Personas A/B: test two versions of the service AI and promote the one that converts more

Personas A/B

Two versions of your AI talking at the same time to find out which one converts more — and the care needed to avoid deciding on a lucky number.

A persona is the way your AI talks: the name it uses, the tone, what it offers, what it answers by heart. Here you put two versions live at the same time and let the leads decide which one sells more — instead of arguing about tone of voice in a meeting.

The screen says atribuição sticky por lead. Sticky means stuck: each lead always talks to the same version, from the first hello to the end. Nobody notices a change of tone mid-conversation, and the measurement doesn't get mixed up.

The screen, part by part

Start an experiment

1. Click "Criar variante B (clonar persona ativa)" (Create variant B (clone active persona)): B starts as an exact copy of the persona that is already live, and the test turns on right away. From then on, each new lead lands in A or B and stays there.
2. Change only one thing in B: A more direct tone, or a different featured offer, or a different first sentence. If you change three things and B wins, you won't know which one won — and you won't be able to repeat the success.
3. Let it run without checking every day: With the normal traffic of most accounts, gathering 30 leads on each side takes two to four weeks. Checking the scoreboard every day only produces anxiety and hasty decisions.
4. Wait for the significance card to clear: While it says "Coletando dados" (Collecting data), there is no result — neither for nor against. Whoever is ahead today easily switches sides when there are few leads.
5. Promote the winner: Only when the "VENCEDORA" (WINNER) badge and the sentence saying it can be promoted with confidence appear together.

Reading the two cards

"conversão lead→qualificado" (lead-to-qualified conversion)

The big number. Of every 100 people who talked to that version, how many became hot or warm. This is the measure that decides the experiment.

"1ª resposta" (1st response)

How long until the first reply goes out. It works as a sanity check: if one of the versions is much slower, compare carefully, because part of the conversion difference may be waiting, not text.

handoff

Handoff: the share of conversations that needed to be passed to a person on the team. Less is not always better — in high-ticket sales, more handoffs with more closings is great. The queue is in "Transbordo para humano" (Handoff to a human).

leads

How many people talked to that version. It is the first number to look at: without people on both sides, the rest is decoration.

Significance, in plain English

The screen warns that "vencedor só é declarado com teste z de duas proporções, exigindo p < 0,05 e amostra mínima de 30 leads por variante" (a winner is declared only with a two-proportion z-test, requiring p < 0.05 and a minimum sample of 30 leads per variant). That is a statistics sentence, and it means three simple things:

• The math compares two rates, A's and B's, taking into account how many people went through each one. Twenty percent of 5 people is not worth the same as twenty percent of 500.
• `p` is the chance that the difference is luck. p = 0,03 means a 3% chance of coincidence. Below 0.05 the screen accepts it as a real difference.
• The "IC95%" (95% CI) is the range where the real difference should lie. If the range goes from -4 to +12 points, it passes through zero — and passing through zero means a tie still fits in the math. No promoting.

Promote, pause, and resume

These three actions have different consequences, and two of them are hard to undo. It is worth reading before you click:

Promover

Ends the experiment. The winner becomes the account's only persona and the other one is paused. In practice there is no going back: to compare again, you start a new experiment from scratch.

"Pausar a B" (Pause B)

Sends 100% of new leads to A. Those who were already talking to B stay with it — the stickiness lasts until the end of that conversation.

"Retomar" (Resume)

Splits the leads again, but restarts the measurement from zero. Everything accumulated before the pause is lost, including the 20-some leads that had already been counted.

The persona editor

"Editar personas" (Edit personas) opens the screen where each version is written. There are few fields, and each one changes a different thing in the conversation:

"Nome da persona" (Persona name)

The internal label, so you can tell A from B in the list. The lead never sees this name.

"Nome da IA" (AI name)

The name that appears to the lead in the conversation. This one, yes, they read.

"Empresa exibida" (Displayed company)

How the AI introduces itself. It should match what is written on your channels.

"Ofertas" (Offers)

What it may offer in this version. It is the field that changes results the most, more than the tone.

"Tom de voz" (Tone of voice)

How it talks, written by you. The four shortcuts — "Direto" (Direct), "Consultivo" (Consultative), "Descontraído" (Casual), and Formal — only fill the box with starter text; they don't save anything. After clicking, read, adjust, and save.

"FAQ (respostas prontas)" (FAQ (ready-made answers))

Recurring questions, with the exact answer you want sent. Useful for price, lead time, and policy — the three where improvising is costly.

"Responder em voz (TTS)" (Reply by voice (TTS))

Reply in audio instead of text. It changes the lead's perception a lot, so test it on its own, in an experiment of its own. Details in "Voz" (Voice).

One persona per account

If you serve through more than one account — two WhatsApp numbers, an Instagram account, the web chat — each one can have its own persona. This is what allows two brands, or two franchises, in the same dashboard: the headquarters number replies as Sofia and the second brand's Instagram replies as Marina, without mixing.

1. Open "Conexões" (Connections): Each account appears with a line saying "Atende como…" (Responds as…) — including those that were never configured, which say a persona da empresa.
2. In the account menu, choose "Escolher persona" (Choose persona): On Instagram, the button is on the account's own row. The list offers the option Persona da empresa (padrão) and all active personas.
3. Save: Applies from the next message onward. Conversations that were already open stay with the persona that was already handling them — switching tone in the middle of a chat is exactly what sticky assignment avoids.

The A/B test keeps running

When you link an account to a variant that is in a test, that account keeps drawing between A and B — the link picks the pair, it doesn't freeze the test. Accounts without a link stay in the company's test, as before.

Only an active persona can be linked

A paused persona answers nobody; allowing it to be linked would create a screen saying Marina replies on a number where she never speaks.

If the linked persona is paused later

The account doesn't go silent: it goes back to replying with the company persona until you choose another one.

If the account is disconnected

The link is kept. When you reconnect the same number or the same Instagram account, it goes back to replying with the same persona — nothing needs to be redone.

In "Configurações" (Settings) → "IA & persona" (AI & persona) there is the same map in read-only form, in the "Quem responde por cada conta" (Who answers for each account) block — useful for checking everything at once without opening "Conexões" (Connections). And the "Base de conhecimento" (Knowledge base) has the equivalent scoping: each document can apply only to the accounts you select.

What is worth testing

• The first sentence. It is what decides whether the person replies. Testing only this usually pays off more than rewriting the whole persona.
• Asking early or late. A version that already asks about the timeline in the second message versus one that chats first.
• One featured offer versus three. Fewer options usually qualify more people, but this changes from industry to industry — which is why you test it.
• Formal versus casual. It is the most requested test and almost always the one with the smallest effect. Leave it for after the other three.

After promoting, confirm the effect where it shows up in money: Analytics, same period, same filters, one week before and one week after.

In the dashboard: "Abrir Personas A/B" (Open Personas A/B)