The order queue only 2 people could clear
A marketplace that sells trust is only as good as its last check. Code checks every placement at delivery and every week after; people see only the failures.
Illustrative example, not a client. Figures are modeled.
- Shop
- Self-serve SEO marketplace, about 15 people, remote-first, founder-led: agencies buy placements on publisher sites, the marketplace brokers and verifies them
- Surface
- Fulfillment. The gap between “publisher says it’s live” and “agency gets a report it can send its client”
- The constraint
- 2 senior ops people verify every placement by hand. Volume grew 40% in a year; they didn’t
- What triggered it
- An agency’s client found 11 of their links had quietly been set to nofollow 3 months after delivery. The agency asked for a refund on all of them, and posted about it
- What we built
- A placement verifier: code checks every link at delivery and weekly after, a model reads the page around it, and a person sees only the ones that fail
Placements checked at delivery and re-checked every week, with no added headcount.
Median time from publisher “live” to the agency’s report, once the ops leads read only the 9% that fail.
Every order ends the same way. A publisher marks the link live. One of 2 ops leads opens the page, finds the link, checks the anchor text, checks it’s followed, checks the page is indexed, checks the article isn’t a thin rewrite of something else, and writes it into the report.
6 to 9 minutes each, about 2,800 times a month.
Nobody checks again after that. Publishers know it. Some flip links to nofollow after 60 days, or delete the article and keep the fee. The marketplace finds out when an agency’s client does.
They had tried a crawler that confirmed the link existed. It confirmed 11 links that were all there, all nofollow, and all worthless.
We started with the 11 links. The first rule: no placement is “delivered” until code has checked the things the ops leads check by eye, and a placement that fails a re-check is flagged to the agency before their client sees it.
| Check | Who runs it | Result |
|---|---|---|
| Link present on the page | Code | Pass |
| Anchor text matches the order | Code | Pass |
| Followed (no nofollow, sponsored, ugc) | Code | Pass |
| Page indexed and responding | Code | Pass |
| Written for a reader, not to hold linksanswers with a quote from the page and a reason | Model | Unsure → person |
Then the weekly sweep. Every live placement is re-checked. A link that changes status opens a ticket with the publisher automatically, with the evidence attached, and the agency’s report updates itself.
In the first modeled week: 14 status changes, 14 publisher tickets, 0 found by a client first.
The model was too strict at first. It flagged 30% of new articles as thin, because publishers in some niches really do write short. The ops leads overrode most of them, so we stopped asking “is this thin” and started asking “is this written for a reader or for a link.” Flags dropped to 9%, and the overrides stopped.
And the sweep found what nobody wanted to know: 6% of placements from the previous year had already been changed or removed. That’s a refund liability the business now knows about, and one a buyer will ask about. The fix was to make good on them before listing, not after.
Modeled. The ops leads stop reading every page and start reading the 9% that fail, which moves median delivery from 4 days to same-day and absorbs the next year’s growth without a third hire. The weekly sweep turns “we verify every link” from a sales line into a log.
Not measured: whether agencies renew more when the sweep catches a flip before they do. That needs a renewal cycle, and a demo has none.
The 2 ops leads writing down what “good” looks like, then letting code hold them to it.
Agreement that a flagged placement is not delivered, even when the publisher insists.
And a founder willing to look at the 6% before a buyer did.
Find out what a buyer will price into your handoff before the buyer does.
If the read finds nothing a buyer would price into your transition, we say so in writing: list as-is. The map is still yours.