Hyppo blog

Fintech onboarding: where users drop off before the first payment

Nicolas Lew Deveali — Sat, 22 Aug 2026 00:00:00 GMT

In a fintech, onboarding does not fall apart where the team assumes. The usual suspect is the signup form, and that is almost never where the real problem sits: whoever made it through filling out their details already has intent. The stretch that gets lost quietly is the one that comes after, between identity verification approved and the first funded transaction.

Why that stretch stays invisible

Most product dashboards measure "registered users" and "active users" as if they were two consecutive steps. In between there are at least two events that almost never get their own name: KYC approval and the first funding load. Without that split, the report mixes someone who signed up yesterday with someone approved three weeks ago who still has not funded a dollar, both showing up as "pending" in the same column.

This matters because the already-approved user has spent the most expensive effort in the process: uploaded documents, waited for review, retried a photo that did not pass. Losing them there costs more than losing someone at the form, yet it gets less attention precisely because it is not measured as its own step.

The three events worth separating

Signup completed. Timestamp of when the user finished entering their details.
Verification approved. Timestamp of when KYC gave the green light, not when the user started it.
First transaction funded. The first real money movement, whether it is a top-up, a transfer or a payment.

Those three timestamps per user are enough to calculate the time between each step and see where each weekly cohort stalls. No more infrastructure than that is needed for the initial diagnosis.

Which message works best at each point

A welcome email at 24 hours treats every user the same, regardless of the step they are on. Against that baseline, a message triggered by the "verification approved" event, pointing to the exact next step (how to fund the account, which method fits the country), converts better because it answers the user's real moment instead of a fixed clock.

That architecture, where the stage lives as a profile attribute and only fires the message when it applies, is the same one covered in customer retention by lifecycle stage. At one cross-border payments fintech client, that approach placed 99.8% of users in a concrete stage, which made it possible to replace separate per-country campaigns with a single message engine.

Where to start

Before automating anything, split the three events in your user table and calculate how long it takes between verification approved and first transaction for the last four weeks. If that number keeps growing week over week, it is not a product problem: it means nobody is talking to the user at the exact moment they got stuck.

That diagnosis, and the flow built on top of it, is the work we do in Growth Automation, applying the same event-and-stage logic that turned Bigbox's setup into production in 15 days, later replicated across five countries.

Customer retention: the three-stage framework we use

Nicolas Lew Deveali — Sat, 22 Aug 2026 00:00:00 GMT

Most of the growth a company loses is not lost at acquisition. It is lost afterwards: in the onboarding nobody finishes, in the second purchase that never arrives, in the customer who leaves without saying anything and without complaining.

The usual answer to that is a reactivation campaign. It almost never works, and the reason is structural: it speaks the same way to people at very different points in their relationship with the business. Someone who never bought, someone who bought once and someone who bought six times have three different problems. One message cannot solve all three.

What we are measuring when we say retention

Retention is the share of customers who repeat whatever defines value in your business within a time window. The definition matters less than the two decisions it hides: which event counts as "repeating", and how long the window is.

In ecommerce the event is usually a purchase and the window is the product's natural repurchase cycle. In a subscription business it is the renewal. In a car dealership it can be a service visit, because the next car purchase is years away. Choosing the wrong event is what produces dashboards full of numbers nobody understands, and they end in the most expensive conversation any marketing team has: why does this number not match that one.

One clarification that saves time: in practice, churn is not the inverse of retention. Retention looks at who came back, churn looks at who is not going to. One is counted, the other is estimated, and confusing them is why so many retention reports read like a crystal ball.

The three lifecycle stages

This is the frame we use to segment a base before touching a single message. Three stages, defined by behaviour rather than by how long someone has been around.

Acquisition: zero purchases

People who left their details, registered, enquired or abandoned a cart, and never bought. They are not customers, they are prospects, and treating them as customers is the most common mistake in a badly segmented base.

What they need is not a promotion. They need the one piece of information they are missing in order to decide, and it is almost never the price. It is the specific doubt that never got answered.

