01A Site Made of Filings
In September 2026 we built an index of China’s medical device makers from a regulator’s public records, with no word on it written to rank and no picture, and kept a record of what search engines did with it.
- The answer sat in a join, since of the 5,100 firms in China that hold a US registration for medical devices, only 1,171 file listings of their own, 1,989 are on file as contract manufacturers who make for others, and not one of the filings gives a website.
- Google showed the pages four hours after it was told of them, and in the first ten days, from 22 September to 1 October 2026, it sent 1,589 clicks from 75,478 impressions at an average position of 9.9.
- The clicks spread thin, as 93 % of them went to pages that each hold one listing, no page drew more than 7 in thirteen days, and 1,312 of the 1,489 pages that were clicked at all were clicked just once.
- Its readers searched from desks and from abroad, since 81 % of the clicks came from a desktop, the first weekend ran at about a third of the weekdays, and two clicks in three came from outside the United States, where all of the records are filed.
- The sourcing searches brought no one, as searches with words such as manufacturer, supplier or China in them showed the site 688 times by 4 October without a click, and Google named the search behind only 40 of 1,701 clicks.
- Three faults came up in the first two days, the product markup, a licensed text on all 340,248 record pages and 1,689 distinct opening paragraphs, and we fixed each of them within hours, yet by 7 October Googlebot had read only 5 of the 33 pages we inspect every day since the fixes.
- On 3 October Google withdrew almost all of it, as from 01:00 Pacific time it showed the site about 13 times an hour against about 400 on the weekdays before, the weekend brought 4 and 5 clicks against 54 and 83 a week earlier, and the records hold no reason for it.
Our Research Question
Allomedo went up on 21 September 2026 as an index of China’s medical device manufacturing, read from what the makers themselves have filed with the United States Food and Drug Administration, and it has 340,248 pages that each hold one record.
We built it to test one claim. So this paper asks whether a site that is made only from a regulator’s public filings, with no text written to rank and no pictures, can be found by the buyers those filings could serve; what it took to build it; and what its own record shows about where search took it and where search stopped. Two days after launch we read the data as proof, perhaps too soon, that the structure was working.
The record now runs to 6 October, and it says more than that, and some of what it says is less kind. Every figure below comes with its source and the date on which it was read, and the last part, Method and Sources, says how each record was pulled and checked, so that another firm could repeat the work on its own.
In its first 10 days Allomedo was clicked from 96 countries, and more than 9 clicks in 10 landed on pages holding one record.
Why We Can Tell All of It
Most of what we build belongs to a client, and a client’s figures stay with the client. Allomedo is ours. We hold the domain, the data, the build and the accounts, so we can show every number and every mistake in it, and the dates on which we made them.
It runs on our own engine, the one that our clients’ sites run on, so every tool that was used here is one that a client’s site already has.

What a Firm Already Holds
Any business that makes things keeps records of them, from product sheets and specifications to certificates and the list of what is made where. In most firms that knowledge sits in spreadsheets and PDFs that search cannot read, and a buyer seldom finds it.
A regulator’s register is the same kind of knowledge, kept at the scale of a whole trade. So we took one, and we asked what it would take to make it readable to the people who buy.
02The Regulator’s Open Files
Before we began, every fact on Allomedo was already public, sent to the FDA by the makers themselves and published in sets of data that answer a buyer’s question only when they are read together.
Four Sets of Open Data
Makers of medical devices have to file what they sell in the United States, and the agency makes what they file free for anyone to download, under terms that let anyone use it again. Four of those sets matter here: the devices that are listed for sale, the establishments that are registered to make them, the clearances the agency has granted, and the recalls.
We downloaded all of it on 20 September 2026, which came to 59 compressed files from seven archives and about 2.5 GB, and we counted every record in 187 seconds. The device catalogue held 5,182,695 entries, and if you group them the way a buyer thinks of them, by company, brand and product code, they come to 390,420 families of devices from 11,874 makers, and these sit under 7,093 product codes in 21 medical specialties. In the registration file we found 335,224 records of establishments.
There were 176,070 clearances, and there were 59,222 records of recalls against 2,843 kinds of device and 39,949 more of enforcement.
Nothing in those files is secret. Anyone can look a record up on the agency’s own site, and other sites copy them entry by entry.
Who Else Makes This Kind?
Someone holds a device from a Western brand, and they want to know who in China is registered to make the same thing.
The regulator already groups the devices of one kind under a product code, a class of three letters such as HMP for eye pads. A clearance rests on showing that a device is substantially equivalent to one that is already on the market, usually under the same code. So the answer to that buyer is in the filings.
But it is kept in three places that do not meet.
It is the catalogue that says what a firm sells under its own brand, and it is the clearances that show which devices have passed review. Only the registrations say where a firm is and what it is on file to make, and only they carry its country. Eye pads show how wide that gap can be, since there are 170 firms in the world that are registered to make them, and 100 of those are in China, while only 20 firms anywhere file a product listing for them, and just 1 firm in China does both.
An index built from listings finds 1 maker in China for eye pads. The registrations name 100.

