Learn from others

1Summary

• One paid product, six free satellites. Pancake charges only for the conversation inbox, priced per page and per staff seat from about 92,000đ a month. POS, chatbot, landing-page builder, team chat, CRM and accounting are free. The free products keep orders, customers and money inside the same account, which makes the paid inbox hard to leave.
• The moat is the closed loop between ads and orders. Every order created in a chat is sent back to Meta and Google as a conversion signal, so a seller's ads get better the longer they stay on Pancake. Meta named Pancake its Vietnam messaging partner of the year two years running.
• Scale came from a free tool, not from marketing. pages.fm was a free fanpage utility in 2016 and 2017. The legal entity was only registered in December 2019, after the product already had its user base.
• The AI strategy is assistant first, agent second. The AI assistant only suggests replies to staff; Botcake AI screens conversations and hands over when purchase intent appears. The chief executive frames this as cost: clients report 60 to 70 percent of marketing budget going into messaging, and link-based order closing cutting operations cost by 70 to 80 percent.
• For FLYER the most transferable ideas are conversion signals fed back to ads from the CRM, a single account and wallet across the product family, per-unit pricing that grows with the customer, and public developer documentation. As a Pancake customer FLYER is under-using the platform: default retail tags, a disconnected page, a median first reply of 161 minutes on the Teado page, and a bot flow that rejects one in five phone numbers.

2The company in numbers

Timeline

3The ecosystem, product by product

Seven products share one Pancake Account, one prepaid wallet and one customer record. The paid inbox sits in the middle; everything else exists to pull data into it or to push data out to the ad platforms.

Two more layers matter. The platform layer is WhatsApp Business API resold with Meta's per-message fees passed through, plus an Open API and webhooks for POS and the inbox, documented at docs.pancake.biz with a machine-readable OpenAPI file. The community layer is a Facebook group for users, a public changelog, success stories, and yearly summits co-hosted with Meta.

4How Pancake instruments the journey

Pancake has no general-purpose analytics event bus and no session tracking. Instead it instruments the four moments where money moves, and it reuses one object's state machine as the event source for everything else. Understanding that choice is the most transferable part of the whole platform.

The four moments, and the key that carries across them

The join key between the conversation world and the order world is the phone number. Pancake detects it in the message text, and an optional setting creates the POS order automatically the moment a message contains one. That single toggle is what turns a chat into a tracked journey without anyone filling in a form.

One state change, many subscribers

Nothing in Pancake fires an event by being logged. Events fire when an order changes status. The same transition is consumed by five or six different subsystems, each configured separately, which is why the platform feels automated without having a workflow builder at its centre.

The statuses that carry side effects are worth naming, because the seller picks one of them as the trigger for the purchase event.

The event catalogue

Attribution rules that decide the numbers

• Seven days from the click. An order counts as coming from an ad only if it appears within seven days of the click. The report separates new ad-sourced orders from processed ones, and the close rate is processed divided by new.
• Conversion campaigns need a conversation. For a conversion campaign, Pancake can only report a Meta purchase if the customer also opened a conversation with the page. Traffic that converts on a landing page without messaging is invisible to this column, which is why Pancake tells sellers to route people into Messenger.
• Offline events are the cross-check. Because in-store events fire without any conversation, sellers are told to compare them against the online columns to catch mapping mistakes and double counting.
• One status maps to one event. A status cannot carry two events, and after creating a row only the event type and value can be edited. This is a deliberate guard against double counting.
• Zero-value orders get a default price. A gift or zero-price order would otherwise send revenue of nothing and corrupt return on ad spend, so the configuration requires a fallback value.
• The dataset belongs to the page owner. Events land in the dataset on the business manager that owns the page, not on a partner's business manager, which is the most common reason a seller sees no data.
• Platform numbers will not match. Pancake states plainly that its conversation counts differ from TikTok's and Meta's because the counting methods differ, and its campaign report shows platform rows and Pancake rows in different colours rather than reconciling them.

Two details show how much of this is handled for the customer rather than by the customer. Pancake and Meta migrated every existing account onto the conversions interface, and the connection is created automatically as soon as a page produces its first order. Most sellers never open the configuration screen at all, which is why the feature has near-universal adoption across 260,000 accounts.