Early lifecycle: between the first purchase and the convergence point

This is where most of the leaking money is. A customer who bought once has not chosen your brand yet: they tried it. The probability they leave at this stage is high, and it drops quickly with each subsequent purchase.

It is also the stage with the least automated work in most accounts we audit, because it belongs to nobody. Acquisition already counted them as a conversion, and the loyalty team does not consider them theirs yet.

Mature: from the convergence point onwards

The customer has a habit. Churn risk has settled at a low floor and the communication changes purpose: you stop convincing and start not annoying, raising frequency or basket size, and asking for the thing only a happy customer can give you, which is a referral or a review.

How to find your convergence point

The boundary between the early and the mature stage is not an industry number. It is yours, and it comes out of your data with a calculation that fits into an afternoon.

Group your customers by order number: those who made one purchase, two, three, and so on.
For each group, calculate the share that did not buy again within your repurchase window. That is your churn curve by order.
Find where the curve stops falling sharply and flattens for two consecutive periods. That order number is your convergence point.

What you are looking for is the moment the relationship stops being fragile. In most ecommerce accounts we work with, that point shows up earlier than the team expects, and always in the same place: it does not move month to month, which is what makes it a reliable boundary for segmenting.

Once you have the number, everything else becomes concrete. You are no longer deciding "who gets the loyalty campaign", you are deciding "what do we say to someone sitting on order two when convergence is at four".

What to automate at each stage

The order matters. Stage first, flow second. The other way round, automation only multiplies message volume.

In acquisition, short sequences aimed at the objection, not at the discount. Plus a cut-off rule: if after N contacts there has been no signal, that person leaves the active flow. A base inflated with people who will never buy ruins every metric you have and raises your platform bill.

In early lifecycle, this is where the best work belongs. Confirmation that the first purchase went well, the right complementary product at the right moment, and a reminder calculated on that product's real repurchase cycle rather than a generic thirty days.

In mature, less frequency and more value. Early access, recommendations that assume the person already knows the catalogue, and the review request that only makes sense here.

The condition for this to work is not the tool, it is that the stage lives on the customer profile as an attribute and updates itself. At Takenos that architecture reached 99.8% of users placed in a stage, and that is what made it possible to move from separate per-country campaigns to a single lifecycle engine.

The three measurement mistakes that ruin the diagnosis

Reading the average. A global retention rate goes up when a large cohort enters, even if every individual cohort is leaking faster than the one before. Read it by acquisition cohort, always, and look at the full curve rather than only the first month.

Counting the wrong base. If your customer table only holds people who bought, any conversion rate you calculate will be fantastic and false. This is the mistake that produces those reports showing 100% conversion that nobody questions.

Crediting the last campaign. The message that precedes the purchase takes credit for a process that started much earlier. It is useful for knowing which message closed; it is useless for deciding where to invest.

How to start next week

You do not need a six-month project or a CDP for the first step.

Define the event that counts as "repeating" in your business, and the window. Write it somewhere the whole team can see.
Calculate the churn curve by order number and find your convergence point.
Split the base into the three stages and look at how many people sit in each. The distribution is usually the first surprise.
Pick one stage, the early one, and build a flow. One.
Measure against a control group. Without a control you do not know whether it worked, you know the month was good.

That work is exactly what we do in Growth Automation Strategy, and at Bigbox the full setup was live in production in 15 days and then replicated across five countries on a single flow library. The slow part is never the tool: it is agreeing on what each stage means.

Answer engine optimization: how to get AI to cite your brand

Nicolas Lew Deveali — Fri, 12 Dec 2025 00:00:00 GMT

For two decades the goal was the click. We optimised for blue links, fought for the top position and watched CTR like a vital sign.

That changed. With Google's AI overviews, ChatGPT and Perplexity in the middle, a lot of people are no longer looking for a list of options. They are looking for an answer. And if your brand is not inside that answer, you do not lose a click: you lose the whole conversation, because you never made it to the table.

What answer engine optimization is

Classic SEO works to rank. Answer engine optimization works to be cited.