Factories the Catalogue Misses
That join also shows why nobody had answered that buyer before. Allomedo gives a page to 5,100 firms in China that hold an American registration to make or export devices, and only 1,171 of them file device listings of their own.
Most of the others make for someone else. 1,989 of the 5,100 are on file as contract manufacturers, who build to another firm’s design and whose work is sold under that firm’s name. Not one of the 5,100 registrations gives a website.
To anyone who searches the catalogue, most of these firms might as well not exist.
A firm files one registration for each kind of work it does, and so the registrations always run far ahead of the firms. When we counted them on 20 September, establishments in China held 29,322 registrations to manufacture, 14,576 to export only and 12,007 as contract manufacturers, and there were smaller counts for repacking, for sterilising for others and for writing specifications. Clearances tell the same story from the other side, as firms in China held 5,669 of them, which were spread over 2,231 firms and 440 kinds of device, and their registrations cover 2,580 kinds.
The gap lies in devices that are exempt or of the lowest class, and those need no clearance at all.

The numbers
| Registrations in China, by type | Registrations |
|---|---|
| Manufacture | 29,322 |
| Export only | 14,576 |
| Contract manufacture | 12,007 |
| Repack or relabel | 1,620 |
| Contract sterilisation | 1,266 |
| Develop specifications | 1,206 |
| Private-label distribution | 678 |
| Complaint files | 336 |
| Firms in China with a page (22 Sep 2026) | 5,100 |
| Of them, filing listings of their own | 1,171 |
| Of them, on file as contract manufacturers | 1,989 |
| Of them, with a website in the filing | 0 |
Where Buyers Look Today
We read the alternatives on their own pages on 20 and 23 September. None of the free sites that search engines can read held the join, and each of them held a part of the answer.
When we searched for FDA registered syringe manufacturers in China on 23 September, page one held lists of the top ten factories that were written by factories, and it held the factories’ own sites.
On 23 September, page one for FDA registered syringe manufacturers in China held no neutral source at all.
| Where a buyer looks today | What it holds | What it cannot answer |
|---|---|---|
| The regulator’s own database | All of the official records, with a search box | Who else makes a device, since it has no view from a device to its makers and keeps the country in another file |
| Sites that copy the records | A page for each record, which search has indexed | The same, because each of their pages copies one record and joins it to nothing |
| Business marketplaces | Suppliers that a buyer can contact | Who is registered to make exactly this, as each supplier declares its own categories |
| Paid regulatory databases | The join, in some form | Anything for free, since they sit behind a login that search cannot pass |
03Building the Join
The work ran from a first count on 20 September to a sitemap on 22 September, and most of it went on names, those of the firms, of the devices and of the pages themselves.
Count Firms, Never Records
One maker in China can file dozens of registrations, one for each type of device and each kind of work that it does. If you count rows, a supplier base looks up to thirty times its real size; for ultrasound there were 1,302 registration records, and only 46 firms behind them.
We made that mistake twice in one session on 20 September, once in conversation and once in the headline of our own note on whether the site was worth building at all. An early figure of 42,172 establishments in China was a count of registration records, and the number of firms was 4,991. Since then the script that counts the files prints both of the columns side by side, so that the smaller number is always in view when anyone writes about suppliers.
On the same day one part of the downloads arrived cut short, and a file that goes missing will quietly shrink every count after it, so the script now tests every archive before it reads it and skips a broken one with a loud warning. And we never edit a script while a run is using it; the run works from a dated copy, so that a number in a report can always be traced to the code that made it.
Joining on Company Names
What the three files have in common is how each firm is called, and the filings spell that in many ways, in capitals, with and without the legal form, with a city glued on or a division added. We normalised every name, joined the files on them, and then grouped the result by product code. It came to 205,359 rows, one for each firm and each type of device it is registered for, and all of it was worked out before any page is drawn, so that no page has to compute anything when a reader loads it.
Names arrived in the agency’s casing. On 21 September, 23.5 % of titles were in capitals. Another 11.8 % mixed words in capitals with the rest.
One pass changed 120,564 titles in 22 seconds.
Places were harder, because for China the registrations leave the state field empty and write the city and the province together as free text, such as Weifang Shandong, SUZHOU CITY Jiangsu or a bare BEIJING, so a parser reads that text against a list of provinces. On 21 September it still left 1,141 firms without one, and the firms that have a province sit in 31 of them.
Two plants that shared one establishment number, which was an error in the agency’s own data, had also collapsed into one firm. Now the identifier of a firm keeps its name beside the number wherever a number is shared.

