Lead Extractor

The full text is printed under this drawing, built from the very constants a run sends, so it can never drift from what actually happens.

A term with 44 products but 2 only-here landed in ground we had already covered. One with 31 only-here opened a door. That is the shape worth repeating — more than the qualifying count, because the goal is finding new niches next to our brands.

Nothing is scraped automatically. A term costs a night of proxy bandwidth, so every suggestion waits for you.
No term is ever repeated — enforced in the prompt and again in code, because a model slips.
The AI never sees a product below the qualifying line, so an on-brand product at $6k/mo cannot inform it.
How many terms is your prompt’s decision. Nothing the code adds asks for a number. There is one runaway guard at 1,000 terms, far above anything a person types, and it says so in the run log if it ever trims.

Every night at 3:30 AM Pacific, with a rescue pass at 9:30 AM that only re-runs terms the night failed on. You can also press Run on the leads page and go immediately.

All of it is read off Amazon’s own search page — no Keepa, no paid API. That is why a product with no “N bought in past month” badge has no revenue: Amazon did not print one.

Nothing is thrown away, and nothing leaves on its own. A known brand is tagged, not dropped; a product that misses the thresholds is stored; and a product is archived only when you archive it — no nightly job ever removes one (owner 2026-08-15).
Only page 1 is read — past it, nobody is buying.
A term whose every location failed stays queued and retries by itself, on the same day and every following run.
A term that came back from one location but was blocked in the other counts as done and is not searched again — you get that term’s results from one place instead of two.

The list — a whole-word match of the title against the brands in ⚙ Settings. A match is tagged with that brand at scrape and hidden by the default chip — never dropped, so taking a name off the list brings its products straight back.
The AI’s judgement — one question: would a shopper go looking for this NAME?
It answers from what it already knows about the brand. That is the strongest signal.
Plenty of reviews, products or sales on Amazon prove nothing.
A supporting hint: a real brand usually puts its name up front in the title, because the name helps sell.
It answers real, borderline or not a real brand, each with a percentage.
A name it does not clearly recognise is called not real rather than guessed at.
A name it DOES recognise it puts to you and never acts on alone; a not real verdict settles itself, and that product is not sent again.
The Known brand? column shows the AI’s ANSWER and nothing else: a name, a quiet No, or empty when nobody has looked yet. You answer it in the AI right? box beside it — including on a No, which had no control on the row at all until 2026-09-03.

Until 2026-09-03 this happened automatically, and that was the problem. Every answer you gave was assembled into the next run’s prompt by code — examples, corrections, both brand lists — and none of it appeared anywhere you could read it. Two sorts a day apart could be judged by different instructions with nothing on screen saying so.
Now it is three deliberate steps.
One — you answer AI right? on the rows you have an opinion about, and say why in the box beside it. That is stored, and it changes nothing on its own.
Two — you tick those rows and choose Bulk actions → Suggest a prompt improvement. The AI reads the current prompt plus what you said, and proposes a rewrite. Rows you marked Correct go too — they are what stops a fix breaking the answers it already gets right.
Three — you see the prompt before and after, edit it if you want, and press Approve. Nothing changes until you do. The old version is kept and is one click away in the History box.
If your answers show no repeated problem it returns the prompt unchanged and says so — a change nobody needed costs more than no change at all.
A brand suggestion you have not answered is not settled, so every run judges it again. Answer it and the row is locked and never re-asked — and the AI’s own not a real brand locks it too, which is what stopped a sort re-buying an answer for 178 of 240 products every time it ran.
Want it stricter or looser right now? Edit the prompt under this drawing — it is the whole rule, in plain words, examples included.

The prompt forbids it: if the list already holds Bonsai Growing Kit, a product of that kind must be given that exact name — not “Bonsai Starter Kit”, “Bonsai Kit” or “Bonsai Grow Set”. Words like starter, growing, complete, DIY, kit, set, pack and bundle do not make a new niche.
And the code checks anyway: any new name that means the same as an existing one is renamed to the existing one before it is written, and the swap is printed in the run log. It works across runs and inside a single run, so two products judged minutes apart cannot split a niche in two. A genuinely different product — Dog Water Fountain beside Cat Water Fountain — is never merged.

Every answer is stored with the value it replaced. ↺ beside the niche puts one back; Undo all on the banner puts a whole run back. Typing over a niche removes the AI mark by itself — the row is yours then.
A named niche is never re-judged; only blank ones, plus any brand still waiting on your answer.
Only qualifying products are sent at all — never an archived one, and never one already tagged with a brand from your list, which the page hides anyway.
A product it will not name is left blank, never guessed at.