Funnel Spy

Funnel Spy

Competitor research tells you what someone sells. Funnel Spy shows you how the selling machine actually works — by walking it the way a stranger would, one page at a time, and writing down everything the machine does to convert them.

For each target you deliver five things:

1. The channel map — every social platform they run, what each one is for,

the exact link each points at, and the single place they all converge. Paid entry points are verified in the ad libraries, not guessed. This is the funnel above the website, and most teardowns miss it entirely.

1. The Walk — the actual page-by-page path from first click to checkout,

captured live: every hop's headline, CTA, price, proof, and friction, with a screenshot as evidence.

1. The Machine — the funnel software, email platform, checkout processor,

and tracking pixels detected from page source. The stack tells you the strategy: a Deadline Funnel script means evergreen scarcity; a ManyChat snippet means comment-to-DM; a Calendly embed means a sales-call close.

1. The Offer Ladder — every rung from free to high-ticket, with prices, and

the order they're presented in.

1. The Scorecard — four gates (Attract / Capture / Convert / Ascend)

scored against a fixed rubric, a 0–100 total, and the weakest gate named. Only what an outsider can verify by looking — email nurture is never scored, because it cannot be seen. The weakest gate is the finding: it's either their leak or your opening.

Then a cross-target synthesis: shared funnel patterns, a Best-Engineered ranking driven by the scorecards, and — if the user gave an anchor — Your Opening and Steal This Week (three concrete moves).

Evidence rules

• Evidence tiers on every claim:
• CONFIRMED — seen first-hand on an owned/authoritative page. A page you

walked and screenshotted is CONFIRMED with proof attached — the strongest evidence this skill produces.

• REPORTED — stated somewhere but self-reported or third-party-estimated,

unaudited ("they claim $80K/mo"). Real signal, order-of-magnitude only.

• INFERRED — your reasoned guess, labeled as such.
• Never fabricate a price, follower count, conversion rate, or funnel step.

"Not found" is reported as not found, and absence of evidence ("no email capture found") is not confirmed absence — say which one you mean.

• The walk is a visitor's-eye capture. When a step needs a submission, an

account, or a purchase to continue, capture that page, mark what's behind it unobservable, and recover it through sitemaps and archives instead of guessing (references/funnel-walk.md has the protocol).

• Short verbatim captures only. Swipe-file items are headlines, hooks, CTA

labels, and guarantee lines — a line or two each — never wholesale page copy.

Step 1 — Normalize the targets

The user may paste a mix: @handle, instagram.com/..., "GoHighLevel", acme.com, or a direct funnel URL (a Linktree, a landing page, an ad's destination). Sort each into:

• Social handle → find the person/brand, then their funnel entrance (the

bio link is the front door).

• Company / brand name → find the site; the funnel entrance is usually the

homepage hero CTA or the pricing page.

• URL → if it's a homepage, start there; if it's already a mid-funnel page

(a Linktree, an opt-in, a sales page), start there and work backwards to find the entry points feeding it.

Infer which is which; strip leading @. Chase mistyped handles via search for the corrected spelling or real name before giving up. If a target genuinely can't be found, say so plainly — never invent a profile.

No list? If the user asks "who are my competitors?" instead of naming targets, run a short discovery first: search their niche + offer type ("[niche] course", "[niche] coach site:skool.com", "[their keyword] alternative"), skim who keeps appearing, and present 5–8 candidates with a one-line reason each. Let the user pick before spending walk effort — spying on the wrong five funnels thoroughly is worse than asking one question.

Step 2 — One question up front: format, depth, anchor

Ask once, as a multiple-choice question (never free-form typing), then start:

1. Output format — the Funnel Dossier PDF (recommended default — the

designed, shareable teardown), chat summary, spreadsheet matrix, or Word doc.

1. Depth — Full Spy (recon + live browser walk + ads check + tech

detection; the real product), Recon only (search-based, no live walking — faster, for a quick read), or Single-page teardown (the user pasted one landing/sales page and wants it decoded now: skip recon and fan-out, walk that page and its immediate chain, run the copy teardown and tech detection on it, score the CONVERT gate only, and deliver — minutes, not a full run). If browser tools aren't available in this environment, say so and run the fetch-based walk instead (WebFetch/curl per page) so the chain, copy, prices, and tech stack are still captured first-hand — only the screenshots are lost.

1. Anchor — optionally: what does the user sell, at what price? This

powers Your Opening and Steal This Week. Facts-only is fine too.

If the user already stated any of these, don't re-ask.

