You asked for an exact plan for each of the four picks, and proof that we know how to deliver it. You also suspected the Apify openings were already taken. We checked all four.
Two of four go ahead. You were right about Apify: both niches are taken. AI search: prove it on ourselves for 6 weeks, then sell. DM setter: build in 2 weeks, pilots in week 3.
The engines' own docs explain the mechanism. The main lever (being named on other sites' pages) moved answers within weeks in controlled tests. Nobody has independently tested a full service, so we test one on ourselves first.
18 copies came after the last winner and share 3.7% of users; the best has 27. The broken official tool keeps growing anyway.
23 copies came after the June winner and share 11%; the best (74 users) launched one day after it.
Legal path through ManyChat (already Meta-approved) with our own AI writing replies. The unproven part is landing the first 3 pilots, not the build.
What changed since the first page. The four ◎ picks become two lanes, run in parallel: - AI-search visibility (prove for 6 weeks, then sell); - the AI DM setter (pilots from week 3).
The Apify picks are dropped. A daily store scan could catch rare openings, but the data puts it at $1–4k a month at best against shops running hundreds of tools, so it is parked.
Why prove first. You said you'd accept a service you don't know yourself only if we have a strategy and can prove it works. The strategy is below with its evidence ranked. The proof is a 6-week test on our own brand plus one past client, with pass/fail rules fixed in advance.
ChatGPT, Perplexity, Google's AI and Claude search the web live and quote what they find. So the brand has to be on the pages they find. That is the whole mechanism.
"Best AI visibility agency for UK Shopify stores?"
We pick 40–60 such questions per clientIt splits the question into several searches ("fan-out") in its own index: OpenAI's crawler, Bing, Google, Brave.
Let the search bots in; rank for the sub-questionsIt reads dozens of pages and leans heavily on third-party pages: "best X" lists, press, reviews, YouTube, Reddit, Wikipedia.
Get the brand named on those pagesIt names the brands those pages agree on. The list changes from run to run.
Measure the share of answers, never one answerSo what AEO really is. About 70% normal SEO and digital PR, plus measuring AI answers. AEO Engine's own monthly delivery shows the same: 30–60 articles, 3–10 backlinks, a PR campaign, 3–8 community mentions. Its guarantee is on organic traffic, not on AI visibility.
Two speeds. - Answers built from a live search can change in days to weeks. A BBC reporter changed what three engines said with one blog post within 24 hours, on a question nobody else competed for. - Answers from the model's memory change only with the next model release, which takes months. About half of ChatGPT's buying questions are answered without a search (third-party estimate).
Measuring honestly. SparkToro found fewer than 1 in 100 runs return the same brand list. So we report the share of answers that name the brand across many questions, each asked several times, never a single "rank".
Every lever, ranked by how strong the evidence is. We sell only the top ones. Schema and llms.txt get sold a lot and do nothing.