Where the source is captured

Three mechanisms cover the entry points that ad platforms do not label by themselves. Trackable short links for Messenger, Instagram, WhatsApp, TikTok and Zalo carry a source name the admin chooses, with a pre-filled first message the customer has to send for the source to register. The website chat widget can raise a form-submission event for Tag Manager. Marketplace and website orders arrive with their own channel recorded. Everything else is inferred from the ad identifier Meta or TikTok attaches to the conversation.

What Pancake deliberately does not track

There is no page-view funnel, no session or path analysis, no cohort tool and no experiment framework. The reports answer a short list of operational questions: which campaign produced orders, which staff member replied, which tag is growing, when the peak hours are. A seller who wants deeper analysis is expected to take the webhook feed or the export and do it elsewhere.

The pattern FLYER can copy

The structure is what transfers, not the vocabulary. FLYER has no orders and no carriers, but it does have a state machine that every customer passes through, and it currently emits nothing when a customer advances.

The minimum useful version is three events, not thirty: lead qualified, trial activated, and paid, each carrying the phone number as the join key and the campaign identifier where one exists. That is enough to change how campaigns are optimised, and it is a week of work rather than a platform project.

5How they track and improve the product itself

The obvious assumption is that a company with this retention runs a heavy product-analytics stack. It does not appear to. On 19/09/2026 the login page, the point-of-sale app and the chatbot site were loaded in a real browser and every outbound request was recorded. Nothing product-analytical came back.

Limit of this check: only surfaces reachable without an account were observed, and first-party telemetry sent to Pancake's own domains would be indistinguishable from ordinary application traffic. The claim is narrow. No third-party product-analytics tool is loaded where a visitor can see it, and the two tags that do exist are marketing tags on marketing pages.

Why they can get away with it

A product-analytics tool exists to answer one question: is this account actually using the thing. Most software has to infer that from clicks, because the software itself holds no record of the customer's work. Pancake holds all of it. Conversations, orders, stock movements, customer records, staff assignments and payments are the product's own tables. Whether a shop is alive, growing or drifting away is a query, not a tracking plan.

That single fact reshapes everything else, and it is the mechanism behind the retention.

Step three is where the retention is actually made. Everything above it is distribution; everything below it is consequence.

What they measure, and who it is for

The reports inside Pancake are not about Pancake. Every one of them measures the shop's own operation, which is why the owner opens the tool daily even when no customer is waiting.

Notice how carefully two of those definitions are drawn. A phone number is credited only when a staff member's reply produced it. A response-time average ignores the hours the shop is closed. Those rules exist because the numbers are used to pay people, and a metric that decides pay gets scrutinised in a way a dashboard metric never does. That scrutiny is free product feedback.

How the product actually improves

With no analytics stack, the input has to come from somewhere. It comes from support and from a public user group, and it turns into small releases at a rate most teams would find uncomfortable.

The changelog is public and dated. In September 2026 up to the 18th there were nine release days carrying nineteen items; August had ten days and fourteen items; July had six days and fourteen. Two to three release days a week, every week, each one a small named improvement rather than a quarterly feature. Nothing on that list would appear on a roadmap. All of it removes a specific annoyance somebody complained about.

The levers that hold customers, ranked by how hard they are to copy

What FLYER should take, and what it should not

Two of these transfer directly, one does not.

• Take the health query instead of a tracking plan, where the data already exists. FLYER knows whether a teacher assigned work, whether a class submitted, whether a school's students logged in this week. A teacher and school health score built from those tables answers the retention question without a single new event, and it can be built this quarter.
• Take the idea of metrics people are paid on. The reports FLYER shows a centre today describe children's progress. The ones a centre's owner would open daily describe their teachers: who assigns work, whose classes finish it, whose parents renew. That is the number that makes the tool part of how the centre is run.
• Do not skip product analytics. Pancake can live without it because its users are paid to be in the product all day and because money changes hands inside it. FLYER's learners are not paid to be there, usage is voluntary and the paying parent is not the user. Where behaviour is discretionary, clicks carry information the order table never will. The lesson is not to avoid instrumentation; it is to stop instrumenting what the database already knows and spend that effort where behaviour is genuinely invisible.

6How the money works

Plan prices from third-party listings, since pancake.vn renders them with JavaScript; the site itself advertises "from 3,000đ a day", which matches the Small plan.

Five design choices stand out.