When someone asks an assistant "what is the best CRM for scaling a SaaS", the model does not scan keywords and return ten links. It synthesises information from sources it treats as reliable and builds an answer. AEO is the work of making your content one of those sources.

The practical difference: ranking means appearing among the options, being cited means being the recommendation.

Why it matters to whoever owns growth

If you run paid media, the reasonable first reaction is to ask what this has to do with your spend. It has to do with it on two fronts.

The first is where demand comes from. The research stage moved. It used to start in a search engine and continue across three tabs; today a good share of it starts and ends in a conversation with an assistant. That stretch has no impressions and no CPC, it shows up in no ad platform, and it is where the buyer builds the shortlist. If you arrive at retargeting, you arrive after the comparison is already done.

The second is brand recognition. Models favour entities with established authority. A brand that shows up cited in its category becomes familiar before the first ad, and that lifts the performance of everything that comes afterwards, paid included.

The five things that actually move it

You do not need to throw out the SEO playbook. You need to add five layers to it.

1. Verifiable authorship, not a decorative byline

Models, like Google, treat content better when it traces back to a specific person. That means a first and last name, a role, an author page of their own, and external profiles confirming that this person exists and works on what they claim.

What to do: sign every article with a real person, give that person a page on the site with their experience, and publish cases that prove you did the work instead of only describing it. Our Grupo Garden and Bigbox cases exist for that reason before any other.

2. Structured data on everything describable

Models are good at interpreting text and better when the text arrives declared. Schema is not a magic ranking signal, it is a translation: it tells the machine what each thing on the page is without relying on it inferring that from the layout.

What to do: mark up articles, FAQs, products and pricing. And one rule that is not negotiable: schema has to describe what the page actually shows. Structured data that contradicts the page is a guidelines violation, and the penalty reaches the whole domain.

3. The answer first, always

A model lifts fragments that stand on their own. A section that opens with three paragraphs of context and defines the term in the fourth is a section that cannot be cited without being cropped badly.

What to do: answer in the first 40 to 60 words of each section, then expand. It works for the engine and it works for the person who arrived with one specific question and does not want to read a thousand words to find it.

4. The entity gets built off your site

A model decides how much authority you have by looking at how the rest of the internet talks about you. It reads forums, directories, reviews, industry coverage. Your own site says what you want to be; everything else says what you are.

What to do: earn third-party mentions and reviews where your category is genuinely discussed. The reviews clients published on the Customer.io partner directory carry more weight than any adjective we could put on a homepage, precisely because we did not write them.

5. First-party data: the one advantage nobody can copy

Models cite numbers, and a number needs a source. If the data comes out of your operation, your clients or your own analysis, whoever uses it has to name you.

What to do: publish what you see and nobody else sees. You do not need a market study with a thousand respondents. One real pattern, well told, is enough: something like "across the ecommerce accounts we audit, lifecycle convergence lands on almost the same order number every time", measured on your own data.

How this gets measured today

Uncomfortably, and that is worth saying out loud. There is still no console that reports AI citations the way Search Console reports impressions.

What works in the meantime is tidy and manual: you build a list of 20 to 30 questions your buyers actually ask, run it once a month across the assistants your market uses, and note whether you appeared, in which phrasing and from which source. In three months you have a trend. It is not elegant, but it is yours and it does not depend on anyone publishing it for you.

What did not change

All of this sits on top of a site that loads fast, can be crawled, has a clear architecture and content that resolves something. An answer engine cannot cite what it never indexed.

That is why the order matters: technical SEO first, then answer structure, then the entity. The other way around does not work, and it is the most common mistake we see when somebody writes to us because they "want to show up in ChatGPT".

AI chatbots for business: how to pick the one that fits

Tomas Lagomarsino — Fri, 12 Sep 2025 00:00:00 GMT

If you run a small or mid-sized business, you wear several hats. The promise of an AI chatbot, someone who answers around the clock and never gets tired, is hard to ignore. The problem is that the market turned into a maze of acronyms and every platform promises the same thing.

