The Sourcing Funnel, Measured: 4,656 Pages to 568 Suppliers

The four stages (and why the numbers drop at each one)

When you click "Start campaign" in Procurea, four stages run in sequence. The numbers below are not a model of what the funnel should do. They are the sum of the eight campaigns published on this site, every one of them, including the two that went badly.

• Stage 1. Strategy: one brief becomes a set of localised search queries, 25 to 40 per country.
• Stage 2. Screener: each candidate page is fetched and read against the brief. Across the eight runs this stage read 4,656 company pages.
• Stage 3. Enrichment: for what survives, find a published contact address and, for EU suppliers who state a VAT number, run it through VIES.
• Stage 4. Auditor: re-read each survivor against its own pages and score the match, 0 to 100.

1,058 candidates were rejected with a reason recorded. 991 as too weak a match for the brief, 38 in the wrong region, 21 as sample or catalogue pages, and the rest one at a time: a government procurement portal, a trade magazine, a caterer whose specialisation did not fit, and a page about a presidential library. The reasons are in the campaign record, which means they are arguable. That is the difference between a filter you can inspect and a number you have to trust.

568 suppliers were kept, 472 of them scored as primary and 96 held as reserve. 300 came out with a published contact address.

So the honest yield is about 10 percent from page read to supplier kept, and roughly 6 percent from page read to supplier you can actually write to. Anyone quoting you a higher number is either counting differently or has not published the run.

The drop-off is the point rather than a defect. A raw search result list for a sourcing brief is mostly resellers, directories, news and unrelated businesses. A shortlist that has not thrown those away is a search results page with a nicer font.

What the aggregate hides is the spread, and the spread is the useful part. The packaging run read 763 pages and kept 170 suppliers. The kitchen equipment run read 91 and kept 7. Same engine, same week. The difference is the shape of the category, not the performance of the software, and the next section is about how to tell which one you have before you spend the time.

Stage 1, Strategy: from 1 brief to 32 localized queries

The Strategy stage turns the buyer's brief (category, geography, certifications, volume) into a search plan. Not one query, dozens, spread across countries and languages, tuned to how real manufacturers describe themselves online.

For cosmetic packaging sourcing across five EU countries, the plan might look like: - German queries: "Kosmetikflaschen Hersteller GMP," "Kosmetikverpackungen Produzent Deutschland," "PET Kosmetik Flaschen Lieferant" - Polish: "producent butelek kosmetycznych GMP," "opakowania kosmetyczne hurt," "butelki PET kosmetyki" - Italian: "produttore flaconi cosmetici GMP," "fornitore packaging cosmetico," "bottiglie PET cosmetici" - French + Spanish equivalents

25-40 queries per country is the typical range. Going narrower misses suppliers; going broader explodes cost and adds noise.

Why this matters: a buyer doing manual sourcing usually runs 5-8 queries in their own language. That is how "I could not find European suppliers for this category" happens, they exist, they just use different words. Localization is the single biggest lever in discovery. A German "Hersteller" searching a Polish "producent" both mean manufacturer; they produce almost no overlap in Google results.

What Strategy does not do: it does not pick up non-indexed directories (some regional portals block Googlebot), very niche niches under ten global suppliers, or shell-company farms that exist purely to appear in searches. We flag those honestly later rather than pretending they do not exist.

A missing certificate never drops a good maker. It lowers a score, and the score is a number you can argue with.

Stage 2. Screener: 4,656 pages read, 1,058 rejected with a reason

Stage 2 fetches each candidate page, reads the homepage and whatever else looks load bearing (products, certifications, about, contact), and scores it against the brief. This is where the cut happens.

Across the eight published runs the screener recorded a reason on 1,058 rejections:

• 991 as too weak a match for the brief. Real companies, wrong ones: the resellers, the adjacent categories, the manufacturers who make something next to what you asked for.
• 38 in the wrong region for the brief.
• 21 as sample or catalogue pages rather than a company.
• 8 individually, and these are worth reading because they show what the screen is actually doing: a government procurement portal, a trade magazine, a caterer whose specialisation did not fit the brief, and a page about a presidential library that a query for "obama" style branding had dragged in.

A note on arithmetic, because the numbers do not close and you would notice. 4,656 pages were read and 568 suppliers were kept, which leaves far more than 1,058 unaccounted for. The rest never reached a rejection decision: dead domains, parked pages, pages the scraper could not read, and pages that were plainly not a company. Only candidates that got as far as being judged against the brief get a recorded reason, and those are the 1,058.

We do not publish a precision or recall figure. An earlier version of this page said 85 to 92 percent precision and about 70 percent recall, measured against buyer feedback. No such study was run, and inventing the number is worse than not having it. What we can offer instead is the files: eight shortlists with every score and every rejection in them, so you can judge the screen on its output rather than on our arithmetic about the screen.

What we do know, and it is not a recall figure: the engine misses suppliers whose site is thin, whose category words never appear on the homepage, or who describe themselves in terms the screen did not weight. If you want those, widen the brief and accept more noise.