1. Charge for the chokepoint only. The inbox is where the seller feels pain every hour; that is the one thing billed. POS, chatbot, landing pages, CRM and accounting are free, and each one moves more of the seller's data into the account.
2. Price on units the customer already counts. Pages and staff are how a shop owner already thinks about size. When the shop grows, the bill grows without a sales conversation. There is no per-message or per-conversation fee on the core product, so heavy use never feels punished.
3. Prepay six months. The default term is six months, which funds the free products and removes monthly churn decisions.
4. Usage fees only on marginal cost. WhatsApp template messages and OpenAI tokens are billed at cost through a prepaid wallet in USD. Pancake does not mark up the thing it cannot control.
5. Enterprise by contract, not by tier. Large accounts such as Yody or FPT Shop are served under custom plans and become success stories; the public price list stays simple for the 260,000 small sellers.

7Why it works

It owns the moment where Vietnamese commerce happens

Vietnam is Meta's leading market for business messaging in the region, and Messenger is the first inquiry channel during live streams. Pancake sits exactly at that moment, and its chief executive reports clients putting 60 to 70 percent of marketing budget into messaging. A tool that manages the moment where money is made is bought before any other tool.

Data gravity from free products

Every free satellite adds a table to the account: orders, stock, customers, staff tasks, books. Leaving the paid inbox means leaving all of it. This is the same play as Google Workspace or HubSpot's free CRM, executed at Vietnamese price points.

The loop back to the ad platform

Sending order events back to Meta and Google as conversion signals is the feature Pancake puts first on its POS page and the one Meta cites when it names Pancake and Haravan as partners it works with on "lower funnel signals". It aligns Pancake with the platform's own goals, which is why Pancake gets early access, summit stages and awards, and why sellers see their ad cost fall the longer they stay.

Free distribution before monetisation

The product was free for its first three years and the company was registered only after the user base existed. Sellers recommended pages.fm to each other in Facebook groups; the company still runs its main support in a Facebook group.

AI positioned as staff assistance, then screening

The inbox AI assistant suggests a reply, detects sentiment and lets staff accept with Tab and Enter. Botcake AI screens the first messages and hands over when intent appears. Mai Hồng Phong describes the outcome as staff stepping in "only when purchase intent becomes clear". The AI does not replace the seller's voice; it removes the waiting.

Shipping speed and documentation

The changelog is public, the docs cover eight products in Vietnamese and English, and the OpenAPI file carries a note telling AI agents to read the YAML instead of scraping HTML. That is an unusually modern developer stance for a Vietnamese SME tool and it made FLYER's own integration in September possible in a day.

8Where it falls short

• Built for retail, not for services or education. Default tags are "Kiểm hàng, Mua hàng, Hết hàng, Trả hàng". FLYER's pages still carry those. Loyalty, stock and carriers mean nothing to a school or a language centre; the CRM has no notion of a student, a class or a renewal date.
• Offline retail is weaker than KiotViet or Sapo. Third-party reviews rank Pancake POS first only when most revenue is online.
• Personal Zalo is a grey zone. Pancake connects staff's personal Zalo accounts by QR code. It works, and for FLYER it carries 76 percent of all conversations, but it runs outside Zalo's API and the account can be locked. The proper channel is Zalo OA, which Pancake also supports.
• Opaque pricing and single-page-app docs. The price list needs JavaScript to show numbers; the docs cannot be read without a browser. Both are minor, but they contradict the open-API stance.
• API rough edges. The conversation endpoint requires both since and until in unix seconds, silently returns nothing on ISO dates, rejects windows over one month, ignores page_number on v2, invalidates the page token when regenerated, and rate-limits early. All of this is recorded from FLYER's September pull.
• Customer counts are not audited. 260,000 businesses likely counts connected pages or accounts. Paid customers are not disclosed.

9Lessons for FLYER as a Pancake customer

FLYER already runs 23 channels through Pancake: 9 Facebook pages, 1 Zalo OA and 13 personal Zalo accounts. The August 2026 data pulled through the API shows the platform is doing the collecting but FLYER is not using the parts that make Pancake valuable.

None of this needs a new tool. It needs the platform FLYER already pays for to be configured the way its best customers configure it. The cs-flow page on funnel.flyer.vn already lays out the three-layer plan; this table is the order to do it in.