| # | Lever | What we do | Evidence | Shows up in | Engines |
|---|---|---|---|---|---|
| 01 | Let the AI search bots in | Allow OAI-SearchBot, PerplexityBot, Claude-SearchBot, Googlebot, Bingbot; check the CDN doesn't block them; submit sitemaps to Google and Bing | Proven (engine docs or experiment) — Engine docs: blocked sites "will not be shown in ChatGPT search answers" | ~24 h to days | ChatGPT, Perplexity, Google AIO/AI Mode, Gemini, Claude, Copilot |
| 02 | Get named on other sites' "best X" lists, press, reviews | Pitch inclusion in existing roundups; answer journalist requests; real reviews on niche directories | Proven (engine docs or experiment) — Brand mentions on the web correlate 0.66 with AI visibility vs 0.22 for backlinks (Ahrefs, 75k brands); in a 30-day controlled test 86% of mentions came via third-party lists | 1 day–4 weeks (narrow); months (memory) | All; strongest Perplexity, Gemini/AIO |
| 03 | Rank for the questions' sub-searches | Normal SEO for pages that answer the sub-questions; titles that match them | Strong correlation — Engines only cite what they retrieve; cited titles match the sub-searches (Ahrefs, 1.4M ChatGPT prompts) | weeks–months | All |
| 04 | One exact "facts" page | Prices, locations, numbers, founding — exact and current | Proven (engine docs or experiment) — In a made-up-brand experiment ChatGPT used the official FAQ 84% of the time (Ahrefs) | weeks | ChatGPT strongest, Gemini, Perplexity |
| 05 | YouTube mentions | 1–2 screen-recorded explainers a month, brand in title and transcript | Strong correlation — Highest correlation measured: 0.74 (Ahrefs, 75k brands); no experiment yet | unknown | ChatGPT, Google AIO, AI Mode |
| 06 | Keep key pages fresh | Monthly refresh with new data and a visible date | Strong correlation — AI-cited pages are 26% fresher than organic results (Ahrefs, 17M citations) | days–weeks | ChatGPT most, Perplexity, Gemini |
| 07 | Wikipedia (only if notable) | Disclosed talk-page requests only — never paid editing | Strong correlation — 48% of ChatGPT's top-10 cited-domain share (Profound) | months | ChatGPT |
| 08 | Stats, quotes and sources in the text | Real numbers, expert quotes, references | Weak / mixed — +22–42% in the GEO lab paper (KDD 2024), but an independent benchmark (NeurIPS 2025) found most such tricks don't work | days–weeks | Perplexity-type retrieval (lab) |
| 09 | Our own "best X" page | Honest comparisons including competitors | Weak / mixed — Vendor cases (Ramp 3% → 22% in a month), but Google's AI now leaves the publishing brand out 83% of the time | ~1 month | ChatGPT, Perplexity, AIO (declining) |
| 10 | Reddit / Quora | Disclosed founder answers to real threads; one account | Weak / mixed — Big in Perplexity; collapsed twice in ChatGPT (Sep 2025, Aug 2026) | days; unstable | Perplexity, Google AIO |
| 11 | Backlinks | Normal SEO hygiene only | Weak / mixed — Weak correlation (0.22–0.33) | months | All |
| 12 | Schema markup | Don't sell it as AI work | No evidence — Before/after test with control pages: no lift in ChatGPT, slight drop in Google's AI (Ahrefs) | no effect | AIO, AI Mode, ChatGPT |
| 13 | llms.txt | Don't sell it | No evidence — No correlation across ~300k domains; 97% of files never requested | none | All |
| 14 | Keyword stuffing | Never | No evidence — −9% to −10% in the GEO paper | negative | Perplexity |
| 15 | Hidden text aimed at AI | Never — banned | No evidence — Works in the lab, breaks Google's spam rules (updated Aug 2026) | banned | Lab, Perplexity |
What "causal" means here. Bot access is a hard gate written in the engines' docs. Earned mentions are marked causal because small controlled tests (a made-up brand, a zero-start brand) showed mentions on third-party pages pulling the brand into answers within weeks. The big studies behind most other rows are correlations, mostly run by tool vendors.
Vendor case studies (self-reported): - Ramp via Profound: 3.2% → 22.2% visibility in ~1 month. - Gruns via AthenaHQ: 2.0% → 12.6% share of voice in 60 days. - Heyflow via Peec AI: 10% → 40% within weeks. - AEO Engine clients: organic traffic ×3–9 over 4–12 months (traffic, not AI visibility).
Independent checks: - Ahrefs schema test: no lift. - C-SEO Bench, NeurIPS 2025: most AI-specific rewriting is ineffective. - GEO paper: the lab gains. - SparkToro: answers are inconsistent between runs. - Chen et al. 2025: engines favour earned media.
Never do: - fake Reddit accounts or bought votes; - fake or incentivised reviews (FTC rule; UK DMCC Act fines up to 10% of turnover, and it covers agencies that offer this); - pages shown only to AI bots, or mass AI-generated pages (Google spam policies).