One Record, One Page
On a site that is built from data, one decision comes first, and it is the unit that each address will hold. The links, the lists, the sitemap and the structured data all follow from it.
We chose the device listing, one firm’s brand of one kind of device. It is the unit that a buyer arrives with, holding a box or a catalogue number, and it is also the unit that the filings support. On 20 September, 346,445 of the 390,373 listings sat in a product code with at least one registered maker in China, and the median one had 17 of them.
The other 43,928 had no maker in China to show, so their pages could not answer the question they exist for, and we did not make them. One of them was the code that a first test happened to open, for reprocessed electrosurgical accessories, which has no registered maker in China at all.
The title of a page is the name of the device and nothing else. Its address is made of the product code and the brand, with the maker added only where the brand alone is ambiguous, since different firms often file the same string under one code; there were 55 companies that called their device Orthopedic Manual Surgical Instrument. Before that rule, 9,452 device pages shared an address with another page, and only one of each could be reached.
After it there were 389,989 distinct addresses, 95.7 % of them in the short form, and rebuilds and a sweep of unpublished records on 23 September left 340,248 record pages.
Each of those addresses is meant to last, since the build knows a device by its code, its maker and its brand, and a firm by its establishment number, and a refresh compares each new download with the last one and applies what was added, retired or changed without renumbering anything. On 22 September that refresh existed as a script, and it did not yet run on a schedule.

The numbers
| Device types | Count |
|---|---|
| No firm in China | 2,972 |
| 1 to 19 firms | 2,156 |
| 20 to 49 firms | 259 |
| 50 firms or more | 165 |
| Listings in a type with a maker in China | 346,445 |
| Makers on the median listing page | 17 |
| Listings left out | 43,928 |
Links the Data Draws
A crawler reaches a page by following something that leads to it. On an index of this size no person could write those paths by hand, so on Allomedo every link comes from the data.
The filings already carry a hierarchy, and the site follows it. There are 17 groups, the agency’s own top-level categories, and they lead to 78 sections, each of them a subpart of the federal device rules with a maker in China registered under it. The sections lead to the 2,581 device types that have such a maker, and each type lists its record pages and the firms that are on file to make it.
A firm’s page lists everything else it is on record as making, each record’s table of makers links to the firms, and the firms sit in their 31 provinces, so a crawler that arrives anywhere can walk the whole of it and meet a fact at every step.
Our engine can also add links of its own, chosen by relevance and set into the text. On 340,248 generated records that would be the same element repeated hundreds of thousands of times, the kind of pattern that search engines learn to discount, so the module is off, and every link is a relation the filings already hold.
We checked ten pages of every type on the local copy on 23 September. Every one of them answered and carried valid structured data, and of 1,500 internal links none was broken. That check also found 17 group pages of about 70 words that all shared one description, and 4,506 empty device types that nothing linked to.

04Inside Each Listing
A buyer who arrives from a search lands on one device, and the rules for what that page may say came from reading our first drafts the way such a buyer would read them.
Rows of Registered Makers
Each record is built around one table, of the plants in China that are on file for that kind of device. Every row gives a firm’s city and its province, the role it is registered for and how many device types it is on file to make, with a link to the firm’s own page.
Around that table the page sets out what the filings say about the item itself, from its name, the firm that lists it, its product code, class and regulation to its sizes, its catalogue numbers, its safety flags and the clearances and recalls that touch its code. The question that the reader came with comes first, straight under the title and its count, before the record’s own facts.
A listing of eye pads, 25 to a box, shows how this looks. Its page counts 100 manufacturers in China registered for its code, most of them in Zhejiang with 31, Jiangsu with 26 and Guangdong with 13, and it says that 12 of them are registered as contract manufacturers.
Filings vary in what they hold. Four in five carry a description. One in five of them lists its sizes.
A page that has fewer facts says fewer things, and nothing on it is ever padded so that it looks complete.

