The label said verified. We called it anyway.

We build a tool that finds people and their phone numbers. This week someone on our own team sat down and called the list it produced, one number after another, and wrote down what happened on each call. That is the whole story. Not a study, not a benchmark - one person, a phone, and a list our own software was sure about.

The list was labelled. Each number came back from the data provider with a confidence score, and a lot of them read 90 percent. Verified. Valid mobile. The kind of number you are supposed to be able to just dial. So we dialled them.

Fewer than one in ten went to the right person.

That is the number that stopped the room. Not "a bit lower than we hoped." The provider said 90 percent confident, and the real hit rate on a live call was under 10 percent. Some numbers rang busy. Some went straight to voicemail with a name that was not the name we expected. Some were disconnected. And a few - this is the one that stays with you - belonged to the right family but the wrong person. One number we had filed under a prospect turned out to be his father's. The label was not lying about a phone existing. It was lying about who would pick it up.

Being honest with ourselves

The easy version of this post blames the data provider. It is their number, their label, their 90 percent. But we shipped it. We put that number in front of a salesperson and let them believe it. So the honest version is that we trusted a label we had not checked, and a label you have not checked is just a rumour with a decimal point on it.

Here is exactly how we got it wrong, because the mechanism matters more than the apology. Our tool did the sensible-looking thing. When a number was missing, it went and searched the web to find one. That part worked. But when a number was already there, already stamped verified by the provider, it left it alone. Why would you re-check something that says verified. That was the bug. Not a broken line of code - a broken assumption. We treated "verified" as if it meant verified, when all it ever meant was that this number was once seen next to this name in some database, at some point, by someone who was not us and was not on the call.

A phone number is not a fact about a person. It is a fact about a moment. People change jobs, change numbers, hand an old handset to a kid, put a landline down as a mobile because a form demanded a mobile. The database captured a true thing once and kept serving it long after it stopped being true. The confidence score described how sure the provider was when they wrote it down, not how sure you should be when you dial it two years later.

What "verified" actually costs

The reason a 90 percent label that is really 10 percent is so expensive is that it does its damage after you have already committed. A blank field, you work around. You see there is no number, you go find one, you move on. A confident wrong number you act on. You spend the call. You spend the salesperson's morning. You dial his father, you leave a voicemail for a stranger, you cross a name off because "we tried them and got nothing," when the truth is you never reached them at all. The bad data does not just fail to help. It actively closes a door that was never really shut, and it does it wearing the costume of good data.

And it compounds quietly. One salesperson calling one list can feel the wrongness by lunchtime - the calls are not landing, the people are not who the sheet says. But scale that up and nobody is on the phone to notice. The numbers flow into a dialer, the dialer reports "attempted," a dashboard turns the attempts into a percentage, and the percentage looks like progress. The only place the truth lives is on the actual call, with an actual person, and that is the one place the system had stopped looking.

The fix is not a better label. It is doing the check.

So we changed what the tool does, and the change is small to describe and the whole point of the thing. Now, before it hands you a number, it verifies that the number belongs to that person - this name, this company, this number, together - rather than trusting that someone upstream already did. It grounds the check: it goes and looks, across the open web, for evidence that ties the three things together, and if it cannot find that the number is this person's and not a relative's and not a general switchboard, it does not pretend. It says so. A number that fails the check is worth more to you as a flag than as a false green light, because now you know to go find the real one instead of burning a morning on the wrong one.

The lesson is not really about phones. It is about where verification lives. A confidence score is a promise somebody else made about a fact they did not check on your behalf. It is worth exactly as much as your trust in the person who wrote it, and usually you do not know that person at all. The only verification that counts is the kind that actually happened. A number you dialled. A reply you got back. A person who answered and was the person. Everything before that is a label, and a label is a thing to check, not a thing to believe.

Why this is the same idea, everywhere we work

We keep arriving at the same place from different roads. The best reply we ever got was not to a clever line, it was to a real moment - someone who had actually just changed jobs. The warmest conversations come from the network we already have, where the person is not a guess. And now the phone list: the number that was worth calling was the one we could actually tie to a real person, not the one with the highest score attached.

Grounding is the through-line. It is why Reach works the network you already have instead of a bought list of strangers - because a connection is a fact you can see, not a label you have to trust. It is why a human approves every message before it sends - because a person on the actual thread is the last, best check that the moment is real. We are not against data. We are against believing data we did not check, and we have now been embarrassed by our own enough times to mean it.

Someone on our team spent a day on the phone finding out that a 90 was really a

1. That was a good day. It is the cheapest way there is to learn that the score was never the thing - the call was. Everything we build now assumes the label is a question, and the only answer that counts is the one you went and got yourself.

---

Sources

First-party (the spine of the story)

Team meeting transcript, 2026-08-21 — Google Drive, folder 1drHKlrUpEv7jqPybpRS3_CVyiVUtgfId, doc 1Nuea_JQcOtOE2Zad17L7wkF3G07p3dz8UUH3HBp10JQ, Transcript tab (not the Gemini "Notes" summary above it). Read in full via the Drive connector this run.

Load-bearing facts, all quoted or closely paraphrased from the transcript:

• A teammate (Mike) called through the list our own Sales Enrich tool produced and recorded an accuracy of less than 10% — Tisha, 00:25:13: "Mike did have recorded the accuracy rate to be less than 10%."
• The failure held even for numbers the provider labelled with a 90% confidence score — Tisha, 00:25:13 / 00:27:09: "including for the phone numbers that were labeled by the provider as ... an accuracy score of 90%."
• Failure modes on the calls: "either busy or they ... were landing to voicemail or getting disconnected or sometimes even ... they didn't belong to the right person" — and specifically a number that "actually belong to their relative like maybe the father of that person" (Tisha, 00:27:09).
• Root cause: phases one and two only web-searched when a number was missing and trusted the provider's label when a number was present and marked verified — Tisha, 00:27:09: "if the provider has provided a valid personal phone number and labels it ... verified and has a confidence score of 90%, then ... it kind of trusts the provider's data ... the labels cannot be trusted."
• Fix: a new server-side "phone ownership" verification service — a grounded LLM + Google search tree that, given full name + company + number, decides whether the number ties to that exact person at that company (Tisha, 00:28:46).

The only party described as failing here is our own tool. No external company is named as struggling. "Mike" is an internal teammate, not a customer; he is anonymised to "someone on our team" in the public post and deck.

Supporting first-party context (not quoted in the deck): the same account's get_results this run still shows the pattern the grounding thesis rests on — the generic company pitch replies 16.3% (137/838) while a real, personal touch replies far higher — reinforcing that the number worth acting on is the one you can tie to a real person, never the one with the highest score attached.

Deliberately NOT cited

A web search for a primary study on B2B phone-data "verified" labels vs real connect rate returned only vendor SEO blogs — Salesfinity ("63–91% across nine providers"), Landbase ("50% accuracy on average"), emarketnow ("vendor claims vs reality"), plus Cognism/Lusha marketing ("87–98% verified"). All are self-interested and none attributes its figures to a named, methodologically transparent primary source. An attempted fetch of the emarketnow page returned HTTP 429. Per the house rule (open every source you cite; unattributed vendor statistics are not used), none of these is cited. The story rests entirely on first-party data, consistent with the 2026-08-20 story's precedent.

No price

No figure resembling a price appears in the case, script, carousel, captions or post.