AEO Engine uses "parasite SEO", which breaks Google's rules. We don't copy that part.
Copied from AEO Engine's own deliverables list, minus the parts that break Google's or your rules. AI does ~85%. You spend 4–6 hours per client per month.
| When | What the client gets | Done by |
|---|---|---|
| Day 0–7 | Onboarding: read-only Google Search Console, GA4 and Bing access, plus a draft-only CMS role (we write, they publish). Brand facts, banned claims, 3–5 competitors. | AI + you |
| Weekly | AI-answer tracking: 150 buyer questions × 4 engines × 3 runs. Share of answers naming the brand vs rivals, and which pages the engines cite. | AI |
| Weekly | Written report (no calls, no camera). A one-pager each month and a review on day 90. | AI |
| Monthly | Up to 30 pieces: answer pages, honest comparison and "best X" pages, FAQ hubs, refreshes. Each carries a real fact from the client. | AI + you |
| Monthly | 3–5 earned mentions: journalist requests, inclusion in existing listicles and directories the engines cite. A press release each quarter. | AI + you |
| Monthly | 3–5 brand mentions from the client's own disclosed accounts (LinkedIn, Quora, Reddit where the subreddit allows). | AI + you |
| Once + check | Crawler access, sitemaps, IndexNow, Bing Webmaster Tools. Entity and review profiles (Google Business Profile, G2/Trustpilot); the client sends review invites to all customers, never paid. | AI |
| Term | 90-day initial term, billed monthly by SWIFT/Payoneer invoice. Guarantee on a number we measure (see below). | you |
| Tracking option | Type | Price | Engines | Cost per client / mo | Pay from Ukraine |
|---|---|---|---|---|---|
| Build: DataForSEO app scraper + Perplexity + Google AI SERP our pick | build | $0.0012 per ChatGPT/Gemini answer; $0.0012 AI Overview; $0.0024 AI Mode | ChatGPT, Gemini, Perplexity, Google AIO + AI Mode | $7–9 | yes |
| Build: official model APIs only | build | OpenAI web search $10–25 per 1k; Gemini grounding $14 per 1k | ChatGPT-like, Gemini-like | $14–25 | yes |
| Rank Prompt (agency) | buy | Agency $119 · Agency+ $239 (white-label) | ChatGPT, AIO/AI Mode, Perplexity, Gemini, Claude, Grok | $52–155 | yes |
| Otterly.AI | buy | $189 (150 agency prompts) · $489 (500) | ChatGPT, AIO, Perplexity, Copilot (+add-ons) | $51–69 | yes |
| Peec AI | buy | €85 / €205 / €425 | 3 of ChatGPT, AI Mode, AIO, Copilot, Gemini | €70–115 | yes |
| Semrush AI Visibility | buy | $99 + $60 per 50 prompts | AI Mode, ChatGPT | $100–160 | yes |
| Bing Webmaster Tools AI Performance | free | free | Copilot (cited pages + queries) | $0 | yes |
What we copy from AEO Engine: tiers, deliverables list, 7-day onboarding, 90-day term, weekly written report, free report, guarantee.
What we change, and why: - "Quality backlinks" become earned mentions: buying links for ranking breaks Google's link-spam policy. - "Parasite SEO" is dropped: it is site-reputation abuse under Google's rules. - Reddit and reviews run only through the client's own disclosed accounts and invites to all customers. Fake or incentivised reviews are illegal under UK DMCC Act and FTC rules. - Schema and llms.txt stay as hygiene, never sold as AI work.
Guarantee. AEO Engine promises 30% more organic traffic in 90 days. We promise a number we measure ourselves: "named in more of your tracked buyer questions within 90 days, or month 4 is free". Working a month free costs only ~$170–470 plus hours.
The free snapshot (~$0.30 per prospect), copied from AEO Engine's free report: - 10 buyer questions × 4 engines; - which rivals get named instead, and which sites the engines cite for them; - a quick bot-access check.