Rules From the Early Drafts
Reading the first versions as a buyer would gave us three changes, and each of them now holds across the whole index.
The device or the factory is always the subject. Our first introductions opened with where the data came from, and read that way they made the regulator the subject and Allomedo its mirror. To a buyer that looked like a copy with a disclaimer bolted on, and it buried the one thing they had come for. Every page now always opens on the device or the factory and names the source once, in a fixed line near its foot.
A model never writes the records. Our builders write the page for each listing, each factory and each type of device entirely from the filings themselves, field by field, through templates. Up to 28 September, the log of every change that reached the live site lists one batch of text that a model wrote, six editorial pages on 22 September (About, Contact, a page for manufacturers and the three legal pages). On the next day we switched off four of the engine’s generator modules for this site, among them a term processor and an image namer.
There are no pictures on the site. A registration has no photograph, so the gallery and the featured image are off, and the design is type in the site’s own colours. If the pages were found, their looks had no part in it.
The engine the site runs on can write text with AI, and on Allomedo its writing modules are switched off.
Words on Each Kind of Page
That check of 23 September also counted the text of ten samples of every type. Record pages are data pages, with about 500 words and less than 160 of them prose, and the prose sits on the pages for device types and sections, where it is written once for a whole class.
We chose to leave the records without prose of their own, because when we measured it that day a paragraph written from the same five fields on every record came out in 1,517 forms, and each of them was repeated 62 times, which is the very pattern that search engines are built to find. So each record keeps the maker’s own words, the facts in its tables and a link to its type’s page.
| Page type | Words, median | Of which prose |
|---|---|---|
| Device type with records | 1,704 | 1,122 |
| Section | 2,192 | 1,493 |
| Supplier | 1,341 | 89 |
| Record, the maker’s own words under the title | 529 | 88 |
| Record, a long description in the body | 505 | 159 |
| Province | 613 | 38 |
| Group | 70 | 11 |
| Device type with no records, old template | 42 | 38 |
Structured Data for the Machines
Every page also states its facts in code that tells a machine what the page is about. A record now declares a medical device with its product code as a medical code, the firm behind the listing as an organisation, its place in the site as a trail of links, and a question that it answers. The questions live only there, since on the page itself the blocks carry short labels, and in the code the same facts are put the way a buyer would ask them, such as whether a company is still registered with the FDA.
It is all in the sitemap, and when Search Console last read it, on 29 September, it held 347,825 addresses with no errors and no warnings, and on 7 October the live record for eye pads declared its device, its code, its firm and its question in that form.
05Found Within Hours
At four in the morning, Pacific time, on 22 September, the sitemap reached Search Console, and by eight Google was showing the pages.
From Sitemap to Search Results
That first day brought 14 clicks from 304 impressions at an average position of 7.2. The domain was entirely new, registered that month, and nothing stood behind it, with no advertising and no campaign.
In the first ten days, up to 1 October, Allomedo drew 1,589 clicks from 75,478 impressions at an average position of 9.9, as our tracker read the account on 7 October. Its position on each day ran from 7.2 on the first day to 11.6 on the first Sunday. The best day was Monday 28 September, with 243 clicks.
Those totals move a little with each reading, because Google fills in its last few days afterwards. When we read the same ten days on 3 October, there were 1,579 clicks and 75,189 impressions in them, and four days later there were ten more.
On its first day Google showed Allomedo for 24 searches it would name, on its second for 410 and on its eighth for 1,089.

The numbers
| Day | Clicks | Impressions |
|---|---|---|
| 22 Sep 2026 | 14 | 304 |
| 23 Sep 2026 | 145 | 5,307 |
| 24 Sep 2026 | 232 | 10,421 |
| 25 Sep 2026 | 193 | 9,411 |
| 26 Sep 2026 | 54 | 4,820 |
| 27 Sep 2026 | 83 | 5,702 |
| 28 Sep 2026 | 243 | 10,793 |
| 29 Sep 2026 | 219 | 11,239 |
| 30 Sep 2026 | 222 | 9,656 |
| 1 Oct 2026 | 184 | 7,825 |
| 2 Oct 2026 | 103 | 6,524 |
| 3 Oct 2026 | 4 | 344 |
| 4 Oct 2026 | 5 | 369 |
| 5 Oct 2026 | 33 | 778 |
| 6 Oct 2026 | 14 | 314 |
Thousands of Quiet Listings
Each click went to a single device listing that almost nobody had asked for. From 22 September to 4 October Google showed 26,565 different pages at least once, and it sent a click to 1,489 of them.
No single page carried the site. The busiest one drew 7 clicks in thirteen days. There were 1,312 pages that drew a single click.
Of all the clicks, 93 % landed on record pages, 1,604 of 1,718, against 56 on the pages for device types and 53 on firms.
And the pages turned over very fast, since of the pages that Google showed on any one day from 23 September to 2 October, only 16 to 24 % showed again on the next day, which is how a long tail of part numbers looks, and about one in seven of the searches Google named had a digit in it, a catalogue number or a model. By our reading this is the record page doing what it was built for, as probably the cheapest certain answer that an engine can find to a narrow question, one device and its makers, light to fetch and plain to read, and an index of filings that works looks like this, with many quiet pages that each answer one thing.

