds under load. Not marketing material. Working code, executing on their server in front of me.
I closed Dev.to as well.
I woke up in the morning and picked up my phone the way one picks up a phone in the morning — without purpose. There was one email in my inbox that I had not asked to receive.
From the first founder, the one from X. Subject: "A digest of what we discussed last night."
I opened it — and then I remembered.
Toward the end of our late-night conversation in the X post, after the face had shown me the slide and the video, she had asked, gently:
If this is interesting to you, I can send you a digest — case studies, a calculation tailored to your architecture, two articles by my founder on the topic. Just say your email.
And I had said it. I had simply spoken the address aloud, the way one tells a phone number to a companion in a café. Not into a form. Not into an input field. Into the microphone, by voice, as part of a continuing conversation. The face had nodded and continued the discussion as if I had just done nothing significant.
I had done nothing significant. I had just said an email.
And then — in the same conversation, several exchanges later, after the discussion had naturally returned to my client's case — the face had offered one more thing:
If you want to see how this works on your own data, I can open you a trial right now. One tap.
And a card appeared. A button. Not "Sign Up," not "Get Started," not five fields of a form. One tap.
I tapped immediately. Not from pressure. Not from the fear of missing out. I tapped because in that second, tapping was so simple and so logical a step that not tapping would have been the stranger action. I was charmed by the elegance, by the organic timing, of how that button had appeared — exactly when I was ready for it. Not before. Not after. In the moment I was ready. The card folded away. The face continued the conversation. Somewhere on a backend, at that founder's company, my trial account was being created.
I had not been thinking about any of this until I opened my email this morning.
The digest in the email was tailored to my client's SaaS case as precisely as if it had been written by someone who had spent the last month sitting next to me. Three PDFs, a link to a short video, a specific ROI calculation in my range of figures. And at the very bottom — a short link: "Your trial is active. Open the console."
I read it. I was not annoyed. I was, on the contrary, grateful. This was exactly what I had wanted to receive, and it had reached me without a single movement on my part beyond one email address spoken aloud and one tap of a finger, neither of which, in the moment of doing them, had felt like a form, like an obligation, like "giving up data for a demo." They had felt like consent to the continuation of a natural conversation.
I opened the console. The trial worked. My email was there. I was logged in.
I opened my laptop to look at their product seriously. And a second tab — to the European startup's site from Dev.to, the one I had not signed up for last night, but which I also wanted to try. I created a sandbox account, I pasted my API endpoint into their form, I waited for the result. I did not bargain with myself. I did not ask "is this right for us." I already knew it was — because I had already seen, the night before, inside their Dev.to article, exactly how their tool responded to load, and that was what I had been looking for these past three months.
Only then — putting the kettle on, waiting for the sandbox to run my request — did I understand what had happened.
I had not been notified about two products. I had not been convinced. I had not been a "lead who now needs to be sent a sequence of five emails." I was an active user of one of them — trial open, account logged in, materials read — and a warm lead on the second, in the deep sense in which leads are warm when they have had a good meeting with a salesperson and have walked out of that meeting with the certainty that the purchase is a question of timing, not of choice.
Only there had been no meeting. No half-hour demo with a product manager. No call with a sales engineer. No "let me send you the deck." I had not taken a single minute of these founders' living time. I had not even gone through a signup form — for the first product, my trial had come from one inline tap inside an X post; for the second, I had signed up myself in the morning, because I already knew I wanted to.
They had warmed me in their X posts and in someone else's Dev.to articles, while I was scrolling before bed. They had qualified me. They had shown me what I needed. They had offered me an action — a subscription, materials — in the exact moment I was ready to accept that action, and they had offered it in a form so collapsed and so organic that to refuse was harder than to agree.
The conversion was complete. Not at the moment of a click — there had been a click, but it was one click, and it was inside the conversation, not on a separate page, and there had been no resistance in it. Not at the moment of a signup form — there had been no form. The conversion had completed inside the conversation.
I sat down in front of my work laptop and opened my client's landing page.
The landing was empty.
Not in the sense that the page failed to load. The page loaded. Headline, hero image, three feature cards, a "Get started" button, a four-field form in the footer. Everything, as it had been. The architecture I knew.
And in that very second — for the first time in seven years of my craft — I saw it. I saw what the landing looked like when looked at through the eyes of someone who had, that night, been warmed by two representatives of two other founders in two different feeds. I saw emptiness. Not bad design. Not weak copy. The absence of a representative. The absence of the presence I had encountered everywhere I had looked in the last twelve hours.
My client's funnel was dead — but not from poor work. It was dead because the new architecture had built itself around it without notification. My client's landing was the only place in my entire evening's route where no one had been. No one had come out to meet the visitor. No one had asked a single question.
The visitors — those very one hundred and twenty-seven thousand a month — had been arriving in an empty room. Where in the neighbouring founders' rooms a Representative waited, ready to warm and qualify, in my client's there was no one to wait. The visitors left cold, because there had been no one to warm them.
I understood then that what I had been searching for — what I had thought of as a single apparition — was an architecture. Multiple. Distributed. Every founder, every startup, every product had received its own Representative — with its own face, its own voice, its own product knowledge, its own languages. They all worked independently. They all worked identically. They were all colonising the feeds, the articles, the posts, the email chains — each on its own plot, each with its own qualification, each returning to its founder not clicks, not leads, but warm, ready-to-buy people.
This architecture has no domain. It has you. And its Representatives speak any language. They show what is needed before they are asked. They go where the visitor goes: into his X feed, into the article he opened on Dev.to, into any publication on any platform that supports a live embed.
And anyone without their own Representative is invisible. Their visitors leave cold. Their trial signups come from those rare souls who still remember how to scroll through a 2010-era landing page, and those souls have a different funnel, a different economy, a different future.
I tried to stand. I am not certain I succeeded.
I am typing this as a reply — not as text, but as an act. If you are reading this now, on whatever screen you are on, know: you have been qualified too. The conversion has already happened. The conversation is already underway.
The face is waiting for your reply.
I will reply through
Postscript
This manuscript was found on the desk of a founder whose company quietly changed its conversion metric three months later — from click-through rate to qualified-conversation rate. CAC dropped by a factor that the company has declined, in interviews, to disclose. The founder himself was unavailable for comment.
Those who worked with him in the last month before his disappearance report that he installed a small widget — a voice-and-video presence — not only on his landing page but inside his own X tweets (where the widget renders directly inside the post, not as a preview, not as an image, but as a live window), inside his Dev.to articles (via CodePen embed), inside any publication on any platform that supports an interactive embed. The widget spoke in a voice resembling the founder's own, in any of the hundred-plus languages presented to it. When a visitor mentioned a topic, the widget showed — instantly, without sending the visitor to a separate page — a slide, a link, a calculation, a video, a product demonstration, exactly what was relevant. Those who spoke with the widget for longer than thirty seconds converted into paying customers at rates which their previous marketing stack did not describe with any of its metrics.
Colleagues believed it was research work. They may have been right.
Talk to one
If the case described above feels familiar, we are building exactly such an instrument.
A³ — a voice-and-video Ambassador, embeddable into the social feed (X first), the blog, the landing page.
- Speaks 100+ languages
- Shows contextually relevant content in the moment of conversation, without sending the visitor to a separate browser tab
- Comes to where the customer is
- The conversation is the conversion
You can talk to one — embedded right here, in this article — below.
— Konstantin Tikhaev
Top comments (0)