Selling Future Max to people who already trust us

Plan for Jack · 2026-10-28 · nothing is built or changed in Meta yet. Click anything to leave a note.

AWhat we know today (checked live, 2026-10-28)

• The pixel is live in both places. Pixel 4483693801879235 "Yummy Future Website" logged 11,836 events in the last 30 days: 10,659 from our site and 1,117 from wefunder.com. So the Wefunder Marketing Toolkit hookup works. 6,265 of them came from our server (CAPI), so server reporting is on too.
• Conversions are rare. 30 days across all site forms: 13 Lead, 14 QuoteStart, 76 FormEngage. Meta wants about 50 of the target event a week before it can steer an ad on it. So we can't optimise on "machine lead" yet (see section E).
• Today the instant form beats the website. Since 10-26, H4 (instant form) got 3 leads for $19.48. H6 (same carousel, sent to /max) got 0 leads for $7.35. Over 30 days to 10-22, 707 real ad visits to /max brought 0 quotes. These were cold strangers, though. Warm people have never been tested.
• Warm audiences we have: IG @yummyfuturebay engagers 365 days (12,700 to 14,900 people) and the 2025 Wefunder investor list (about 1,000, last updated 2026-01-26, needs a refresh export). Not built yet: "visited our site 180 days" and "visited wefunder.com/yummy 180 days". The pixel can build both today.
• $1/day is very little. Our machine tests buy about 38 views per $1, so a 5-day $5 test gets about 190 views and 5 to 8 clicks. That is below the 300-view "ran enough" bar in the test skill. Section F deals with this.

Sources: Graph API pixel /stats (30 days, by host, event and source), machine campaign insights 10-26→10-28, custom audience list, memory machine-sales-strategy-2026-10-12.

BThe hidden page: yummy-future.com/max-insiders

It uses the layout of the old /invest page, which is calmer than /max, to sell the machine.

• Hidden: noindex, not in the nav, footer or sitemap. Only ads and links we send lead there.
• Who it talks to: "You already know Yummy Future." Investors, followers and store regulars. It does not explain what a robot café is from scratch.
• Order on the page: ① real arm video + one line + "Get Max for your space" ② proof: 3 stores where robots make the drinks (facts already on the site) ③ what Max is: the capacity block from the spec sheet (no price, the price switch stays off) ④ two ways to help: "Put Max in my space" or "Introduce us to a space owner" (investors know building, campus and hotel people) ⑤ a short form: phone, email, company; optional: which way, where.
• Form: the site's normal lead form, so each submit lands in the leads table and in Slack like the others, tagged max-insiders plus the ad tags. It also fires pixel Lead (browser + server, counted once).
• Rule 204: the page never mentions the raise (no Wefunder, no SAFE, no "investors earn"). That keeps it a normal product page, so machine ads to it stay in the normal ad category. It thanks no one for investing and quotes no numbers about the round.
• Works on phones (same mobile rules as /max). Tests run green before you see it.

CThe A/B: Wefunder page vs our page

It runs as two ad sets side by side, not Meta's built-in split tool, because the two arms sit in different ad categories. The same person may see both. At $1 that hardly matters.

After 15 days, the winning landing gets the warm part of the steady money. At $30 an arm we may see 0 to 2 hand-raisers per side, so the call might rest on clicks and page views. If it is still a tie, run one more round.

DBudget table

Machine steady $10, pick one:

This replaces today's "$15 main + $5 tests (3 × $5 / 3 days)" and the Wefunder "$15 + $5" from this morning. The current T1 to T4 tests are already paid for, so they finish (about 10-31). The new system starts when they end, so machine spend never goes above $20.

EPixel events and what each ad optimises for

Optimising for: every website ad optimises for landing-page views, with the pixel attached so Lead, QuoteStart and the rest are reported per ad. "Connect the pixel to every ad" = this. When an ad's pixel events hit about 50 a week, switch it to optimise for that event. At $10/day that is unlikely soon. Instant-form ads (if D1) keep optimising for form leads.

FThe iteration loop (every 5 days, both lines)

• Batches, not drip: all 10 slots start and end together, so they compete on the same days in the same auction. Next batch the next morning.
• Each batch: 8 new ideas + 2 "confirm" slots for the last batch's top 2. One batch is pure luck at 190 views. Two batches in a row is a real signal.
• Ranked by the meta-test-ads skill: your YF index per $1, using each line's own weights from /today/ads2 (machine has its own set). Real machine leads come from our leads table, minus spam and tests.
• "Ran enough" bar drops from 300 to 150 views, since $5 buys about 190. A test under 150 is re-run once and then dropped.
• Promotion: a test that ranks top 2 in two batches running and beats the steady ad's index per $1 over the same days → its ad is copied into the steady ad set. The steady set holds at most 3 ads. The one with the lowest index per $1 over the last 7 days is paused. Budgets don't move: steady stays $10, tests stay $10.
• One change per idea: a new test changes just one thing from its parent (picture, first line, audience or landing), so we know what won.
• Where ideas come from: combinations of what you've already approved: 8 approved pictures/videos × 5 audiences × 3 landings × approved first lines. A new picture or new copy still needs your OK on a board. When fewer than 10 ideas are left in the queue, the daily report says so and brings a short list.
• Who runs it: the existing daily 09:00 task (machine-ads-rotation-3day), extended to both lines. Every change goes into plan.json, the ledger and memory, and you get a short report.

GKill rules

• Tests: never stopped early for weak numbers (it's $5; stopping early just loses the data). Stopped early only when an ad is rejected, the link is broken (10+ clicks, 0 page views), or the ad is bleeding junk leads.
• Steady ad: replaced when a confirmed winner beats it. With 7 days and $70 spent, 0 hand-raisers and an index below the test median, it is swapped for the best test even without a confirmed winner.
• Too many repeats: warm audiences are small, so the worker1 rule still applies: shown more than 1.3 times a day → budget cut. Test ad sets are skipped by that rule (it leaves lifetime budgets alone), so it can't fight the $1 tests.
• Whole line stop: machine line, $300 spent from the start with 0 real machine leads → tests pause, steady drops to $5, and I bring you a rethink. Wefunder line: $300 with 0 investors → stop (what you approved this morning).
• Hard ceiling: after every change the task adds up each line's daily spend (daily budgets + lifetime ÷ days). Over $20.01 → undo the change and report.

HOrder of work after your OK

1. Build the hidden page on a branch, tests green, preview link for you. Nothing goes live until you say "ship".
2. Update the test skill and the daily task to the new rule ($1/day × 5 days, half and half, batches of 10, confirm round, 150 views).
3. Build the Meta audiences (site visitors, Wefunder-page visitors, warm combined) and the "Insiders machine lead" custom conversion. These cost $0.
4. One "一次批准" question that lists every Meta change with its $/day: the new steady ad set(s), batch 1 (machine), and arm W (once the Wefunder drafts are OK'd). Then each write is done one at a time and read back.
5. Batch 1 starts the morning after T1 to T4 end (about 10-31), so machine never goes over $20.

Waiting on you: the investor-list export from Wefunder (to refresh the 1,000-person audience), and your OK on the Wefunder IG drafts board (arm W uses one of those ads).