Do not buy the list yet.

The sentence is this: these are the people who should hear from us.

Almost the entire industry treats that sentence as the starting conditions rather than the work. You open a tool. You choose an industry from a dropdown. You choose a company size. You choose a country. You choose a job title. You choose a number. Everything downstream — the enrichment, the sequences, the personalisation, the deliverability engineering, the reporting — is sophisticated machinery built on top of an assumption that took ninety seconds and was never examined again.

Market Scan is the attempt to spend real effort on those ninety seconds.

It is one deep research operation that runs before a single contact is purchased. It produces a defended map of every market the offer can plausibly enter, the evidence and economic reasoning behind each one, the vocabulary each of those markets uses to describe itself, the roles that actually own the problem inside them, and a matrix of precise search instructions a sourcing system can execute.

The output is not a list of people. It is the reason a list of people should exist.

Put ten experienced people in a room. Explain the product. Ask them where it could be sold.

They will produce a sensible list quickly, and it will be a good list. Existing customers. The customers of direct competitors. Two or three neighbouring industries. One genuinely interesting idea somebody remembers from a previous job.

After an hour the whiteboard looks excellent. Everybody leaves satisfied.

That was the easy part, and it is roughly the first ten.

Now ask the same room to prove each direction with current sources. Ask for twenty more defensible markets on top of those. Ask how each of those markets names its own companies. Ask for every plausible title held by the person who owns the problem. Ask them to repeat all of it across the United States, Britain, Germany, the Gulf and Southeast Asia. Ask for the result as a search matrix.

At this point the meeting ends.

Not because the people are not good. Because the work has stopped resembling a meeting.

Every market after the tenth costs more appetite than the one before it.

This gets mistaken for a failure of intelligence. It is not. It is a failure of appetite, and it is completely reasonable.

The first ten directions arrive on their own, because they already exist somewhere in the room's collective experience. The eleventh requires somebody to go and look. The fifteenth requires somebody to go and look, find nothing useful, and go and look again. The twenty-fifth requires somebody to defend an unfamiliar market to four colleagues who would rather go back to the ten that already sound sensible.

Nobody sustains that for long. And there is rarely any pressure to. A team holding ten workable directions can spend an entire year inside them and never once feel a reason to widen the field.

A machine has no appetite to lose. That is its single most useful property here, and it has almost nothing to do with intelligence.

Finding a market is one layer. Underneath it are ten more.

For every direction that survives, somebody still has to answer:

1. Why would this buyer care?
2. How does the problem cost them money?
3. Who owns the consequence when it does?
4. What would make this hypothesis wrong?
5. What evidence supports it today?
6. What does this type of company call itself?
7. What does this role call itself?
8. Which geographical terms change the search?
9. Which records should be excluded, and on what signal?
10. And what, specifically, should this market be told?

Then the combinations begin multiplying.

Take a deliberately conservative illustration.

And that is before company size, operating model, installed technology, legal entity type, local language, subsidiaries, parent groups, or any qualifier separating an operator from an adviser.

The difficulty was never typing one search query. The difficulty is building the complete map without losing the reasoning that produced each part of it. A search path with no reasoning attached is indistinguishable from a guess, and six weeks later there is no way to tell which was which.

Three words do most of the work on this page. They sound more technical than they are, and the distance between the first and the third is the distance between an opinion and an instruction.

Everything begins with what the client can tell us. What the product does. Who already buys it. Where the business can legally and practically operate. What makes the offer worth money. What must be avoided. And what decision we ultimately want another person to make.

That is the seed. It is also the least reliable material in the entire exercise, and most founders reading this already know why. A business describes itself in the vocabulary of the people who built it. The market describes that same business in the vocabulary of the people who pay for it. The two are rarely identical, and only the second one can be searched.

So the scan goes outward. It reads nine kinds of public material, and it reads each of them for a specific reason. Not to collect links. Anyone can collect links. Not to summarise an industry. To answer one question repeatedly: where can this offer matter, and who would feel it if it did.

The pass is finished when the system can explain the market. Not when it has listed the websites inside it.

This is the part where a machine is genuinely unlike a person.

None of those nine sources is exotic. A capable researcher would name the same nine, and would read them better than any model reads them.

They would read perhaps forty sources for a direction, across several days, and would do it thoroughly for the first three or four directions. By the eighth, the reading gets faster and the notes get shorter, because the work is long, the returns feel diminishing, and there is a meeting on Thursday.

