Paresh Shrivastava — Growth marketer who builds AI agents. Built with Claude Code.
646 companies scored, 74 ranked by how likely they are to buy this quarter, 2 worth calling today. The most useful part is the 18 I threw out — starting with the one that was my own second-best lead.
Three things from your own job ads:
So I built what that seat builds in week one. This isn’t a proposal. It’s the output.
Their share of America’s dispute complaints jumped by a third in one month. Textbook trigger. I’d have called them first.
It was fake.
Capital One bought Discover. From that month, Discover’s complaints stopped being filed under their own name and started being filed under Capital One’s. Nobody’s disputes got worse. The complaints just changed their name tag.
Put the two companies back together and the line is flat.
One company in 646 does this. No statistical test catches it — the numbers it corrupts look perfectly healthy. You only find it if something goes looking.
Another 15 are out because their share is falling — several of them while their raw complaint count goes up. That’s the difference between counting complaints and measuring share. Counting just finds big companies, and you already know who those are.
| Tier | N | What your rep does | What the email may claim |
|---|---|---|---|
| Hot | 2 | Call today | Share is rising. An AI tried to kill it and failed. |
| Warm | 1 | Sequence this week | Share is rising. It hasn’t cleared every check. |
| Watchlist | 15 | Check back monthly | Something is moving. Can’t prove it yet. |
| ICP | 38 | Normal cold outreach | Right company. No reason to hurry. |
| DQ | 18 | Do not touch | Out — and the reason is on the row. |
All 74 companies land in exactly one tier with a one-line reason. Nothing disappears quietly. And ICP isn’t the clever part — that is a list anyone could build from the raw data, and it is labelled so the easy half doesn’t take credit for the hard half.
Rising share, a second signal that agrees, and they survived an AI whose only job was to kill them. The line under each is what I’d actually send.
Americans complain to the regulator about their bank, and the regulator publishes every one with the company named. I read 109,004 of them going back 46 months and asked one question:
Share, not count. The whole pile more than doubled over four years, so “their complaints went up” is true of nearly everyone — that is the tide, not the company. Dividing by the national total each month cancels the tide, and cancels Christmas with it.
The second signal is what happened to each complaint: did they refund the customer, and did they reply late. Nationally both moved the wrong way over the same period, so a company only counts if it moved against the market, not with it.
Every test and threshold is on the workings page, including the parts that argue against me and the one safety check that turned out to be contaminated.
One Claude agent writes the best case for each company. A second agent, on a different model, is told to destroy it — and never sees the first one’s own list of doubts, because anyone handed a list of known flaws checks that list and stops looking.
181 recorded tests. The analyst’s own list of doubts got within one step of the merger and stopped. The attacker, shown none of it, scanned every company in the database and found the one that had vanished.
| Step | Who does it |
|---|---|
| Find the account, prove the trigger | This. Out comes a CSV with the tier, the trigger, the evidence and the date |
| Find the human | Clay. Waterfall enrichment onto this list. Not worth rebuilding — you already have it |
| Send it | Outreach. Tier maps to sequence, evidence becomes a merge field, so the opener writes itself |
| Learn from it | Yours — and it’s the part that makes the scoring get better |
| If this moves | It’s telling you about | So you change |
|---|---|---|
| Open rate | the subject line | the subject line, and nothing else |
| Reply rate, Hot vs ICP | the scoring | the account logic |
| Reply rate inside one tier | the writing | the copy |
| Meetings booked | whether the whole idea holds | the premise |
Watch replies, not opens. And the blunt version: if Hot doesn’t beat ICP on reply rate, my scoring is wrong and you should bin it.
A full refresh — re-pull, re-score, rebuild every page — takes about four minutes and costs nothing: the data is public, and the only paid step runs on the handful that survive the free filters.
That is the half of this that doesn’t fit on a page — and it’s the half that decides whether a list like this is worth your reps’ time.