Stage 3. Enrichment: 568 kept, 300 with an address you can write to

Stage 3 takes what survived and looks for a contact address on the company's own pages, plus a short capability summary. For EU suppliers who publish a VAT number, that number goes to VIES.

Across the eight runs, 300 of 568 suppliers came out with a published address. Call it 53 percent, and then ignore the average, because the spread is the whole story:

• Packaging: 134 of 170, 79 percent. Converters publish a sales inbox because that is how they get orders.
• Event catering, Chicago: 50 of 77, 65 percent.
• Cold chain logistics: 14 of 23, 61 percent.
• Medical disposables: 28 of 58, 48 percent. Regulated manufacturers route enquiries through forms and regional offices.
• Kitchen equipment: 1 of 7. That run failed for reasons covered in its own report, and the contact rate is a symptom rather than the cause.

What Stage 3 does not do. It does not guess. A supplier with no published address is exported with that field empty, not filled with a constructed procurement@ that will bounce or land in a reception inbox. It does not query trade registries: there is no Handelsregister, KRS or Companies House lookup anywhere in the pipeline, and an earlier version of this page said there was. It does not infer names from LinkedIn.

That is a smaller claim than the one this page used to make, and it is the one the files support. If half your shortlist arriving without an address sounds like a problem, it is: it is the single most useful number for planning what happens after a search, and it is why the number is on every run report rather than buried.

Stage 4. Auditor: every survivor read again and scored

Stage 4 re-reads each surviving supplier against its own pages and scores the match to your brief from 0 to 100. Across the eight runs the median score per run landed at 85 or 90, with a best of 95. Nothing scored above 95, which is worth knowing: the engine does not award certainty it cannot justify from a website.

139 of the 568 had a certificate named on their own pages, ISO 9001 and its relatives mostly. That is recorded as a claim, with the page it came from. It is not verified. There is no lookup against IAF CertSearch or any issuing body anywhere in this pipeline, and this page used to say there was.

The same goes in the other direction. Where an EU supplier publishes a VAT number, VIES answers, and that answer is real. Where they publish nothing, the field is empty. Registry status is not checked, website dormancy is not checked, and certificate expiry is not tracked, because none of those are built.

Nothing is dropped quietly at this stage. A supplier the engine scored low is in the file with its low score, which is the point of publishing the files rather than a summary of them. You get to disagree with the score, and the score tells you which pages it read to arrive at it.

What you hold at the end is a spreadsheet: name, country, city, website, contact emails, score, specialisation, certificates. Sending the enquiries is your job, and there is no button here that does it for you.

Where the pipeline still misses (honest failure modes)

No one ships perfect software. Four places this pipeline underperforms:

1. Local-language directories we do not scrape. Some regional portals (certain Polish wholesale portals, Turkish B2B sites behind login walls) are not in our scrape set. Suppliers who list only there will be missed unless they also have an indexed website, which many do.

2. Very niche niches with fewer than ~10 global suppliers. If your category has, say, six manufacturers worldwide, highly specialized medical imaging components, very specific aerospace alloys, the Strategy stage generates queries that work for broader categories and may miss the long tail. Manual supplementation is required.

3. Shell companies and listing farms. Some entities exist only to appear in search results. They have a domain, a thin website, and nothing behind it. Our Screener catches most of these because the website is thin, but sophisticated shell operations occasionally pass. Auditor usually catches them at the VAT or registry step; occasionally one slips through.

4. Non-English local-language content that is nonstandard. Regional dialect variations, southern German versus Swiss German industry terminology, various Portuguese variants, sometimes produce thinner results than the main language mode. This improves as we expand language models, but it is a real gap today.

We ship these disclosures rather than hide them because buyers who rely on the pipeline without knowing its limits make bad decisions at the edges. Knowing the failure modes makes you a better user of the tool.

Why this matters to you as a buyer

Three reasons a buyer should care how the pipeline actually works, beyond "it finds suppliers faster."

Reproducibility. Run the same brief in January and again in July, you get an updated list with new entrants, removals, and status changes. Manual research is not reproducible at that fidelity; two researchers running the same brief produce different lists. If your category strategy depends on seeing the supplier universe evolve, the pipeline version is the one that actually tells you that.

Audit trail. Every URL the system looked at, every reason a candidate was dropped, every verification decision is logged. If internal audit or a customer compliance review asks "how did you qualify this supplier?" you can answer with specifics. Manual sourcing produces an "I looked at some websites and sent emails to the ones that looked good" answer, which is not what compliance wants to hear.

Language coverage without hiring. The biggest practical reason to run an assisted pipeline is language. If your category requires reaching German, Turkish, and Italian suppliers, doing that manually means either hiring multilingual researchers (€€€) or accepting that your English-only process misses 60-70% of the real supplier base. The pipeline handles that without the hiring step.

What we can put a number on is our side of that comparison, because it is published: eight briefs, 4,656 pages read, 568 suppliers kept, 300 with an address, 83 minutes of compute in total. The slowest run took 16 minutes and the fastest 4. We have no measured figure for how long the same eight briefs would take a person, and the ones that used to sit here were invented, so the comparison stops at the half we can evidence.