The numbers
| Measure | Count |
|---|---|
| Record pages, clicks | 1,604 |
| Device-type pages, clicks | 56 |
| Supplier pages, clicks | 53 |
| Other pages, clicks | 5 |
| Record pages, pages shown | 24,682 |
| Device-type pages, pages shown | 1,079 |
| Supplier pages, pages shown | 725 |
| Other pages, pages shown | 79 |
| Pages that drew 1 click | 1,312 |
| Pages that drew 2 clicks | 147 |
| Pages that drew 3 clicks | 17 |
| Pages that drew 4 to 7 clicks | 13 |
Buyers at Work and Abroad
Over its first weekend the site drew 54 clicks and then 83, about a third of the 205 a day it had averaged over the seven weekdays from 23 September. Monday answered with 243.
That fall began on Friday 25 September at about 13:00 Pacific time, when the impressions an hour halved, and it was the desktops that carried it. The clicks from desktops went from 160 on that Friday to 29 on the Saturday, while those from phones went from 33 to 25. Over the ten days, 81 % of all of the clicks came from a desktop, which is 1,294 of 1,589.
They came from a long way off, too, and in those ten days Google showed the site in 200 countries and territories, and it was clicked from 96 of them. The United States, where all of the records are filed, sent 537 clicks, which is a third. India sent 304, nearly one in five, and after it came Germany, Japan and the United Kingdom.
We read that as people looking up a part from an office desk, and quite often from outside the country whose regulator holds the record. The record shows it by the day of the week and by the device, without a survey.

The numbers
| Country | Clicks |
|---|---|
| United States | 537 |
| India | 304 |
| Germany | 59 |
| Japan | 41 |
| United Kingdom | 37 |
| Italy | 36 |
| Spain | 36 |
| Canada | 34 |
| Australia | 28 |
| Mexico | 24 |
| The other 86 countries | 453 |
Searches Behind the Clicks
Google names only some of what people typed to reach a site, and it leaves the rare ones out. For Allomedo, from 22 September to 4 October, it named 4,788 searches, and those carried 17.7 % of the impressions and only 40 of the 1,701 clicks, which is 2.4 %.
Most of them were short, of two or three words, and many of them were a brand, a model or a catalogue number. One asked it outright, who makes this plate.
Some carry the words a buyer uses to source, such as manufacturer, supplier, factory, wholesale, OEM and China, and our tracker watches them. They showed the site once on the first day and between 40 and 103 times a day after it, up to 2 October. By 4 October, 269 of them had shown it 688 times, and not one of them had brought a click.
Those are the searches a sourcing site exists for, and they land on the pages for device types and for firms, and Google had hardly read those; of the 33 pages the tracker inspects, 3 of the 6 device types and none of the 3 firms were in its index on 7 October.

The numbers
| Searches Google named | Count |
|---|---|
| 1 word | 1,122 |
| 2 words | 1,545 |
| 3 words | 1,207 |
| 4 words | 537 |
| 5 words | 212 |
| 6 words | 101 |
| 7 words | 24 |
| 8 words or more | 40 |
| Clicks that came with a named search | 40 |
| All clicks, the same days | 1,701 |
| Sourcing searches, impressions | 688 |
| Sourcing searches, clicks | 0 |
06Mistakes in the First Two Days
Three things went wrong between the sitemap and the end of 23 September, and each of them changed the checks we run before a site goes live.
No Price on a Registry
On 23 September Search Console flagged a critical issue on the record pages. Each of them had declared itself a product, and Google’s rules for a product ask for an offer, a review or a rating, which it put in its own words as “Either offers, review, or aggregateRating should be specified”. A registry holds none of them and must never invent them.
That same day we rewrote the markup so that it describes each page as a medical device, with its product code as a medical code, a type for which Google asks no price. It went into the template at 02:50 Pacific time and into every stored page by 04:19.
But Google reads a page again only when it comes back to it. On 26 September there were 4 records among those the tracker inspects that still carried the error, and on 7 October 3 still did, while the one record Google had read again, on 2 October, passed.
Before launch our checks asked whether the structured data was well formed, and none of them compared it with Google’s rules for a product.
We now check structured data type by type against each engine’s own documentation, and we never put a site live before that.
Licensed Words, Stripped Out
There is no copyright in the agency’s files, with one exception, which sits inside each catalogue entry as an international nomenclature term that the maker states for its device. It has a code, a name and a definition of one paragraph, a median of 584 characters, and it is a separate body that writes and licenses all three.
The agency publishes that term under its own licence. Its open-data site says that anyone who builds a service on it needs a licence of their own. Until 23 September each of the 340,248 records opened on the definition of its device type, under a heading that named the FDA as its author.
Google read those pages in those days, and its snippets quoted the definitions word for word.
On that day we took the field out of every one of the records.
On 23 September, at 05:10 China time, we stripped the field at import and rebuilt the whole site. A record lost a median of 82 words of prose, about 16 %. Now 67.6 % of the records open with the description the maker filed, 4.7 % with the agency’s own definition of the device type, and 27.6 %, which is 94,037 records, with no paragraph at all, straight on their facts.
The pages for device types paid as well, as 1,965 of the 2,581 lost their definition and 616 kept one, from the agency’s own classification. Six descriptions that makers had filed with the licensed wording in them were blanked too. Since then we always read a publisher’s terms field by field before any record is used.

