Why Vanity Metrics Kill AI Startups: 700 Customers and 60,000 RMB From One Niche Account

Written with AI assistance. Figures without a traceable source were cut before publishing.

The same piece is posted as a thread on GitHub — that copy has a reply box under it, and this one does not.

In special-purpose vehicles — tankers, sweepers, the trucks nobody films for fun — one effective sales lead costs somewhere between 100 and 1,000 yuan, and a buyer who actually signs wires a hundred thousand or more. I think that single ratio is why the follower dashboards in this category lie to you, and why every acquisition tactic borrowed from consumer apps inverts the moment you carry it into a low-frequency industrial niche where the whole country holds maybe a few thousand real buyers. A handful of operators in this category have built their funnels around that inversion, and their numbers are unusually specific, so I want to walk through them and mark the places where they stop being checkable.

Breaking the Traditional Trap: Why Generic Followers Fail in High-Ticket Niches

The trap is that the generic playbook still looks like it is working.

Scroll the category and you find the same video on every account: a slow pan around the vehicle, background music, someone reading engine displacement, wheelbase and tonnage off a spec sheet. Prospects swipe away inside three seconds, and I’d argue the reason is not production quality but the absence of a person — nobody wires a hundred thousand yuan to an account that recites parameters without once explaining what those parameters do for a business that hauls water or sweeps roads. Everyone stacks specs. Nobody translates them. The view counts stay respectable while the pipeline stays empty, which is the exact failure mode that makes vanity metrics dangerous rather than merely useless.

So the first move is arithmetic, not content. If a qualified lead runs 100 to 1,000 yuan and the deal behind it is six figures, then the budget question stops being “how do I get cheaper impressions” and becomes “how many of these can I verify per week.” I think most teams never make that switch, honestly, because impressions are the number the platform hands you for free.

Volume has an honest use, though. It just isn’t the one being sold.

The builders getting results treat AI as the thing that runs the volume inside boundaries a human wrote down, and they wrap it in three principles: customer-centricity first, human-defined standards with AI executing at scale, and an independent quality-inspection layer that is not the same system that produced the work. Miss the third and you have an agent grading its own homework. I suspect that is where most of these projects quietly break, though I can’t prove it from the logs I have.

There is a version of this that scales past one operator. Enterprises chasing the same shift have to stop treating AI as a copywriting expense and tie it to metrics the CFO already tracks — inventory turnover, new-product success rate — and a niche narrow enough that the tie is legible. Elsewhere, developers generate local SEO pages with AI and rent the finished sites to local merchants for 500 to 3000 dollars monthly, which is the same structure at a smaller unit price: narrow buyer, obvious value, no follower count involved.

Scale of the target changes nothing about the method. Breaking a $1,000,000 annual revenue goal into a daily intake of $2,777 — eighteen standardized units at $150 a day — is the same discipline applied with a bigger numerator, and it works for the same reason: it converts an ambition into a countable daily event. Pair it with hard demand validation, where every release is aimed at a buyer whose willingness to pay has been confirmed rather than assumed, and the follower count stops being a variable anyone tracks.

Building the Automated AI Knowledge-Base Funnel and Lead Routing System

Now the part with receipts.

An automated intake system handled 287 users, and 158 of them actually made it into the group — the entry rate went from 9.1% to 55.1%, and the median time from adding a contact to landing in the group dropped from 164.84 minutes to 13.88 minutes. Read those two numbers together, because separately they are less interesting than they look: the conversion did not improve because the content improved, it improved because the gap between a stranger’s first click and a human reply shrank from nearly three hours to under fifteen minutes. High-intent industrial buyers do not wait around in that gap. They go back to searching.

Routing is where the leakage hides. Unmanaged groups held 83.1% of the total group inventory and contributed 2.3% of new incoming traffic; the managed ones recorded 170 new entries against 4 from everything else combined. I find that split clarifying, because it kills a plausible-sounding worry — the unregistered groups are not a silent reservoir of missed demand, they are just storage.

Storage is not a funnel.

Building the thing is the least mysterious step, and I’d argue it is the one people over-engineer. An eight-step framework from project setup through validator hardening produces an agent that handles standard operations; the sequencing advice attached to it is that you build a minimal demo, test the core interactions against real business constraints, and only promote to production once the reliability holds under live use. Going straight to a full deployment is how you end up with architecture you cannot debug and a conversion pipeline that fails in ways nobody instrumented.

