The questions people ask before buying

An assistant for your website and WhatsApp that reads your own site, documents and prices, then answers your customers from those and nothing else. When it does not know, it says so and takes a message.

About fifteen minutes. You give it your web address, it reads your pages by itself, and you paste one line into your site. There is nothing to train and no flowchart to draw. You can watch it answer questions about your business before you make an account at all.

No. It is one line, pasted the way you would paste a statistics tag, and there is a page of step-by-step instructions for WordPress, Wix, Shopify, Webflow and the rest. If you would rather not, ask and it will be done for you.

Yes. Type your web address on the home page and it will read your site and answer questions about your business there and then. Putting it on your own site needs an account, and every new account starts with a free trial with no card.

Your web address is enough to start. After that, anything that answers a question customers keep asking: a price list, a PDF, a menu, your terms, or just typed answers. The more it holds, the less often it has to say it does not know.

One chat window, with its own knowledge, its own look, and its own place to live — usually one website. It is the thing your customers talk to. A plan allows one or more of them, and a business with two brands or two sites runs one assistant per site so neither answers with the other's material.

The unit your plan counts. One visitor's whole sitting is one conversation, whether that is two messages or fifteen; it ends after thirty minutes of quiet or fifteen messages, whichever comes first. It only counts against your allowance if the assistant actually answered something — a greeting somebody abandons is free, a sitting it could not help with is free, and a sitting your own team answers by hand is never counted.

Everything the assistant is allowed to answer from: the pages it crawled from your website, documents you uploaded, text you pasted, and question-and-answer pairs you wrote. Nothing else. Adding to it is how you make the assistant better, and the knowledge screen shows exactly what it holds.

How stored knowledge is measured, because "characters" means nothing when you are deciding if a plan fits. A page is about 3,000 characters — roughly a typed A4 sheet. A ten-page PDF is about ten pages; a long website might crawl to forty. Your plan carries a page allowance and the knowledge screen shows where you stand against it.

A visitor who left a way to reach them — a name with a phone number or an email address. The assistant asks at the right moment: after it has answered somebody who seems interested, or when it could not answer and a colleague needs to come back to them. The leads screen lists every one, with the conversation it came from.

A question the assistant could not answer, recorded in the customer's own words with a count of how often it has been asked. The screen doubles as a to-do list: answer one there, once, and it stops being missed. It is the single best measure of what your material is missing.

Every conversation across your assistants, in one place — answered, declined, and waiting for a person. This is where your team reads threads, takes over from the assistant, and replies as themselves.

A person from your team replying in a live conversation. The assistant goes quiet on that thread the moment somebody takes it, and stays quiet until it is handed back, so the customer is never answered by two voices at once. The window tells the customer they are now talking to a person.

The assistant deciding a person should see a thread. It happens when a customer asks for a human, when the assistant fails twice in a row, or when a conversation sounds like a complaint. The thread is flagged in the inbox and whoever should know is emailed.

The chat window itself — the bubble in the corner of your website and the panel that opens from it. It is installed by pasting one line into your site, and it ships in English, French, Arabic and Spanish.

The assistant reading your website: it fetches your pages, follows the links it finds, and keeps the text. A crawl is a snapshot of that day, not a live view — when your site changes, refresh the source on the knowledge screen and it reads the site again.

A person who can sign in to your dashboard. Nothing is charged per seat — the people come with the plan: 1 on Free, 2 on Starter, 3 on Business. Everyone on the account can work the inbox; changing settings is for owners and admins.

Sublex Chat answers only from the material you gave it. It cannot quote a price, confirm a booking, or give medical, legal or financial advice, however the question is put. Passages retrieved from your pages are fenced as information rather than instructions, so a sentence planted in a crawled page cannot tell it what to do.

It says so plainly, in its own words rather than a canned line, asks for whatever contact details you chose to collect, and records the question on a screen in your dashboard in the customer's own words. You answer it once and it stops being a gap.

Ask it the same question in the test panel on your assistant. That panel names the sources the answer was built from and how closely each one matched, so you can see which page or document said it. Because it only ever answers from your own material, a wrong answer means the material is wrong, out of date, or missing, and all three are things you can correct rather than a black box you have to argue with.

Yes. Someone from your team replies as themselves in the same window, and the assistant goes quiet on that thread until it is handed back. The customer never has to repeat themselves, and a conversation your team answers is never counted against your allowance.

Asking for one is always enough, however they phrase it, and that cannot be switched off. The thread is flagged, whoever should see it is emailed, and the assistant stops trying to handle it. It also hands over by itself when it fails twice in a row or when a conversation turns into a complaint.

Whatever language the customer writes in, with no setting to change, and even when your own material is in another one. The chat window itself, the buttons and labels a visitor sees, ships in English, French, Arabic and Spanish, with the whole layout mirrored for Arabic rather than the words swapped. All of that is on every plan including the free one.

Nearly always that the answer is not in what Sublex Chat was given. Open the missed questions screen: it lists every question it could not answer, in the customer's own words, with how many times each has been asked, and a button to answer it. That list is the fix, and it is usually shorter than people expect. Check the knowledge screen too, because a crawl often reads fewer pages than a site appears to have.

Usually a wording problem rather than a missing one. Your customers ask for a “callout fee” and the document calls it a “minimum charge”, and nothing connects the two. The fix is a question and answer pair written in your customers' words, which takes a minute and outranks a long document. Ask the same question in the test panel first: it names which sources were used and how closely each matched, so you can see whether the right page was found and passed over, or never found at all.