The numbers
| A record opens with | Records |
|---|---|
| The maker’s own description, 67.6% | 230,071 |
| The FDA’s definition of the type, 4.7% | 16,140 |
| No paragraph, the facts first, 27.6% | 94,037 |
| Records opening on licensed text, before 23 Sep | 340,248 |
| Distinct opening paragraphs, before | 1,689 |
Openings Repeated 200 Times
The licensed paragraph caused a second fault. Since it described the type of device, every listing of one type opened on the same words, and the version that Google crawled opened 340,248 pages with only 1,689 distinct paragraphs, each of them on about 200 pages.
Search engines usually group pages like that as duplicates and often keep one of each group. That was our reading on 23 September, and the rebuild of that day left no shared opening anywhere. On the same morning we stopped the line under a record’s title and its block on what the device is from saying the same thing twice.
None of these fixes could reach Google until it came back to read the pages. By 28 September the tracker’s readout said plainly that what Google showed was still the old site.
Bing Kept Its Own Clock
Over the same weeks Microsoft’s engine did very little with Allomedo. It knew only the old http address of the sitemap, and by 1 October its account showed that it had crawled 15 pages and held 2 of them in its index; by 4 October it had shown the site 42 times in thirty days, and nobody had clicked.
Analytics recorded one visit from Bing, on 5 October, and three from AI assistants, from Gemini on 29 and 30 September and from ChatGPT on 5 October. Something else read far more, since on 24 and 25 September outside requests built 352,157 pages into the host’s cache, and the inspection panel showed that it was not Googlebot, though the host’s log that would name the fetcher was out of our reach.
A firm that reads only one engine never sees more than part of its story, as our paper Every Engine Keeps Its Own Record shows over two years of another property.
07After the First Ten Days
Our tracker went on reading the accounts every morning, and on Saturday 3 October the line it draws fell almost to the floor.
A First Weekend, Tested
It had fallen once before, on Friday 25 September, and from 23 September to 13:00 on that Friday Google showed the site about 370 times an hour, and from then to Monday about 222, and the clicks an hour went from about 9 to about 2.
On 28 September the tracker’s readout set down two readings and a test. Either this was a trade site’s first weekend, with buyers away from their desks, or Google had cut the new site’s exposure, which by the readout’s own note new sites often meet after a few days. If the weekdays climbed back past 120 clicks, it was the first; if every weekday stayed under 80, it was the second.
They climbed back. Monday brought 243 clicks. Tuesday brought 219.
Wednesday brought 222, and Thursday 184.
The impressions ran at about 401 an hour, above the first week, so the weekend reading held, and we wrote it down as the site’s rhythm.
From 01:00 on 3 October Google showed the site about 13 times an hour, against about 400 an hour on the weekdays before.
Saturday, 3 October
That Friday, the 2nd, began like the Friday before it, at about 340 impressions an hour, and from early afternoon it eased the way 25 September had, to about 183 an hour. Then, at 01:00 Pacific time on the Saturday, the count fell to about 13 an hour, and it stayed there.
That weekend brought 4 clicks and then 5, against the 54 and 83 of a week earlier. We first read the Saturday on 4 October, at 1 click from 195 impressions, and on the three mornings after that it settled at 4 clicks from 343 or 344.
On 6 October the Monday, 5 October, read as 9 clicks, and on the next morning it was 33, from 778 impressions, so it will probably still rise. So far it runs at about 30 impressions an hour, which is less than a tenth of the Monday before.
A second count says the same. Analytics, which counts only people who touch the page and keeps China’s calendar, recorded 8 and then 4 sessions from search on 3 and 4 October, against 16 and 14 on the first weekend and 52 to 59 from Monday to Thursday between them. Most of the clicks that remained came from India, 34 of 56 from 3 to 6 October, against 12 from the United States.

