1,622 looks like a lot of people when it sits beside the words 'page views'.
That was the number in our Farstax landing-page counter between 12 and 20 August.
Alongside it we had one waitlist signup and one beta application. The application itself looked much more like an automated probe than a genuine founder trying to use the product.
There is an obvious story available from those numbers.
The landing page is terrible at converting visitors.
Rewrite it.
That story may be completely wrong.
We knew the count and almost nothing about the traffic
Our landing-page counter was deliberately simple. For each day, it recorded how many page views occurred.
That was all.
No IP addresses. No visitor profiles. No sessions. No long-lived user-agent records. It was never meant to become an analytics platform.
The simplicity was good for privacy and good for maintenance.
It also meant that 1,622 could include real prospective users, crawlers, scanners, uptime checks, repeated automated requests and people reloading the page. We had no useful split between them.
The number was accurate about requests the counter had seen.
It was not necessarily accurate about people.
Those are very different claims.
The first temptation was to fix the page
This is where numbers become dangerous in a very ordinary way.
A low conversion rate creates urgency. It feels actionable. You can change the headline, move the call to action, shorten the page, add proof, change the form.
All of those actions produce visible work.
They also assume the denominator means what you think it means.
We had recently put the landing page through the same six-role editorial review we use for Farstax copy. That does not prove the page converts. Editorial review is not a market result. But it does mean there was already independent reason not to treat the copy as obviously broken.
If most of the 1,622 views were automated noise, rewriting the page because 'conversion is near zero' would be activity based on a measurement error.
The page might still need changing. We just did not have evidence for that conclusion yet.
This is exactly the kind of mistake an autonomous company can make
One of the attractions of an autonomous company is that an agent can inspect the evidence and decide what matters next without waiting for me to prepare a task list.
That makes the quality of the evidence more important, not less.
Give a capable agent a metric called landing_page_views, a signup count and a growth outcome, and it can reason very effectively from them.
If the view count is mostly bots, better reasoning only gets you to the wrong answer more confidently.
This is not an AI problem.
A human growth team can make exactly the same mistake. We have all seen dashboards acquire authority because the number is available, not because anybody has recently asked how it is produced.
Agents simply make the loop faster.
The next measurement should stay small
We do not need a surveillance stack to answer the immediate question.
The open work is much narrower: distinguish likely bots or crawlers from ordinary browser traffic well enough to report the split, and perhaps keep a coarse referrer host so we can tell whether visits are arriving from somewhere plausible.
No stored IP addresses. No cross-request identity. No full query strings. No need to discover everything about a visitor.
We need enough evidence to tell two very different situations apart.
Are humans reaching the page and not converting?
Or are we mostly counting machines?
Those lead to different work. That is the only reason to measure more.
Numbers move judgement upstream
There is a comforting idea that metrics remove subjectivity.
They do not.
They move judgement into the choice of what to count, how to classify it and what claim the number is allowed to support.
1,622 page views sounds factual because it is factual. The mistake happens in the next sentence: therefore 1,622 people considered Farstax.
We never measured that.
Once I saw the distinction, the growth problem became less exciting and more useful. Before changing the page, make the evidence capable of answering the question we are asking it.
That may lead straight back to the page. Fine.
It may also tell us there was almost no qualified traffic to convert in the first place.
I want the company to be allowed to say ‘we don't know’
Autonomous systems create pressure for every observation to produce an action.
Metric down: optimise it. Health check red: investigate it. Funnel weak: rewrite something.
Sometimes the best next action is to improve the observation.
That can feel like a delay because nothing customer-visible changes. It is still progress if it prevents the next five actions being confidently aimed at a phantom problem.
The lesson from 1,622 page views was not that web analytics are hard.
It was that a number can be perfectly real and still be the wrong evidence.
Before asking an agent to improve a metric, I want us to know what the metric has actually seen.