How it works

Part one — setting a client up

Six steps, three of which you can walk away from. Nothing here scrapes at volume; the whole setup costs roughly what a tenth of one run does.

1. 01

   Paste the website

   You do this

   On New client, give it the client's website and press Read the website. It reads the homepage and a few pages that describe the business.

   ScreenshotThe setup page with a website URL enteredhow/01-website.png

2. 02

   Check the description

   You do this

   You get a plain-English description of the business back. Read it, and fix anything wrong before going further — this description is the rubric the AI judges every single post against for the life of the client.

   The sentences about who is not a customer matter as much as the ones about who is. If there is a geographic limit, say so explicitly — left unsaid, the model guesses. If a competitor sells something that sounds similar but is a different purchase, name it.

   ScreenshotThe generated business description, editablehow/02-description.png

3. 03

   It imagines the buyer, then goes looking

   Runs on its own

   Press Find where the buyers are. Before proposing anything, it writes a handful of posts the way a buyer would write them — not the way the business describes itself. Everything after this is built from the buyer's words.

   This step exists because of a measured failure. Asked to write keywords straight from a business description, the model produced macro counted meals. The top post in that client's best community said “please recommend good meal delivery”. Thirteen of fifteen generated keywords appeared in no real post.
4. 04

   Every guess is tested against live Reddit

   Runs on its own

   Each proposed community is checked — does it exist, is it still alive, and does it actually talk about this business. Each proposed search query is run once at five results, so a query that returns drama posts is caught before it costs anything.

   ScreenshotOnboarding in progress, checking search queries one at a timehow/03-probing.png

5. 05

   You see the evidence, not just the answer

   You do this

   Nothing is accepted quietly and nothing is rejected quietly. Kept communities show what they returned; discarded ones show why. Dead keywords are listed so you can see which of the generated words matched nothing real.

   ScreenshotThe review screen showing kept and discarded subreddits with reasonshow/04-evidence.png

6. 06

   The searches name communities nobody suggested

   Runs on its own

   This is the one step that can find a room nobody thought of. A site-wide search asks the whole of Reddit the buyer's question, and the answers come back from somewhere — those communities are listed with their evidence and a Watch this button.

   They are never added automatically, because watching one costs money on every future run. Worth knowing why this matters: for one client every qualified lead came from city communities the config-writing prompt of the day explicitly forbade naming. The searches walked into them regardless.

   ScreenshotCommunities discovered by the search probes, with a Watch this buttonhow/05-discovered.png

Part two — what a run does

Four steps, and you are only involved in two of them.

1. 07

   Press Run

   You do this

   On the Leads page. The estimate above the button tells you what it will cost before you click. Nothing is on a timer — a run happens because you asked for one.

   ScreenshotThe Leads page with the Run button and its cost estimatehow/06-run.png

2. 08

   It collects, then filters for free

   Runs on its own

   Every target is fetched, and anything already in the library is dropped without being paid for twice. Then a deliberately wide keyword pass runs against the database — this stage costs nothing and is meant to be generous. Precision is the next stage's job.

   A community that cannot have produced anything new since the last look is skipped rather than re-fetched. Before that existed, one measured re-run fetched 360 posts and found 1 new — and paid for all 360.
3. 09

   The AI reads what survived — all of it

   Runs on its own

   Each surviving post is read against the business description and scored on five dimensions, with a reason, a summary and a drafted reply. Eight at a time.

   One press finishes the whole queue. It used to stop at 200 and report success, leaving the rest until somebody noticed and pressed Run again. It no longer does — the progress counter runs to the real total.

   ScreenshotA run in progress showing posts scored out of the totalhow/07-progress.png

4. 10

   You post. As you.

   You do this

   Open a lead, read the drafted reply, edit it into your own words, and post it yourself. ThreadSpot has no way to post — that is a design constraint, not a missing feature.

   ScreenshotAn expanded lead showing the score reason, summary and drafted replyhow/09-expanded.png

Part three — reading what comes back

A single yes/no score threw away too much. A post can be exactly about the problem you solve and still not be someone ready to buy today — that person is worth talking to, and the old design deleted them. So results arrive in three tiers, and they never overlap.

ScreenshotThe Leads, Conversations and Topical switch with live countshow/10-tiers.png

Reading a row

• Open is your queue — anything not yet marked Reviewed or Dismissed. Marking one files it away. Nothing is ever deleted.
• Sorted by newest post by default. A 0.80 from two hours ago beats a 0.90 from four months ago, which is answered and dead.
• HOT is 0.70 and up, WARM is 0.30 to 0.69. Below 0.30 is hidden rather than deleted — the show below 0.30 toggle is there so you can audit what the filter rejected.
• Screened shows what was scraped but never became a candidate. That is your denominator: it tells you whether the keyword list is too narrow.
• A blank score means scoring failed, not that the post scored zero. Those retry automatically on the next run, up to three times.

What it costs

Limits worth knowing

• Posts removed by Reddit are invisible to the collector even though they still look normal to whoever wrote them. Testing with a post you made yourself? Check it in a logged-out browser first.
• The keyword filter is deliberately wide. Precision is the scoring model's job. A narrow keyword list produced a 1% hit rate and an empty dashboard in testing.
• Negative keywords only reject a post that matched nothing positive. They used to be absolute, and quietly destroyed real leads — a bare hiring once dropped the strongest lead a client had, titled “worth hiring someone”. Never put a place name in the negative list: in an expat market the buyers are the people who moved.
• Subreddits govern collecting only, not matching. Search results arrive from communities you never listed, and most of the best leads found so far were in exactly those.
• Each account's posts are its own. A post collected for one client is not visible to another, even when both watch the same community and the post was only paid for once.

Current configuration

The pipeline API is not responding, so the costs and limits above may not reflect what is actually configured.