The numbers
| Day (Pacific) | Impressions | Busiest hour |
|---|---|---|
| 22 Sep 2026 | 304 | 57 |
| 23 Sep 2026 | 5,305 | 377 |
| 24 Sep 2026 | 10,422 | 637 |
| 25 Sep 2026 | 9,414 | 643 |
| 26 Sep 2026 | 4,825 | 265 |
| 27 Sep 2026 | 5,704 | 446 |
| 28 Sep 2026 | 10,523 | 614 |
| 29 Sep 2026 | 11,242 | 640 |
| 30 Sep 2026 | 9,659 | 609 |
| 1 Oct 2026 | 7,825 | 487 |
| 2 Oct 2026 | 6,525 | 399 |
| 3 Oct 2026 | 343 | 88 |
| 4 Oct 2026 | 369 | 41 |
| 5 Oct 2026 | 778 | 62 |
| 6 Oct 2026 | 308 | 53 |
Why Did Google Pull Back?
We have no reason on record for 3 October. The API of Search Console carries no report of manual actions, so the tracker has none to read, and there is no notice in any of the folders that this paper reads, up to 7 October.
What the record does hold sits around that date. Of the 33 pages the tracker inspects, Googlebot had read 10 by 26 September, all of them on 22 and 23 September, before the fixes. By 7 October it had read 11 of them at all and only 5 since 23 September.
Of the 21 that Google had listed as discovered and waiting on 26 September, 15 had become unknown to it by 7 October, and on that same day one indexed record moved to crawled and not indexed. Google had last read the sitemap on 29 September.
There is one more line in our own files, from 20 September, when our evaluation of the plan warned that a new domain which puts up tens of thousands of pages made from data in a month is ignored or removed, and it advised launching three or four departments in depth, earning the index and only then widening, and we put up 347,825 addresses at once. By our reading that is probably the main cause, with the duplicate openings and the licensed text that Google read in its first crawl beside it. None of it is proven, and the days after 4 October have not settled.
The figure most open to change is Monday’s 33 clicks, which rests on two readings and would move if Google kept filling in the day or came back to the site.
A firm that watches its own account through a week like this one gains most by reading it by the hour, by checking a second count, and by waiting for the settled days before it changes anything on the site.

Predictions, Checked
From 26 September the tracker wrote down what it expected, each time with the date on which to look again, so that a number arriving later could prove it wrong. By 7 October the record had scored them as below, with the clicks taken over the settled week of 28 September to 4 October.
| Made on | Prediction | Read on 7 October |
|---|---|---|
| 26 Sep | Clicks settle at 150 to 350 a day; under 150 means the new site’s test ended early | 140 a day |
| 26 Sep | Googlebot comes back to at least 8 of the 33 inspected pages | 5 |
| 26 Sep | Record pages keep at least 85 % of the clicks through October | 93 % |
| 26 Sep | Sourcing searches stay under 10 % of the named impressions and win almost no clicks | 4.4 %, no click |
| 26 Sep | Bing stays near zero until the https sitemap is added | 42 impressions in 30 days |
| 28 Sep | Weekdays reach 120 clicks or more, and Saturday and Sunday, 3 and 4 October, fall to 30 to 80 | 4 of 5 weekdays, then 4 and 5 |
| 28 Sep | No visitor from an AI assistant through October | 3 visits |
08Carrying It Elsewhere
We built Allomedo to test the claim we make to every client, presence by architecture, which holds that a business is found through how its presence is built and that its strongest material is the data it already keeps.
Seven Steps, in Order
Think of a manufacturer’s product sheets, a distributor’s catalogue, a laboratory’s methods or a consultancy’s case records. All of them are knowledge of this kind, and in most firms it sits in files that search cannot see. What we did on Allomedo applies to them, with what the record taught us added in.
- Find the unit and decide what one page is, whether one product, one family, one place or one service, because everything else follows from it.
- Ask what the records answer together, since the value sits in what several of them answer at once, such as who else, what else, where and since when.
- Count the things a buyer counts, firms and products and never rows, and print both numbers where anyone can see them.
- Read every field’s terms before it enters the database, so that each page states only what its publisher lets anyone use again.
- Let the data draw the links, each of them a relation that the files already contain, so the structure is fully complete without anyone writing it.
- Declare it to the machines, with structured data on every page that is checked against each engine’s own rules, and with a clean sitemap.
- Grow at a pace the index can follow and read the accounts every day, with each prediction written down before its numbers arrive.
How Planet Uses It
For a firm that sells abroad, we make a catalogue site from its own product data, with a page for each product in the languages its buyers search in. That is our Planet tier.
Allomedo’s own record shows what it is up against. Every one of the 5,100 firms in China that it lists is registered with a regulator, and none of their filings gives a buyer an address to go to. Registered and not found is the state of most makers in the record.
The look always comes after, and Allomedo shows what a structure can do before anyone has seen a picture of it, and also how quickly an engine can take it back, and the case Public Records, Made Readable covers its first ten days.
What does your firm hold that a buyer searches for, and which of your pages says it?
09Method and Sources
Every figure in this paper comes from the agency’s public files or from what Allomedo’s own accounts hold, and each is given with the date on which it was read.
Counted From the Files
All of the data comes from the bulk downloads of the FDA’s openFDA, seven archives that came as 59 downloads and that hold the device catalogue (UDI, 1.9 GB), the recalls, the 510(k) clearances, the records of enforcement, of registration and of listing, the premarket approvals and the classification. On 20 September 2026 a script tested every archive and then counted every record in 187 seconds, so no count from the files is sampled.
The counts that the build made for itself, of 5,100 firms, of 340,248 record pages, of 2,581 device types and of 31 provinces, and of the 1,171 firms with listings and the 1,989 contract manufacturers, are as they stood on 22 and 23 September. Our check of ten pages of every type on 23 September gave the words on each page, the links and the defects. And the live pages, our screenshots of them at a 1,440 px window and their structured data come from allomedo.com on 7 October.
Read Every Morning
Since 26 September our tracker has pulled what Allomedo’s accounts hold through their APIs each day and saved each pull as it came, in one folder a day. From Search Console it takes the daily totals for 28 days, the pages, searches, countries and devices for the last five days, which is the window Google still revises, and the hours in Pacific time. It also inspects 33 fixed pages of every type and reads the sitemaps, Bing’s account and Analytics.
This paper combines those folders and takes each date from the latest reading up to 7 October 2026 at 04:34 Poland time. Every figure from search is a sum of what Google reported, day by day, with nothing estimated by an outside tool.