The same ordering holds inside companies, reversed from what most rollout decks propose. Let staff run personal agent instances on low-risk, high-frequency daily tasks first and let the genuine usage logs accumulate; the centralized knowledge base, the permission hierarchy, the evaluation monitors all get built afterward, out of patterns that actually showed up. Top-down first is how you get a governance layer for work nobody does.

One design detail is worth stealing outright: keep platform API responses and business facts in separate layers. Systems love to treat an order query as a qualification gate when, in the real commercial process, an order is just an early trigger — and conflating the two produces automated rejections of qualified buyers that read, to the buyer, as a door closing for no reason.

Engineering the Execution Framework: Human Standards and Independent Quality Control

Documentation is the boring answer, and it is probably the right one.

Standardizing the requirement document before scaling token-heavy work eliminates the rewrite loop where an agent re-derives the same context every run. Define the service boundary, the current functional version, what is explicitly out of scope, and the project constraints that exist but were never written down — that document does more for output quality than another round of prompt tuning, because an agent hitting an undefined requirement does not stop, it invents. To be fair, I think the prompt-versus-spec framing is overstated online; both matter, and the spec is simply the one people skip.

Multi-agent setups are org design wearing a technical costume. The failure is letting a single agent define the requirement, implement it, and then declare it finished, which removes independent verification from a process whose entire value was independent verification. Split the roles the way you would split them among people, and keep the inspector separate from the builder.

For solo builders the same principle shows up as a budget constraint. A minimum viable demo iterated against real operational data costs a fraction of what a speculative full build costs in tokens and rework, and I suspect the token bill is what finally teaches this lesson to people the process arguments never reached.

And when a manual process becomes an automated one, the human-defined standard has to survive the transition intact. Otherwise the efficiency gain is real and the business intent quietly is not.

Scaling and Verifying the Micro-Funnel

Here is the outcome the rest of it was building toward, and it is smaller than the headline numbers people usually quote at you.

Running a closed loop across a content platform, visual assets, and a private community, a vertical sub-account booked over 700 customer conversions and 60,000 RMB in gross merchandise volume within 3 months. That is roughly 86 yuan per conversion — a modest average order value, from an account with no follower count worth screenshotting. What makes it interesting to me is not the size but the shape: precise content, one narrow audience, a routed funnel, and a closed measurement loop from first touch to payment.

Speed of iteration is what keeps that loop alive. A modular production pipeline built on Feishu, Obsidian and scripted automation cut document management from hours to minutes, and a single trigger word now generates the draft, the illustrated body and the promotional copy in one pass, replacing the 2 to 3 hours of daily manual compilation that used to sit in front of every publish. I think that recovered time is the actual asset here, more than any individual automation.

Small and verified beats big and vague, and the comparison case makes the point better than I can: StoryShort, an AI short-video tool, matched in 3 months the cumulative revenue that the B2B tool useArtemis took 2 years to accumulate — around $22,000 in monthly Stripe-verified revenue against nearly $500,000 cumulative. Narrow execution against a specific buyer outran broad scale.

One honest caveat before you copy any of this. Every figure above comes from operators reporting their own results, and the entry-rate and timing numbers are the only ones with a real before-and-after attached; the revenue figures are single snapshots, and I could be wrong about how much they generalize. What I would take from them regardless is the ordering: fix the routing, measure the gap between click and reply, and count conversions instead of followers.

Also readable on Telegraph.

Read next

All 49 write-ups

The 15 figures in this piece — each with the sentence it came from — are in the figures table, alongside 393 more, as JSON and CSV.

Topics: Automation Systems · Niche Market · Revenue Growth

Part of llm-api-pricing — field notes on AI coding agents.

Did this save you an afternoon? A star on the repository is the whole ask — it is what puts these in front of the next person looking; the data is CC BY and does not require starring.

One thing this piece could not settle: How many minutes does it take for a human reply to a stranger’s first click in your niche, and is it under 15 minutes? Reply with a yes or no. The reply box is on the thread copy of this piece.

Want a figure that is not in here yet? Say which metric, which provider, which unit — in one line. One required field, and the page you came from is already filled in. Got a better number? Open an issue — that form knows which write-up you came from too; corrections and counter-data are the point.