10Lessons for FLYER as a product company

These are the mechanisms behind Pancake's growth, restated as questions for FLYER's own product family: flyer.vn, teacher.flyer.us, Study, Exam, IELTS, Teado, Trường Số and the LMS.

Find the chokepoint and bill only there

Pancake bills the inbox because that is where a seller feels the hour-by-hour pain. For a teacher the equivalent is the assign-and-grade loop; for a parent it is the exam result; for a school it is renewal and reporting. FLYER prices per student per year, which is easy to compare with a textbook and hard to grow. The Pancake question is which single workflow the customer would pay for every month, and whether everything around it (the parent app, the library, the placement test) should be free to pull data toward that workflow.

Price on units the customer already counts

Pages and seats are legible to a shop owner. For a centre the legible units are classes and teachers, not students; for a school they are campuses and teachers. A per-class or per-teacher plan with a student allowance would grow with the customer without a renegotiation, the way an extra Pancake page does. This is worth a pricing study before the next B2B renewal season.

Send the signal back

Pancake's most defensible feature is the order event returned to Meta and Google. FLYER runs lead-form ads, then a CRM, then a trial, then a payment, and none of those later events reach the ad platform. Sending "trial activated" and "paid" from the CRM back to Meta as conversion events would let the campaigns optimise for buyers instead of form fills. The 09/09 analysis found 79 percent of phone numbers come from lead forms; a signal on the step after the form is the cheapest lever on cost per paying customer that FLYER has not pulled.

One account, one wallet, one customer record

Seven Pancake products share one login and one prepaid balance. FLYER's products carry separate accounts, separate admin panels and separate payment flows; the domain inventory lists dozens of hostnames. A single FLYER account with a wallet that any product can draw on (AI grading credits, exam attempts, Teado seats) is the structural change behind Pancake's "free satellites" strategy, and it is a prerequisite for the enterprise architecture work already in the PRD.

Free distribution before monetisation

pages.fm was free for three years and grew through seller groups. FLYER's free teacher tier is the equivalent and it already reaches most of the 1,131 active teachers who assign work. The Pancake lesson is to keep that tier generous, instrument it, and let the paid product be the thing teachers ask for rather than the thing they are blocked by.

Ride the platform's goals

Pancake's expansion into the Philippines, Thailand, Cambodia and Laos follows Meta's regional partner events rather than its own marketing spend. FLYER's Spanish-language and Malaysian market work would move faster attached to a platform with regional reach, whether Cambridge English, Google for Education or a national LMS programme, than as a standalone launch.

AI that assists first

Pancake ships the AI assistant as suggestions the staff member accepts, and only then as an agent that screens. FLYER's Trợ giảng AI and the planned Bingo Bot follow the same order: suggest to the teacher, screen for the CS team, hand over when intent or difficulty appears. The 60 to 70 percent budget figure Mai Hồng Phong quotes is a reminder that the AI's job is to shorten the wait, not to replace the person the customer wanted to talk to.

Public docs, public changelog, public case studies

Pancake documents eight products, publishes a changelog and lists nine success stories including an education company (ILA). FLYER's partner API for schools and the /whats-new page are steps in this direction. A public docs site for the LMS integration API and two or three named school case studies would cost a sprint and change how procurement conversations start.

11Proposed next steps

12Sources

• pancake.vn, pricing, features, success stories, About Pancake
• pos.pancake.vn, botcake.io and Botcake pricing, webcake.io, crm.pancake.vn, pancakework.vn, fintab.pancake.vn, docs.pancake.biz
• VietnamBiz, interview with Pancake and PMAX leadership, 08/2025
• VIR, Meta on business messaging in Vietnam
• Pancake on LinkedIn, TopCV company page, business registry
• Tân Hưng Hà, Pancake POS review 2026, Tôi Là Quản Trị, 2017 comparison with Page365, third-party price listings (Nhân Hòa, Onshop, DooPage)
• Pancake OpenAPI specification, webhook specification, and the product documentation for the inbox, POS, Botcake and CRM at docs.pancake.biz (conversion events, order statuses, ad reporting, tag settings)
• Observed 19/09/2026: third-party requests from pages.fm/login, pos.pancake.vn and botcake.io, and the dated entries in the Pancake changelog
• FLYER internal: Pancake API pulls of 09/09 and 13/09/2026 (23 channels, August 2026), the Meta Ads dashboard on funnel.flyer.vn, the cs-flow plan on funnel.flyer.vn