SLSBMB AI vs a chatbot with a good prompt
A chatbot gives you a draft in minutes. We tested exactly how far that draft is from outreach you can send, on one real brief, and published both sides of the comparison.
Updated 2026-10-30 · Markdown
How we tested
On 29 October 2026 we took the published Revopush playbook and gave the same task to one of the strongest AI models available, called directly with no tools, no web access and none of our instructions or memory. That is what chatbot products run on underneath.
The model got the same website text our system read and the human-readable part of the brief the client confirmed. That is generous: a founder who opens a chatbot alone has no such brief, because our interview is what produces it. The request was the one a capable founder would write: find the audiences I should email, with who they are, why they would care now, evidence and size, priorities and budget shares, including ones I haven't thought of; write an opener and two follow-ups for each; and give me search filters specific enough to run in a contact database.
The two results
| Chatbot with one good prompt | SLSBMB AI playbook | |
|---|---|---|
| Time | 2 minutes 51 seconds | 2 hours 5 minutes of agent work after the interview |
| Price | About $0.40 at API list price | $500 at the launch price |
| Tokens | About 25,000 | Tens of millions |
| Sources read | None; the model said it had no web search | 107, with 73 direct links to 62 websites in the text |
| Audiences | 8 | 14 |
| Research | 1,777 words across all audiences | 36,292 words, 2,200–2,800 per audience |
| Market size | Orders of magnitude "for planning" | A reach range for every audience; 3–9 million people in total |
| Emails | 24, averaging 64 words | 42, averaging 243 words |
| Search | 5–11 titles and 3–12 industries per audience, about 700 combinations plus keywords and manual checks | 31,847 company-type × job-title combinations: 41–127 title variants and 9–37 company types per audience |
| Where sending starts | 1,750 test contacts to check by hand | The system keeps finding people from all combinations as sending goes on |
What the model did well
It produced a decent draft fast: eight sensible audiences, several of which matched ours (white-label app platforms, React Native agencies, regulated industries that need self-hosting, delivery and marketplace apps on weak networks, betting and live events). It kept an unverified price comparison out of the emails, as we did, and it said plainly that it had checked nothing and that its market sizes were estimates.
Where the difference is
- Evidence. The model worked from memory and the brief. Every claim in our playbook rests on a source: the client's documentation and prices, the retirement of Microsoft App Center, competitors' price lists, engineering posts and job descriptions that show the real pain and the exact language of each role.
- Audiences you can actually find. Two of the model's eight audiences are defined by traits no contact database has, such as "former App Center users" and "large Expo users". Ours are markets a search can reach, and five of them are missing from the model's answer altogether: B2B software, device companion apps, fund operations teams, developer-tool ecosystems and internal apps with 100,000+ users. The model even advised excluding developer-tool ecosystems as partners; our playbook made them a separate path to shared customers.
- A search that executes. The model itself noted that contact databases can't filter by technology stack or monthly users, so it suggested keyword searches and checking a sample by hand. Our playbook builds its search from what databases can do: company types in every variant of their name and job titles in every spelling, 31,847 combinations from which the system finds people by itself.
- Emails you can send. The model's emails are short and tidy, but the third email in all eight sequences is a goodbye, seven of them asking who else to write to, and two sequences still contain bracketed placeholders that can't be sent as they are. Our 42 emails go out as written, and each follow-up brings its audience a new argument: for fintech, what to tell the security team about code signing; for app platforms, why their release calendar is the sum of every brand's calendar; for internal apps, what it means to push gigabytes through one store channel at the start of a shift.
A chatbot gives you a draft in minutes. The work is hours of research, verified facts, a search that runs and emails you can send, and that is what the playbook is.
Frequently asked questions
Can I just use ChatGPT for cold email?
For a first draft, yes. For outreach that runs, the draft is where the work starts: evidence for every audience, searches a contact database can execute, and emails written to be sent as they are. That work is what the $500 playbook contains, and you can read three complete ones on the examples page before you buy.
Why does the playbook take hours instead of minutes?
Because it reads and checks. For Revopush that meant 107 sources, more than 36,000 words of research and 31,847 search combinations, produced by specialist AI agents working from more than 100,000 lines of instructions and doctrine. One playbook takes nearly fifty million tokens of work; a chatbot answer takes a few thousand.
Was the comparison fair to the chatbot?
It was generous. The model received the client-confirmed brief that our interview had produced, which a founder working alone in a chatbot wouldn't have, and a request covering everything a good outreach plan needs. This is one test on one brief, run on 29 October 2026; the full playbook it was compared with is public.