Step 3 — Set up the run

• Create one task per target plus one synthesis task so progress is visible.
• Create a run folder for evidence: funnel-spy-run-/ with

screenshots/ and portraits/ inside. Screenshots are named --.png (e.g. thedankoe-03-optin.png); portraits are portraits/.jpg.

• Grab each target's portrait per references/target-portrait.md — the

face goes on the dossier cover and on that target's spread. Look at each downloaded image to confirm it's real and the right subject; use the initials fallback plate rather than any invented or stock image.

• Every run also writes one machine-readable dossier.json per target into

the run folder (schema in references/re-spy.md) — it's the single source of truth the deliverables render from, and it's what makes future runs diffable.

• Check for a previous run first. If a funnel-spy-run-*/ folder with

dossiers for these targets already exists (or the user asks to re-check), this is a re-spy: run the same pipeline, then deliver the diff per references/re-spy.md — what changed is worth more than the snapshot.

Step 4 — Fan out: recon agents (parallel)

One research agent per target, launched in parallel in a single turn, each following references/recon-prompt.md. The recon agent's job is to come back with the funnel chain — the ordered list of URLs a stranger would traverse — plus offers and prices, entry points, a compact quantitative snapshot, and a links directory. Recon finds the doors; the orchestrator walks through them.

If subagents aren't available, do the recon yourself per target, same standard.

Enough is enough: recon should confirm the funnel entrance, the core offer and its price, and the capture mechanism. Don't burn searches chasing every minor product — note what wasn't found and move on. The walk will fill gaps.

Step 4c — Map the channel layer (the funnel above the website)

Before walking the site, map the social layer per references/channel-map.md: every platform the target runs, its audience, its job (primary engine / reach / authority / B2B / repurposed), and — the key field — the exact URL each one points at. Then the convergence: the one or two pages every channel feeds.

Hunt for per-platform landing variants (/offer-yt, /offer-ig, utm variants). Probe the obvious siblings, always probe a nonsense control to prove the site 404s on unknown paths, and diff the variants: identical content means attribution tracking, different content means that channel gets bespoke effort. Record cross-channel routing too — short-form feeding long-form, a site-wide socials directory, or multiple principals each running a full stack into the same funnel.

This lands in channels, channel_destination, channel_read and channel_cross in the dossier, and renders as the report's first diagram. It also tells you which entry point to walk first in Step 5.

Step 5 — Walk each funnel (the core of this skill)

Follow references/funnel-walk.md exactly. In brief: start where a stranger actually lands (the URL a channel links to, or an ad's landing page — usually not the homepage), follow the primary CTA at each hop, and at every page capture the URL, its role, the headline and CTA verbatim, prices, proof elements, urgency mechanics, form friction, and a screenshot.

The chain is the observed path, not the site map. A page enters the chain only when you can name what routes traffic to it — record that in from. If the only way to reach a page is the navigation menu, it is not a funnel step: put it in other_doors instead. And expect parallel funnels — organic and paid usually differ, and paid traffic often skips the free layer entirely; give each its own branch rather than flattening them into one line. The checkout is the last stop — capture it fully. Mine sitemap.xml, robots.txt, and the Wayback Machine for the pages the funnel hides (thank-you pages, upsells, old pricing).

On the money pages (primary opt-in, core sales page, checkout), run the structured copy analysis in references/copy-teardown.md — awareness match, hook type, offer presentation, risk reversal, urgency type, proof density, CTA read. It's what justifies the CONVERT score, and its Steal: lines feed the synthesis.

Walks run in the main loop (the browser is one shared surface), one target at a time — recon fans out, walking doesn't.

Step 6 — X-ray the machine

For each key page in the walk (entrance, opt-in, sales page, checkout), fetch the raw page source and match it against references/tech-stack.md — funnel builders, email platforms, checkout processors, schedulers, DM automation, and every tracking pixel. Record what was found and what it means strategically (the reference has the meaning table). Pixels are a big tell: a Meta pixel + long-running ads = paid acquisition; GA4 alone + no pixels = organic-only.

Step 7 — Ads intelligence

Follow references/ads-intel.md: check the Meta Ad Library, Google Ads Transparency Center, and (where reachable) the TikTok library for each target.