It goes inside the first cold email.
The $500 audit (credited to month 1): - 100 questions × 5 engines × 3 runs; - share of voice vs 5 rivals; - the 30–50 sites the engines cite (the PR target list); - a 500-page crawl and a 90-day plan.
Sales tooling: ~$400–600 a month shared across all clients (cold-email tool, lead credits, snapshots, domains).
Tracker caveat. DataForSEO's app scraper reads the ChatGPT and Gemini apps; the scraping is its service under its own terms. We don't scrape ChatGPT ourselves. API-only tracking overlaps with what app users see on only 21–32% of brands (Surfer, Aug 2026).
Access stays low-trust. Read-only Search Console, GA4 and Bing access; a draft-only CMS role; no DNS or CDN admin. The client's developer applies our robots.txt change.
A 6-week test on two subjects before we sell a single retainer: our own new agency brand, and one past client who agrees. Rules are fixed before we start.
We don't sell "AI visibility" until it has moved for us, measured against questions we left untouched.
Sell the retainer if target questions rise ≥10 points more than the untouched control questions (95% confidence), on at least 2 of 4 engines, and it holds through the freeze week. If only our own brand passes, sell only in narrow niches.
Don't sell "AI visibility" if neither subject passes. Fall back to selling only the measurement (monthly AI-answer report) or drop the lane.
Subjects. - S1: our own new agency brand, on narrow questions we can own, e.g. "AI visibility audit for UK Shopify stores" or "ChatGPT visibility for Bali villa managers". - S2: one past client who agrees in writing. Never named publicly without consent.
Size. 60 questions × 4 engines × 5 runs = 1,200 answers per subject per round, about 21,600 answers over ~9 rounds. That detects a change of ~4–10 points. Two baseline rounds set the noise floor; if they differ by 5 points or more, we add a third.
Statistics. Target-minus-control lift with a bootstrap 95% confidence interval and a paired test across questions.
How we collect. From the engines' normal apps, not their APIs: API answers overlap with app answers on only 15–24% of brands. We use a tracker tool that does this, not our own scraper of ChatGPT. Claude is checked by hand only, on the subscription, never through the API.
What we publish. Method, the locked question list, before/after per engine with confidence intervals, a dated action log, and what didn't move. The same format becomes the free audit for prospects.
Cost of the test. About 21,600 answers at ~$0.001–0.004 each ≈ $25–90 on the build-A tracker, plus content and PR time. The tracker we build for the test is the same one we sell with.
Your doubt, tested on store data: 42,487 tools, with creation dates for 3,402. Short answer: yes, for both tools we named.
Both store niches are taken by fast copiers. Late entrants got a median of 8 monthly users.
18 tools came after the last winner (7 Jul 2026). Together they hold 225 of 6,008 monthly users (3.7%); the best has 27. And the broken official tool kept growing (2,452 → 2,831 users in a week at 55% success): users don't leave a broken leader.
23 tools came after the winner (11 Jun 2026). They hold 424 of 3,735 users (11%); the best (74 users) launched one day after the winner. Since August: 17 entrants, best 49 users.
Only 6 of 249 tools with 500+ users are below 80% success, and in one week none newly broke. The real wins are rising waves (Taobao copies: 929 users in 15 days), and shops running 100–465 tools already chase those. Ceiling ≈ $1–4k a month.
Each dot is one tool (colour = share of runs that succeed: green ≥95%, yellow 80–95%, red <80%). The official tool is 6.8 years old and is drawn at the right edge. Everything in the yellow band came after the 7 Jul 2026 winner, and none got past 27 users. Data: Apify store API, 27 Oct 2026.
The winner launched 11 Jun 2026; 23 tools came after it and hold 11% of users combined.