The uncomfortable truth: there is no best chatbot. There is the one that solves your most expensive problem. This guide orders the decision the opposite way from how it usually gets made, starting from the problem and ending at the tool.

The four questions that order the decision

Before you look at a demo, answer these four. The answers rule out half the market before you trial anything.

Where do people message you today?

This is the question that most determines the outcome and the one least often asked. Look at where conversations actually arrive, not where you wish they did.

If most of them come through WhatsApp or Instagram, you need a platform that was born in messaging. An excellent web chat widget will solve the small slice of the problem and you will end up with two inboxes. If contact is mostly on your site, the equation flips.

What volume do you have?

Fifty conversations a month and five thousand are different products, not the same product on another plan. At low volume almost any tool is enough and what matters is configuring it fast. At high volume the pricing model starts to bite (per conversation, per resolution, per agent) along with the ability to report.

Does the bot need to know who is writing?

This is where the market splits in two.

A bot that only reads your website answers questions. A bot connected to the customer profile knows whether they already bought, what they bought, whether a delivery is in progress and what you promised them last week. That is not an incremental improvement: it is the difference between support and selling. At Grupo Garden that connection is the central piece, because the bot does not only respond, it also generates the signal that later decides the automated follow-up.

What happens when the AI does not know?

The most underestimated part. A bot that cannot pass a conversation to a person, with context and without making the customer repeat everything, creates more friction than it saves.

Ask any vendor how escalation works and ask to see it live. The answer is a good quality filter.

The platforms, by job to be done

With those four answers on the table, this gets short.

If you sell over WhatsApp and Instagram

This is the most common case in Latin America and the one worst covered by tools built for the US market. You need real WhatsApp API handling, a shared inbox across several agents, and AI that understands a sales conversation rather than only an FAQ.

It is the ground we built GoChat for: it centralises WhatsApp, Instagram and Messenger into one multi-agent platform, with AI agents that hold the conversation and record what happened. If 90% of your contact is on your website, it is not the right choice, and we would rather say so.

If you already live inside a CRM

If the team works in HubSpot or Salesforce all day, that ecosystem's native chatbot has an advantage no integration matches: every conversation, lead and ticket is logged with no glue in between.

The cost is usually double. The free versions are genuinely basic, and the real capabilities sit in the higher tiers. It is worth it if the CRM is already the centre of the operation; it is not worth it if you are adopting one because of the chatbot.

If you want to start cheap and learn

There is a category of all-in-one tools combining live chat, AI automation and a multichannel inbox, with free plans that hold a moderate volume. They are the fastest and least risky way to find out what your customers actually ask.

That is a sensible first step, with one condition: treat it as an experiment with an end date. What you learn in three months is worth more than the tool, and that information is what lets you choose well afterwards.

If you sell B2B with a high ticket

When a qualified lead is worth thousands, the conversation changes purpose: it is not about saving support time, it is about identifying the right account and getting it talking to a salesperson as soon as possible.

Platforms in this category pay for themselves in that scenario and are wildly expensive in any other. The maths is simple: if one extra meeting a month justifies the cost, it is your category.

The mistake that keeps repeating

Almost every project that goes wrong starts the same way: the tool gets chosen first and the process gets thought about afterwards.

An AI chatbot does not fix a process that does not exist. If nobody today knows what happens to a lead that arrives on a Saturday, automating the Saturday reply will produce leads nobody attends to on Monday either, faster and in larger numbers.

The order that works is the reverse. Define what information you need recorded from every conversation, who acts on it, and within what time. Then pick the tool that best supports that. It is less entertaining than running demos, and it is the part that decides whether the bot is still switched on in six months.

How to trial it without overspending

Three steps, two weeks.

Pick two candidates, not five. With five you will not finish the trial.
Load them with the same material and ask the five hardest questions your customers actually ask, including one the AI should not be able to answer. That is where you see how it escalates to a person.
Look at two numbers at the end: how many conversations resolved without a human, and how many leads were captured with usable data.

What you are looking for is not the most advanced AI. It is the one that gives you time back and does not leave you paying for conversations a person ends up handling anyway.