The scan reads hundreds of sources per direction, and it reads the twenty-eighth direction exactly as carefully as it read the first. Not better than a good researcher. The same as itself, thirty times in a row, across five geographies.

That consistency is the entire mechanical advantage. It is worth describing plainly rather than dressing it up as intelligence.

Everything read in the first pass is written into a single document. Not a summary. Not a slide. A working record, with sources, on top of which the rest of the scan is built.

The Chronicle is the memory of the operation. Every hypothesis produced later has to point back to something inside it, and anything that cannot point back to something inside it does not survive the Judge.

It records:

Three kinds of statement, kept apart on purpose.

Keeping those three apart is not academic caution. It is the thing that prevents a research system from producing a confident, fluent, beautifully formatted document in which nothing at all can be checked.

That is the characteristic failure of any language model handed a research brief and no obligation to show its work. Ask a model for market research and it will always return market research. The only question that matters afterwards is whether a single line of it can be traced back to something real.

"Sell to logistics companies" is not research. The Chronicle has to state which logistics companies, operating in which way, carrying which problem, at what scale, triggered by what, decided by which role, and on the strength of what evidence.

Everything in this section is a hypothesis until the letters come back. We write hypotheses in the indicative because hedging every sentence makes the weak ones look like the strong ones, which is the opposite of useful. Where a direction is a genuine guess, it is labelled experimental, and that label is doing the hedging for the whole paragraph.

By the end of the first pass there is a long list of directions and a document explaining where each one came from. They are not equal, and treating them as equal is how a scan turns back into noise.

So we place every hypothesis at one of three distances from the offer. The distance is not a quality ranking. A distant market can easily be the better market. It describes how much has to be true for the direction to work, and therefore how much should be spent finding out.

Every hypothesis carries a stated confidence, and the reason for it.

Alongside the distance, each direction carries an explicit confidence and a written justification for it. What supports it. What weakens it. Which single piece of evidence would move it furthest in either direction.

The number matters less than the obligation to produce one. A system required to state how confident it is, and why, cannot quietly set a guess and a certainty in the same typeface. That is the standard failure of research documents that read beautifully and turn out, on inspection, to be one strong finding surrounded by twelve agreeable assumptions.

Some directions share a problem, a buyer and an argument. Splitting them produces twenty-five nearly identical letters, and nobody on earth writes the twenty-fifth one properly.

Other directions look almost identical and are not. Merging those produces one letter that is vaguely true for everybody and compelling to nobody. That is the standard output of the industry, and the reason most cold mail reads as though it were addressed to a category rather than to a company.

So we test every accepted hypothesis against the others on three questions.

1. Is the economic problem the same?
2. Is the person who owns it the same?
3. Does the same evidence persuade them?

Three yes answers and they can share a campaign. A single no and they separate, no matter how similar the industry label looks from the outside.

Northline Industrial, in full.

Northline sells conveyor condition monitoring. The dropdown answer was factories.

But conveyors also run inside parcel hubs, fulfilment centres, food plants, cold-storage facilities, distribution networks and third-party logistics operations. The same equipment can be sold through integrators who never operate a conveyor themselves. And it can be introduced across a dozen businesses at once by a private-equity operations team that directly owns none of them.

This is the point at which a single industry label stops holding together.

A normal lead list puts all five into industrial companies, sorted by headcount.

Five different problems. Four different buyer functions. Five different letters. And a private-equity operating partner who would have been removed entirely by the industry dropdown, because the fund is not an industrial company and has never operated a conveyor in its life.

A warehouse operator, a conveyor integrator and a private-equity operations team should not receive the same letter merely because all three can be made to care about conveyor failures.

They buy for different reasons. They carry different objections. Different people own the decision. Different evidence moves them.

So each accepted vertical receives its own commercial argument, its own sequence, and its own definition of what a good reply would even look like.

What the letter is allowed to claim.

Selling potential barely opens a door any more. We have AI. We have a platform. We have infrastructure. We have the best team in the category. The active market has heard every one of those, fifteen layers deep, and the people who somehow have not heard them are usually either impossible to work with or have no money to spend.

What still works is naming a specific economic consequence inside a process the recipient already lives with every day. Not: we noticed you have a challenge in this area.

But: this is where the money leaks, this is roughly what it costs, this is who normally owns it, and this is what changes if it stops.

The difference between those two sentences is not tone. It is whether the sender did the research.

Three constructions this system will not produce:

The first announces that a template has been personalised. The second attempts to frighten somebody who is not frightened. The third has been transmitted so many billions of times that it now carries no information at all, and survives only as a reliable marker that the message can be deleted unread.