Scope Meta to the advertiser, never to a keyword. A keyword search matches text inside ads and returns unrelated advertisers. Use the ad-library typeahead to select the advertiser (its rows show the verified Facebook/Instagram handles and follower counts — use the handles to disambiguate common names), then keep the resulting view_all_page_id and store it as meta_page_id. Record: are they running ads at all, how many are active, how long the oldest running ad has been live (an ad running 90+ days is a proven winner — say so), and the ads' destination URLs. Feed those destination URLs back into Step 5 — an ad's landing page is a funnel entrance the bio link never shows you. If an ad library is unreachable from this environment, report that honestly instead of guessing.

Step 8 — Offer ladder and funnel math

Assemble every offer found into a ladder: free → low-ticket → core → high-ticket/continuity, each rung with its price and where in the walk it appears. Missing rungs are findings ("no low-ticket bridge — they jump $0 → $2,000").

Then, only where the inputs exist, sketch the funnel math: audience or traffic × benchmark rates × price = an order-of-magnitude revenue band. Use the ranges and the honesty rules in references/benchmarks.md — compute with full ranges (never midpoints), name every benchmark used, tag the result INFERRED (modelled), prefer any real observed number over a benchmark, and skip the model entirely when more than two links in the chain are unknown. Never decorate a guess with math.

Step 9 — Score the four gates

Four gates, 0–25 each, totalling 0–100. These are only the things an outsider can actually verify by looking. Email nurture is deliberately not scored: what a brand sends its subscribers cannot be observed from outside, and a score you cannot evidence is a guess wearing a number. Note in the gaps that it wasn't scored, and move on.

This is a rubric-scored qualitative read — always label it as such, never as a measurement. Interpolate between the anchors.

• ATTRACT (getting found) — 0: no discoverable traffic engine. 12:

consistent output on one platform, or active ads. 25: multi-platform consistency plus long-running ads and an owned audience feeding the top.

• CAPTURE (getting your details) — 0: content leads nowhere, no capture

mechanism at all. 12: a lead magnet exists but is buried or high-friction. 25: dedicated low-friction opt-in with a specific promise and proof, or comment-to-DM automation feeding it.

• CONVERT (making the sale) — 0: no offer page found. 12: offer page with a

price but thin proof. 25: specific promise, stacked proof, a real guarantee or clear qualification, honest urgency, and a checkout whose friction matches the price.

• ASCEND (selling bigger) — 0: single-product dead end. 12: two rungs.

25: full ladder free → low → core → high/continuity with visible cross-sells, order bumps, or a continuity offer.

The weakest gate gets named in every deliverable. Against a competitor it's the user's opening; on the user's own funnel it's the leak to fix first. Record a one-line justification per gate in gate_notes — that's what makes the score defensible.

Step 10 — Synthesize

Across targets:

• Funnel patterns — where the machines converge ("all three run short-form

→ comment-to-DM → Skool").

• Divergence — where they split, and who's alone in what.
• Best-Engineered ranking — ordered by scorecard total; break ties with

momentum from the quantitative snapshot. This ranking is about the machine, not the brand's size.

• Your Opening (anchored runs) — the gap between the user's funnel/offer

and the field, stated in one or two sentences.

• Steal This Week (anchored runs) — exactly three moves, each copyable in

under a week, each traceable to something observed in a walk ("add an order bump at checkout — X does this at $27 and you have nothing between $0 and $497").

• Gaps — anything unverifiable, named honestly.

Mirror mode — the user's own funnel

When one of the targets is the user's own funnel (they'll say so, or the anchor domain appears in the target list), run the identical pipeline on it — same walk, same copy teardown, same rubric; no softened scores. Two things change in the output:

• The weakest gate is reported as the leak, and the dossier ends with a

Fix This Week list instead of Steal This Week: three moves, ordered by gate impact, each pointing at a specific observed page ("your opt-in has 5 fields; every walked competitor has 1").

• If competitors were walked in the same run, add a **gate-by-gate mirror

table** (user vs. field) — that comparison writes the user's roadmap, and it's the single highest-value output this skill can produce.

Some of the user's own stages are invisible from outside (their email sequence, their DM flows). Those aren't scored for anyone — say so plainly rather than guessing at them.

Re-spy — the diff over time

When a previous run's dossiers exist for a target (or the user asks "what changed?"), deliver the change report from references/re-spy.md instead of a fresh snapshot: price moves, new/removed funnel steps, stack additions, ads turned on or off, gate drift — each delta with both ends evidenced, plus the strategic read on what they're testing. A target with no previous run can still get a one-off drift read via Wayback snapshots of the money pages.

Step 11 — De

…

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Mahanaicoach
• Source: Mahanaicoach/funnel-spy
• License: MIT

Install and usage instructions live in the source repository linked above.