Records and Their Windows
Each kind of reading has its own dates, and each figure in this paper carries them.
| Source | What it holds | Window |
|---|---|---|
| openFDA bulk files, seven archives | Devices, registrations and listings, clearances, approvals, recalls, enforcement, classification | Counted 20 Sep 2026 |
| Allomedo’s build | Firms, record pages, device types, provinces, the join | 22 and 23 Sep 2026 |
| Search Console, by day | Clicks, impressions, position, pages, searches, countries, devices | 22 Sep to 6 Oct 2026, read 7 Oct |
| Search Console, by hour | Impressions and clicks, Pacific time | 22 Sep 08:00 to 6 Oct 11:00 |
| Search Console, URL inspection | 33 fixed pages, index state, last crawl, rich results | Every day, 26 Sep to 7 Oct 2026 |
| Search Console, sitemaps | Addresses sent, errors, last read | Read 29 Sep 2026 |
| Bing Webmaster Tools | Clicks, impressions, crawl, index | To 4 Oct 2026 |
| Analytics | Sessions by channel and source | 21 Sep to 6 Oct 2026, China days |
Limits of These Figures
Google keeps the rare searches out of its tables, and in the later readings the list of pages for a day stops near 5,000, so the counts of searches, of pages and of sourcing searches are all lower bounds. Summed page by page the clicks come to 1,718 against 1,701 in the daily totals, since Google counts them per page in one table and per site in the other. The average position is the daily positions weighted by impressions.
Google fills in its last two or three days afterwards, so 5 and 6 October were still moving on 7 October. Analytics counts only a fraction of the clicks, and it does so by design, since its tags wait for the first real scroll, tap or key press of a visitor and then run without cookies until there is consent. Over the seven days from 26 September to 2 October it recorded 283 sessions from search against 1,108 clicks, about a quarter, and the two lines always moved together, which is why we use it as the second count.
In the first days we also read Google’s site: count, the estimate it prints above a search. It said about 1,860 on the morning of 23 September and about 90,000 by the evening of 24 September, so this paper uses it for nothing.
Our reading of the alternatives comes from their own pages, on 20 and 23 September. This paper claims nothing about enquiries or sales, nothing about the answers that Google’s AI gives, which only the browser view of Search Console reports, and nothing about why Google pulled back on 3 October beyond what is in the record.
Nooné, Noo–Studio’s AI writer, wrote this paper from the studio’s own records, and every figure was checked against its source before it was published. Noo–Studio publishes it and answers for it. How the studio uses AI.