All three are cheap to write. That is precisely why they are everywhere.

There is now a fourth participant in every deal.

There used to be the buyer. Then the buyer and their team. Then the buyer, their team, and the market around them.

Now there is the buyer, their team, the market, and a model sitting beside each of them that will research your company in ninety seconds and deliver a verdict. Frequently: this looks expensive, this looks unproven, you probably do not need this.

Charisma no longer carries a weak argument through. Neither does a confident deck or a well-designed website. There is something in the room that does not get tired, does not get flattered, and locates the soft part of the pitch immediately.

The only durable answer is to be right about the market before writing anything, and specific enough that the research confirms the letter instead of taking it apart.

Why this part is not handed to people.

This is not a claim that machines write better prose than good writers. They do not.

It is a claim about something else entirely. Ask a talented salesperson for one excellent letter and you will usually get one. Ask for twenty-five, one per vertical, each built on its own evidence, each avoiding the objections specific to that buyer, each holding the same standard from the first to the last — and it does not happen. It has been paid for at every reasonable salary, for months at a time, and it still does not happen.

People think in situations. They are remarkable at reading a room and improvising the argument that lands. That skill does not convert into twenty-five systematic documents produced to an identical standard on a Tuesday afternoon.

Systems are the one thing a machine is genuinely suited to. So the machine drafts every vertical to the same standard, and the human does what humans are actually good at: deciding which arguments are true, which are worth money, and what finally goes into somebody's hands.

By now we know that a market deserves attention, why it deserves attention, and what should be said to it. None of that helps if we cannot locate a single company inside it.

Companies name themselves inconsistently. Industries name themselves inconsistently. Job titles are worse than both. The vocabulary shifts with geography, company size, business model, and whichever rebrand survived the last board meeting.

So the second research pass goes looking for the language.

Leave the conveyors for a moment. Gambling shows this better than any other industry, because it cannot agree what it is called even in its own filings.

One market, eleven names.

Suppose the target is gambling. The companies describe themselves as:

And in several jurisdictions, a local regulatory category nobody outside that jurisdiction has ever said out loud.

Search only for gambling companies and the database returns exactly the narrow market your narrow vocabulary requested. It will return it quickly, cleanly, and with no indication whatsoever that anything is missing.

That is the dangerous part. An incomplete search does not look incomplete.

The same problem, worse, with people.

A partnerships decision inside that market belongs to:

Head of VIP is the one worth stopping on. In parts of that industry it denotes the person who personally owns the most commercially significant relationships in the business. Nothing in the words tells you that. Nobody would type it into a filter. And an outsider searching for founders and C-level titles in the conventional sense will cleanly miss the person who actually decides.

Every industry contains a dozen of these. It is exactly why the market pays a premium for people who already learned one industry's vocabulary at a previous employer's expense, and why buying that knowledge through hiring is slow, narrow and expensive.

The useful property of AI here is not that it holds one magic list of job titles. It is that it will repeat this research for every accepted market, in every relevant geography, without deciding somewhere around the fourth segment that three titles are probably enough.

Then it becomes a matrix.

Each accepted vertical is crossed with every company type inside it, every naming pattern for that company type, every buyer function, every title variant, every relevant geography, every meaningful scale band, and whatever qualifier keeps the search honest.

The result is thousands of Market Arms. On a large scan, tens of thousands.

That number is not the system losing control. It is what the market looks like when nobody has quietly rounded it down to the first four ideas that came to mind.

The purpose is not to buy every record from every arm. Most arms will never be purchased at all.

The purpose is to see the entire searchable market before a human, a database or a budget quietly reduces it to whatever was thinkable on the first afternoon.

Everything described so far runs as four separated passes. They are separated because each demands a different kind of attention, and a system attempting all four at once performs all four adequately.

The Judge was the part that took longest to get right.

Software can validate a file. It can confirm that the JSON parses, the arithmetic adds up, the required fields are present and nothing is malformed. All of that is useful, and all of it is nearly worthless here, because none of it examines whether the reasoning is any good.

Meaning is not a file format.

So the Judge asks the questions a validator cannot. Does this hypothesis actually follow from what the Chronicle contains? Does the cited evidence say what it is claimed to say? Do the markets that were merged genuinely share a buyer? Did the markets that were separated genuinely need to be? Does this letter address the buyer it was written for, or a nearby buyer who was easier to write to? Are these Market Arms useful, or merely numerous?

When one part fails, that part goes back. Not the whole scan. A weak letter for one vertical does not invalidate the other twenty-four, and rebuilding everything because one segment came out thin is how a system becomes too expensive to run at all.

