Every number on this site has to trace to a build we can point at — a count you could go and check, rather than an average you could not. Here is what that rule costs us, and why we keep it anyway.
Most agency sites carry a version of the same sentence: our clients see an average multiple return inside a short window. The variable parts change. The structure never does.
Read one and ask what you would need in order to disprove it. The client roster. The baseline each of them started from. The definition of "return" — revenue, profit, booked jobs, impressions. The definition of "client" — everybody, or everybody who stayed. The window, and whether it was chosen before the results came in or after.
None of that is available to you, and none of it is going to be. The sentence is not necessarily a lie. It is something more slippery: a claim built so that nothing can land on it. You cannot check it, so you cannot disagree with it, so the only thing left to do is decide whether you like the tone of the website.
Numbers drift in that direction on their own, without anyone deciding to deceive. The bigger and vaguer a figure, the better it reads and the harder it is to verify, so at every fork there is a softer, rounder, more flattering version of a true-ish sentence sitting there. Take the softer branch often enough and you end up with a page where nothing is checkable and nothing, technically, was invented.
We wrote the opposite rule into this site before we wrote any of the copy. Every number on it has to trace to something real: a metric on a case study, a stage in a system we have actually shipped. If a figure is not in those files, it does not go on the page. No rounding up, no "most businesses", no benchmark half-remembered from a webinar.
What a traceable number looks like
The strongest figure we have is not a lift. It is 36,786.
That is how many URLs the old Shopflooring store served before we replatformed it, every one of them mapped to a live destination before the DNS cutover so that no ranked URL became a 404. Over 23,000 of them redirect straight to the same product. The rest go to a search or a brand landing page that was checked to answer the query, and a 300-URL sample came back 300 for 300 against production.
Set that next to "migrations that protect your search rankings", which is the same claim with the evidence taken out. The count can be wrong. The method can be argued with. You could pull an old sitemap out of an archive, pick twenty URLs at random and see where they land today. That is the point of publishing it: a number worth putting on a website is one a sceptical reader could go and break.
The rest of the site works the same way. Asisly answers in 33 languages, and the seven the primary transcriber does not support are routed to a second provider rather than pretended over — the kind of detail that only survives when the number has to hold up. Contractor OS keeps no money in floating point anywhere: numeric(14,2) in Postgres, decimal strings in JavaScript, every calculation in integer cents, and a demo balance sheet that comes out out by exactly 0.00 on two different Postgres implementations. Toolkit performs 48 operations, reads 12 OCR languages, and deletes everything an hour after processing.
Not one of those is a business outcome. That is the trade.
| The sentence we could write | The sentence we do write |
|---|---|
| Migrations that protect your rankings. | 36,786 legacy URLs mapped before the cutover, none left to 404. |
| Enterprise-grade reliability. | 643 backend tests, 70 versioned migrations and 13 scheduled jobs on a quarter-hour cadence. |
| Speaks your customers' language. | 33 languages answered natively, seven of them routed to a second transcription provider because the first cannot do them. |
| Job costing you can trust. | Zero floating-point money values, and a demo balance sheet out by exactly 0.00 on two Postgres implementations. |
The left column is more persuasive. We know that. It is also the column where a prospect has nothing to take hold of, and where we would have nothing to say in the second meeting except the same thing more loudly.
What the rule costs
Three things, and they are worth naming rather than pretending the discipline is free.
The headline is weaker. A count of artefacts does not sell like an outcome percentage. The proof strip on our homepage is a row of counts — products in a catalogue, workflows in an estate, subscribers on infrastructure we own — where a competitor's is a row of percentages. Ours reads flatter. It is also the only version we can hand you a way to check.
There are pages we cannot build. No testimonial wall, because we are not writing testimonials. No results page with cohort averages, because we do not have a cohort we could describe honestly. Of the fourteen systems on this site, three link to a public URL you can open; the rest say "Private", because a machine-generated hosting URL is not a public reference and we are not going to dress one up as something you can go and inspect.
Most of what we can prove is unglamorous. "2,257 supplier photos matched on exact SKU" is not a headline. It is a Tuesday. But it is the actual texture of building on a deep catalogue, and anyone who has tried it recognises the problem on sight — which turns out to be a better filter for the clients we want than a percentage would be.
WHERE THE DISCIPLINE PAYS FOR ITSELF
In the second meeting. A claim with a method attached gets more specific under questioning. A claim without one gets quieter.
"We do not know yet" beats a plausible percentage
The calculator on our engagement page does arithmetic on numbers you type in, and shows the sum it used: hours a week, times people, times loaded hourly cost, times 52 weeks. That is the whole model. There is deliberately no efficiency multiplier on the end of it.
We could add one. Pick a share of the process that a system "typically" removes, apply it, and the output stops being a cost and becomes a business case. It would look reasonable. It would also be invented, and you would have no way to tell which half of the result was your arithmetic and which half was our assumption. So the page says what it is doing and then stops: it does not know your margins, your seasonality, or whether the process can be automated at all.
The hourly-cost field carries the same warning in its help text. The default is a placeholder for salary plus payroll burden, benefits and overhead — an assumption, not a statistic, and you are meant to replace it with your own figure.
Then there is the intake. The brief on our contact page ends in an assessment that is allowed to come back as "possibly too early": low volume, a small budget and not much manual handling rarely means a build that pays for itself, and the useful thing is to say so before anyone has spent anything. A qualifier that never disqualifies is a lead form wearing a costume.
How to check anything we have written
- Open the three that are public. Search the catalogue, price a room, start a vendor onboarding, ring the number. Count things. The claims are structured as counts precisely so that you can.
- Ask how a figure was produced. Every metric sits on a case study that sets out the work behind it. If the method does not survive a question, the number should not have been published in the first place.
- Take the audit. It is free, it is written, it arrives inside a week, and it is yours to keep whether or not you hire us: findings ranked by estimated revenue impact, a scope and cost range for the work worth doing, and a recommended first slice small enough to judge us on. If the honest reading is that the work is not worth doing at your volume, the dossier says that instead.
- Notice what we will not give you. Ask for our average client return and you will get an explanation of why we do not publish one, rather than a figure invented on the spot. That is a slower answer. It is also the only one that would still be true a year from now.
WHAT IS DELIBERATELY ABSENT
No money-back guarantee, no ROI guarantee, no ranking guarantee, and no aggregate rating in our structured data to win a star in search results. All of those were available. None of them would have been earned.
The build as the last editor
The proof strip pulls each of its metrics out of the case study file by slug when the site compiles, and throws if one is missing. Rename a case study and the build fails rather than quietly serving a claim whose source has gone. The system matcher behind the brief behaves the same way: a renamed slug is a build error, not a dead link a visitor finds six months later.
That is a small piece of engineering and it is not there for elegance. It is there because good intentions decay. A rule that lives in a style guide gets broken in a hurry by someone with a deadline. A rule that fails the build does not get broken at all, which is the most durable form of editorial policy we have found.
The same logic runs through the rest of it. Anyone can promise to be honest about numbers. The question worth asking a studio is what they have built that would stop them being otherwise — and then whether you can go and check that too.
If you would like to see this applied to your own business rather than to ours, that is what the audit is for. It is free, it is written down, and it is allowed to tell you that you are too early.