Bars are scaled for reading, not to a common axis. Store-wide numerators are exact (every tool with 50+ users has a creation date); the ~46,000 denominator is an estimate from a 1-in-10 sample of the newest list. At ~$2.5 net per monthly user, 8 users ≈ $20 a month.
| Tool | Monthly users | Runs succeed | Good alternatives | Opening |
|---|---|---|---|---|
| lukaskrivka/google-sheets | 841 | 45% | 0 | maybe — check if failures are users' own sign-in |
| apify/google-trends-scraper | 2831 | 55% | 2 | taken |
| pizani/taobao-product-scraper | 2736 | 68% | 6 | taken |
| disarming_screw-owner/1-2-000-leads-apollo-scraper-replacement-b2b-b2c-leads | 577 | 70% | 10 | taken |
| zerobreak/google-hotels-scraper | 523 | 76% | 0 | thin |
Only one has no reliable alternative (Google Sheets import/export), and its failures may come from users' own Google sign-in, not the tool.
Our recommendation. Drop both Google tools. Park the daily scan: at $1–4k a month at best, racing shops that run 100–465 tools, it isn't worth your focus now. Revisit only if the two sold lanes stall.
Limits of the data. - The store API stops paging around 12–15k results, and its totals vary between calls. - Break and recovery rates come from one week (20 → 27 Oct). - Ages "at #1" are upper bounds.
Raw data: `research_notes/Copy what works/apify_data/`.
Yes, we know how to build it, and it's legal. Build on ManyChat, which Meta has already approved, so we skip Meta's app review. Our own AI writes the replies.
Build: 2 weeks. Cost per client: ~$123 of $490 (75% margin). The hard part is landing the first 3 pilots, not the tech.
| Way to build it | How | Cost per client / mo | Meta app review needed? | |
|---|---|---|---|---|
| ManyChat Pro per client + our AI | Coach connects Instagram + WhatsApp by login button (no passwords). Every DM goes to our AI by webhook; the reply comes back through ManyChat. | $39 | No — ManyChat is already approved | our pick |
| GoHighLevel (if the coach already uses it) | Plug our AI into the coach's own GHL workflow; native calendars | $50–97 | No | |
| Chatfuel / SendPulse | Bot flows calling our AI over an API | $9–18 | No | |
| Resell DM Champ white-label | Their AI under our brand: fastest, but we own nothing | ~$30 | No | |
| Our own Meta app (month 2–3) | Instagram API + WhatsApp Cloud API as a tech provider; filed in week 6 | ~$10 | Yes — Business Verification + App Review |
AI may reply and follow up within 24 h of the person's last message (7 days after a click-to-DM ad). We never cold-message from the coach's account.
Only for real humans. Using it for AI breaks Meta policy; our AI never uses it.
One private reply per comment, within 7 days ("comment GUIDE and I'll DM you").
Meta's Jan 2026 ban hits general-purpose assistants like ChatGPT. A business's own lead-qualifying and booking bot is allowed, with a clear path to a human.
EU AI Act Art. 50 transparency has applied since 2 Aug 2026; Meta also asks for bot disclosure. It may lower replies, so we measure it in the pilot.
We process the leads' DMs for the coach. Ukraine has no EU adequacy decision, so each client signs a DPA + standard contractual clauses (UK addendum for UK). 90-day retention.
The AI that answers. - OpenAI's small model writes most replies (~$0.0026 each); hard turns go to a bigger model. - About $20 a month per client at 400 new conversations. - The Anthropic API is not used, per your billing rule. You approve a $50/month hard cap on the OpenAI account for the pilots.
Copied from the original (likely Upster: €290 a month plus credits, or €990 unlimited; inferred from TrustMRR): - answers in the coach's voice, trained on their own Instagram data export; - answers only from their offer document, with a blocklist of claims (prices, health or income promises); - qualifies the lead, books the call in their calendar, and follows up at 4 h and 23 h; - hands over to the coach when needed; - weekly report: response time, qualified leads, booked calls, show rate.