What this looked like before it worked.

The current architecture is four sequential passes, each running up to roughly fifty minutes.

An earlier version of the same scan used between two hundred and ninety and five hundred sequential model sessions for a single product. The longest complete run took two and a half days.

It produced good work. It was also unshippable, in the way that anything requiring two and a half days and continuous supervision is unshippable.

Getting from five hundred sessions down to four was not a matter of deleting steps. It was a matter of writing down what each pass is actually responsible for, in enough detail that it stops needing to be told twice. Most of the work of building this system was that, and almost none of it is visible in the output.

A scan is not good because the number came out large.

Discovery should be broad. Spending should be suspicious. Those are different instincts and they belong to different stages.

Before a single contact is purchased or a single letter produced, the following are removed:

The scan widens the world first, then removes everything unable to defend itself. Both halves are required. Breadth without pruning is just an expensive list. Pruning without breadth is the dropdown menu again, wearing a longer document.

Every accepted Market Arm becomes a sourcing instruction, and candidate companies come back against it. A record returned by a database is a candidate. It is not yet a company worth writing to.

The company.

We check each candidate against the hypothesis that produced it:

1. Does this company actually perform the target activity?
2. Does it operate in the required geography?
3. Does its scale fit the argument?
4. Is the relevant facility or business unit real?
5. Does it show the operational signal the hypothesis requires?
6. Is it still trading?
7. Is it a duplicate of a record already accepted under another name?
8. Is it part of a larger group, and does that change who decides?
9. Is it an operator, a vendor, an adviser, or a directory entry?

The company file records why the organisation was accepted or rejected, and which Market Arm produced it.

That last field matters more than it sounds. It is what allows a weak arm to be identified months later by its results rather than by somebody's opinion of it in a meeting.

The person.

We do not send letters to markets. We send them to people, and not necessarily to the most senior person available.

The target is whoever sits closest to the decision. Who owns the problem. Who feels the cost when it happens. Who controls the relevant budget line. Who can approve a next step. Or who can carry the proposal to the person who will.

A title is evidence. It is not proof. Profiles go stale. Conference biographies describe jobs people have already left. A group executive can hold complete authority and no interest whatsoever in one site's problem. A site manager can feel that problem every single day and have no permission to change anything across the group. Where those signals conflict, the person stays unresolved until a current source closes it.

The record is strong when several signals agree at once: the person is current, the role is relevant, the geography makes sense, the operating scope matches the hypothesis, and the company context supports the argument the letter is going to make.

The physical route.

Here physical mail stops resembling email, and being wrong starts costing money.

A correct name and a headquarters address are not a delivery route.

The registered address often belongs to a formation agent. Headquarters can sit in a different country from the recipient. The person sometimes works from a regional office, a plant or a subsidiary trading under a different name. Companies move. People leave. A shared reception accepts a package and has no reason on earth to route it onward. Parent companies refuse mail addressed to employees of their own subsidiaries.

So we verify the route as one connected record: current company, current person, current role, the relevant receiving office, the physical mailing address, the sources supporting each of those, and anything that contradicts them.

A weak route is not tested with the client's postage.

That single sentence is most of the difference between physical outreach and email. A bad email address costs nothing, so it gets sent anyway, and the failure is invisible. Paper, an object, production, handling and postage all cost money, and every wasted send is a small physical monument to a decision that should have been made three steps earlier.

The constraint is not a disadvantage. It is the entire reason this system runs in the order it does.

By this point we know why this market matters, why this company fits it, why this person was chosen over the three others in the same building, what economic problem the letter is entitled to discuss, what evidence exists behind it, and where an envelope can physically arrive.

Only now can a letter be specific in a way the recipient is able to verify.

Each vertical carries its own argument. Each recipient gets one clear reason for the approach. Each letter asks for exactly one next step.

When an object goes into the envelope it has to carry part of the argument, not decorate it. A worn part, a material sample, a coin, a map, a card that makes replying trivial. Which object, and why that one, is the other page's job.

This is where most of this work normally goes to die.

Research lives in a document. Company sourcing lives in a spreadsheet. The people live in a second spreadsheet. Address verification happens in messages. The copy lives in a separate file. Production happens in a supplier portal. Tracking happens in a carrier account. Replies arrive wherever replies feel like arriving.

Four months later somebody asks why one particular company was ever contacted, and the answer has to be reconstructed from six tools and the memory of a person currently on holiday.

coin.im is built around that chain as one connected record.