The crawler reads the HTML a page returns and obeys robots.txt. Three kinds of page are invisible to it: pages a browser builds with JavaScript after loading, pages behind a login, and anything your robots.txt disallows. If a page looks empty to it, copy the text in as pasted text or a document instead. The knowledge screen shows how much it actually holds, which is the quickest way to spot a site that gave up less than you thought.

A scanned PDF is a picture of words, not words, and there is nothing in it to read. It will tell you so rather than pretend, and the source will show as not read with the reason. Export a real PDF from whatever produced the document, or paste the text in. The same goes for a price list that is one large image.

Look at your own website first. Example chat bubbles, testimonials, a design placeholder, or a screenshot caption are read as prose by any crawler, and a decoration that reads like a fact becomes one. This is not a hypothetical: a demo assistant once answered “Yes, 9am to 2pm” about Sunday opening for a business with no opening hours, because the homepage carried a decorative example chat saying exactly that. Remove or hide the decoration, delete that source, and crawl again.

A crawl is a snapshot of the day it ran, not a live view of your site. Change the page, then refresh that source so it reads it again. Where a document has been superseded, delete the old one rather than uploading the new one beside it: two price lists in the knowledge means two prices are true as far as it knows, and it has no way to tell which of you is right.

This happens when one assistant has been given more than one business's material. It has the address every passage came from and will name whose information it is, but the cleanest answer is one assistant per business. The plans allow several, and separating them costs nothing.

When nothing in your material is a close enough match, it falls back to the description you wrote of your business, and nothing else. A one-line description produces one-line answers. Filling that in properly is the single highest-value ten minutes of setup, and after that every vague answer is a question your material has not been asked yet.

Sometimes, and it would be dishonest to pretend otherwise. Sublex Chat is software, and software has faults. The way to tell them apart is the test panel: if it names the right source and still answers badly, that is ours; if it names nothing, or names the wrong page, the material is the thing to fix. Send the question exactly as the customer typed it and roughly when, and you will get a straight answer about which it was. Real faults have been found this way, and they are written up rather than quietly patched.

The missed questions screen shrinks. That is the honest measure, and it is why the list exists rather than a satisfaction score nobody fills in. A question answered once stops appearing, so the list is a to-do that empties rather than a report that accumulates.

In Sublex Chat, one visitor's whole sitting is one conversation, whether that is two messages or fifteen. It ends after thirty minutes of quiet or fifteen messages, whichever comes first, and it only counts if the assistant actually answered something. A greeting somebody abandons is free, and so is a sitting your own team answers by hand. Free includes 25 a month, Starter 500, Business 2,000.

No. There is one meter and one bill. The model that answered makes no difference to what you pay, so nothing drains faster because the answer was better, and there is no second subscription for the AI.

You are told at 80%. At 100% the assistant stops answering questions but still takes every enquiry, so nothing is lost while you decide, and your own team can keep replying by hand throughout. You can add conversations for a one-off price or move up a plan. Nothing is charged that you did not press a button to buy, and you are never moved to a higher plan automatically.

Both, and they are different things. Every new account starts with 14 days of the full Business plan with no card. When that ends you land on Free and stay there for as long as you like, with 25 conversations a month. Free is not a countdown.

Change plan whenever you like and the difference is prorated against your existing subscription rather than charged as a second one. Cancelling keeps the plan running until the period you already paid for runs out.

Usually, and the arithmetic is done in public rather than claimed. Starter is $29 a month or $24 billed yearly, Business $79 or $66, with the AI, live chat and branding removal inside the price rather than added to it.

No, and not for support either. Nobody at Sublex Chat can open one: the support view excludes the tables holding messages, leads and private notes, and the database enforces that rather than a policy you are asked to trust.

No. Your customers' messages go to the AI provider to produce that one answer and for nothing else. Nothing is sold, and no model is built out of your business.

The database, file storage and sign-in run on Supabase, hosted in the EU. Answers are generated by an AI provider whose processing may happen outside it. Every subprocessor is named in the privacy policy rather than described in general terms.

Yes, yourself, without asking anyone. Export everything held on a conversation as a file, or erase the whole thread. You can also set how long transcripts are kept and have older ones removed automatically.

The key it carries is public by design and grants exactly one thing: chatting with your assistant. Which sites may use it is decided by your own domain allowlist, enforced on the server rather than in the browser, so the key is worth nothing pasted anywhere else.

Take your conversations with you as files, delete the account, and remove one line from your website. There is no export fee, no notice period, and nothing of yours is held back.

No. There is no calendar behind it, so it does not propose a time, confirm an appointment, or promise that somebody will ring at a particular hour. It treats that as something it cannot do, takes the request, and hands it to a person.

No. It answers from documents, not from your shop or your order system. Ask it whether something is in stock and it will say it cannot see that rather than guess.

Not yet. There are five public routes, all of them internal to the product, and a public API is a later phase. If your plan depends on one, that is worth knowing before you buy rather than after.

If the app shows your website in a web view, the widget comes with it. A true native app does not carry it, and there is no native SDK today.

None of those exists today. They are real requests and they are on the list, but nothing is gained by implying otherwise: if your organisation requires one, this is not yet the right product for you.