Why it lifts results. HBR (2011, 2,241 firms): answering within an hour made qualifying a lead ~7× more likely than an hour later, and 60× more likely than after 24 hours. 23% of firms never answered. All AI-setter before/after numbers are vendors' own, e.g. Upster: 3.4% booking rate on Instagram vs 22.1% on WhatsApp, and 52% of bookings come from follow-ups.
Costs that can surprise. WhatsApp: only the first 1,000 service messages per number per month are free since 1 Oct 2026, then about $0.022 (UK) to $0.055 (DE) each. A heavy client costs ~$234 a month (52% margin).
2–3 coaches in the UK, Ireland or Germany who get 30+ new DMs a week and sell a €1k+ offer. Free for 14 days, then $490 a month.
Each coach's own Instagram export is the baseline. The pilot must beat their own numbers.
Scale it if: median first reply ≤ 2 min, unanswered after 24 h ≤ 2%, booked calls ≥ 1.5× baseline and ≥ 5 extra calls per coach in 14 days, show rate ≥ 60%, the coach would send ≥ 90% of replies as written, zero Meta warnings, and 2 of 3 coaches start paying.
Stop or re-aim if booked calls are under 1.2× baseline, any false claim or Meta warning, or 0–1 coaches pay. Then point the same AI at WhatsApp leads for Bali/Phuket villa managers through your network, or fold it into the studio offer. No ads.
Pilot cost: under $200 (3 × ManyChat $39 + ~$30 AI + ~$20 hosting), inside the cap you approve.
Honest risk. Instagram cold DMs can't be automated legally, so someone sends ~25 a day by hand: ~30 minutes for you, or a VA. Cold email to UK/IE coaches gets 0.2–0.6% positive replies. Nobody has published a reply rate for cold Instagram DMs to coaches.
Not confirmed yet: - whether Meta's Business Verification accepts a Ukrainian FOP (the ManyChat path doesn't need it); - ManyChat's exact plan features (its pricing page blocked our reader).
Two lanes in parallel, with a go/stop gate for each. Your recurring time: ~30 minutes a day in weeks 2–3 (sending DMs by hand), plus signing.
≥300 touches and 3 coaches signed. Miss → aim the same AI at WhatsApp leads for Bali/Phuket villa managers through your network.
Pilots beat their own baseline and 2 coaches pay $490. Miss → stop the DM lane.
Target questions up ≥10 points vs control and it holds through the freeze. Pass → sell retainers. Only our brand → narrow niches only. Fail → don't sell AI visibility.
≥$2k a month signed across both lanes (e.g. 1 retainer + 1–2 DM clients). Miss → back to the board for the next copy.
Why parallel. The AI-search lane can't sell until week 6, so the DM lane is the earlier money (paying clients from week 5 if pilots pass). Both use the same AI-run back office and the same invoicing to your FOP.
What could break the plan: - no coach agrees to a pilot (manual DMs, unknown reply rate); - the AI-search test shows no lift beyond the control questions; - Meta Business Verification refuses a Ukrainian FOP. The ManyChat path doesn't need it.
Every link on this page, re-checked when it was built.
Evidence labels. Engine docs (OpenAI, Google, Perplexity, Anthropic, Meta) are primary. Ahrefs, Profound, SE Ranking and Seer studies are large vendor studies (correlation). GEO (KDD 2024) and C-SEO Bench (NeurIPS 2025) are peer-reviewed. Every before/after case study is the vendor's own claim.
Apify numbers come from the public store API on 27 Oct 2026 (42,487 tools crawled; raw data in research_notes/Copy what works/apify_data/). Revenue per user (~$2.5 net a month) is an estimate from Apify's own payout totals.
Not checked: ManyChat's exact plan features (its pricing page blocked our reader; third-party pages used), Meta's official app-review timing, whether Meta Business Verification accepts a Ukrainian FOP, Peec and Profound prices (third-party), incoming SWIFT/Payoneer fees.