Open a company and see why it entered the campaign. Open a person and see why they were selected over the three others in the same building. Open an address and see what evidence approved it. Open a letter and see which market logic produced it. Open a delivery event and trace it back to the original hypothesis and the sources behind it.

The CRM is not an address book bolted on at the end. It is the memory of how the campaign was built.

And then the results come back and change the map.

A scan built once is not a scan frozen once.

Replies, silences, deliveries, refusals and returned envelopes all attach to the arm that produced them. Across months this does something a purchased list is structurally incapable of doing.

Weak markets shrink. Strong markets get more precise. Title patterns appear that nobody predicted. Old addresses die and are marked as dead rather than quietly resent next quarter. A hypothesis rated at low confidence turns out to answer at three times the rate of an obvious one, and the reason why gets written down instead of being remembered vaguely by one person.

The business stops repeatedly buying lists and starts accumulating knowledge of its own market.

What actually arrives.

This is not a general benchmark and it is not an opinion about which model is cleverest. It measures one thing: how each model performs the actual passes of an actual Market Scan.

In August 2026 the complete scan — all four passes — was run against three real business inputs, on five models, repeatedly, and scored on what came out.

The clients are anonymised here. The numbers are not adjusted.

The headline is in the gap between two columns.

On segments, the models are close. Forty-eight against seventy-three. Every one of them finds roughly the same set of markets, which is reassuring and also unsurprising: the obvious directions are obvious to anything that reads.

On arms, they are not close at all. Six hundred and sixty-one against four thousand four hundred and seven.

Same markets. Nearly seven times the searchable surface inside them.

That difference is not cleverness. It is persistence — the willingness to keep asking how else this type of company might name itself, in this geography, at this scale, after already producing eleven answers that looked sufficient. It costs about twenty-five extra minutes of reasoning per run.

For a scan built once to plan a year of campaigns, twenty-five minutes is not a consideration. That is why the breadth model runs the research and the arms, a faster and more controlled model handles the copy, and a separate model sits as Judge over both.

What this measurement does not prove.

Stated plainly, because a benchmark that only reports its wins is marketing:

A wider searchable market is a wider searchable market. It is not revenue, and anybody presenting it as revenue is selling something.

What it does mean is that the map is not being quietly truncated by the tool drawing it.

Where the doctrine came from.

The rules the scan follows were not derived in one sitting. They were written down across outbound work with more than six hundred and eighty businesses, in the ordinary way: something failed, the reason was identified, and a rule was added so it would not fail the same way again.

Most of this page is that accumulation, restated as a system a machine can execute. The individual observations are not remarkable. Having six hundred of them in one place, applied consistently, is the entire product.

Perfect leads do not close deals.

We have delivered hundreds of leads a client could find no fault with whatsoever, and watched none of them convert. Not because the targeting was wrong. Because after the reply there is still the economics, the delivery method, the SLA, the risk, the internal research, the legal review, and the question "do we actually need this" repeating at every floor.

A good outbound operation can turn a good offer into an excellent one. It cannot turn nothing into something. Physics is not negotiable on this point.

The scan finds markets. It does not create demand.

If nobody in a market has the problem, no amount of research produces a buyer. What the scan can do is find the markets where the problem exists and nobody has thought to look, and prove that they exist before money is spent on them.

It does not replace an experienced salesperson.

It allows experienced judgment to operate across a far wider field than one person could ever read. The machine does the repetitive reading, comparison and vocabulary expansion. Humans still decide which evidence matters, which markets deserve money, which people should be approached, and what finally goes into somebody's hands.

Physical mail is slower and costs more per touch.

If the requirement is fifty thousand touches next week at the lowest possible unit cost, this is the wrong system and we will say so. Paper works where the recipient is hard to reach, the deal is worth reaching them, and the inbox has already failed.

The full product is not live yet.

The scan runs. The campaigns run. The self-serve version, where a conversation with a bot produces a complete scan without anybody at our end pressing a button, is still being built. This page describes what exists and what is being finished, and does not blur the boundary between them.

This page describes Market Scan and the full coin.im workflow around it.

The scan runs today. Campaigns run today. The automated self-serve version, the sourcing system, physical-route verification, letter production and the connected CRM are being built as one product rather than assembled from separate tools later.

There is no demonstration dashboard on this page. There is no decorative login opening a sample project. The interface will appear when it has real work to show. Not before.

Until then, the job of this page is narrower: after reading it you should understand what Market Scan does, why doing it properly is difficult, and how it turns one business into months of precise physical outreach.