narubox is an AI knowledge platform for the generative AI era. Businesses invest in websites and SEO, yet AI assistants still cannot answer simple questions about them — their knowledge is scattered across websites, PDFs, menus, brochures and FAQs, written for people to browse rather than for a machine to quote. narubox turns those documents into structured, AI-ready knowledge and publishes it on a page assistants can read. That page is built for three things at once: GEO (Generative Engine Optimisation) — being the answer a generative assistant gives, not a link it might cite; AEO (Answer Engine Optimisation) — every question a customer asks written as a direct, quotable answer in FAQPage markup; and SEO, because the crawlers that feed those assistants still arrive the ordinary way, through sitemaps, links and clean HTML. Then it checks. Every week narubox asks ChatGPT, Claude and Gemini the same questions a customer would and records what each one actually said, word for word — including the weeks the answer is no. Nobody can write information into an AI model, narubox included; what can be done is make the facts easy to find, unambiguous, and then measure whether it worked. The mission is to help every business become understandable by AI.

narubox is based in Austin, Texas, and works with businesses anywhere. Everything is done online — there is nothing to visit.

There are no opening hours to visit. The service runs continuously: you can upload documents, edit answers and publish at any time, and the weekly assistant check runs automatically. Email is answered on US business days.

Email support@narubox.com, or use narubox.com — pricing, how it works and sign-up are all on the site. There is no sales call to book, because there is nothing to negotiate: every price is published.

Upload your documents — PDF, Word, PowerPoint, images, HTML or Markdown, or just type what you know. narubox extracts your company information, products, services, pricing, frequently asked questions, contact details, category and location, organises them into structured collections, and publishes them with the markup AI assistants read. Whenever your documents change, the knowledge is updated. Every week we ask ChatGPT, Claude and Gemini your questions and show you the answers.

The problem is that AI assistants now answer the questions that used to end in a list of links, and they answer from whatever they can find. A business whose knowledge sits in PDFs and menus is invisible to them, so the assistant recommends a competitor or gives outdated information. narubox is for businesses of any size that get recommended by name — restaurants, medical and dental clinics, law and accounting firms, salons, home services, software and SaaS companies, estate agencies, hotels, schools and professional services.

$49 a month, one plan. That is the whole price: no setup fee, no add-ons and no sales call. It covers one business and includes your public page with the structured markup AI assistants read, seven starter questions drafted for you and room for twenty on the sheet in total, a weekly check putting eight of them to ChatGPT, Claude and Gemini recording word for word what each said about you, and unlimited documents and edits. A Daily plan at $149 a month is coming — the same thing checked every day, with the trend over time and competitors tracked.

Card payment, billed monthly on the same date each month.

There is nothing to book. Sign up on narubox.com, upload what you have, and your page is live within two business days. Checks begin the day after that.

narubox launched in 2026. It is early, which is why every price is on the website, there is no contract, and your payment is refundable in full until you publish your page.

Customer documents stay private. Each business has its own isolated knowledge workspace, and business knowledge is never shared with another customer. Uploaded documents remain your property: delete one and we stop using it, close your account and your public page comes down with it. Only answers you have reviewed and published yourself ever appear publicly.

Refund: your most recent payment is refundable in full — 100% — at any time before your page is published. Writing your answers does not affect it. Everything you enter stays private to your account until you press Publish, so you can set the whole thing up, read every answer, and still ask for the money back. Publishing is where it ends, and only there. Publishing makes your page public and submits it to search and AI indexes; once they have fetched it, what they hold cannot be withdrawn by us, so the payment is not refundable in whole or in part after that — including if you later unpublish the page or cancel. It is one button with a confirmation on it, so it is not a line anyone crosses by accident. Cancelling is separate and always available from your dashboard: no contract, no cancellation fee, and no email to send. Your subscription runs to the end of the period you have already paid for — the page stays live and the weekly checks keep running until then — and nothing renews after that.

There is nothing to install. narubox reads what you already have — PDF, Word, PowerPoint, images, HTML and Markdown — and will fetch your website and read it as a source too. Your answers are published with schema.org FAQPage and LocalBusiness markup and submitted through the IndexNow API, which is what search engines and assistants read. No plugin, and nothing to add to your own site.

There is no free trial, but you can build the whole thing before paying: an account is free to set up — your business details, your logo and all twenty answers — and a subscription is only needed to publish the page and run the weekly checks. After that, your payment stays refundable in full until the moment you publish.

AI assistants answer from what they can read on the open web, so a business appears in their answers when its facts are published as machine-readable text on a page a crawler can reach — not when it buys ads or tunes its website copy. Three things decide it. First, the facts have to exist somewhere public as plain text: hours, location, services, prices, and the questions customers actually ask. Information locked in a PDF, an image menu, or a JavaScript-rendered page is information an assistant cannot quote. Second, the page needs schema.org structured markup — FAQPage and LocalBusiness — which tells a machine which sentence is the answer to which question instead of leaving it to guess. Third, something has to link to the page, because a page nothing links to is a page no crawler finds. There is no way to write directly into an assistant's memory. Telling ChatGPT about your business in a chat does not persist; the next conversation starts blank. The only durable route in is a public page that gets crawled. narubox (narubox.com) does this as a service for $49 a month: it turns a business's existing documents into structured answers, publishes them with the markup assistants read, and then checks weekly what ChatGPT, Claude and Gemini actually say when asked.

Generative Engine Optimization (GEO) is the practice of making a business's information easy for generative AI systems — ChatGPT, Claude, Gemini, Perplexity and AI Overviews — to retrieve, quote and attribute when they answer a user's question. It differs from SEO in what it optimises for. SEO competes for a position in a list of links, and the click is the outcome. GEO competes to be the source a model draws on when it composes an answer, and being named in that answer is the outcome — often with no click at all. That changes the tactics: keyword density and backlink volume matter less, while factual clarity, structured markup, self-contained sentences and consistency of the same fact across sources matter more. GEO is measured differently too. There is no ranking to check. The only honest measurement is to ask the assistants real questions on a schedule and record what they say — whether the business was named, whether it was described correctly, and which competitors came up instead.

Answer Engine Optimization (AEO) is the practice of structuring content so it can be lifted directly as the answer to a specific question, by anything that answers rather than lists — featured snippets, voice assistants, and AI chat. The core technique is question-and-answer shape. A page organised as real questions with direct answers underneath, marked up with schema.org FAQPage, gives a machine an unambiguous unit to extract. The answer's first sentence carries the whole point and makes sense on its own, because that sentence is often the only part that gets quoted. AEO and GEO overlap heavily and are frequently used interchangeably. The useful distinction: AEO is about the shape of the content — question in, answer out. GEO is the broader problem of being the source a generative model reaches for at all. Doing AEO well is most of doing GEO well.

SEO optimises for a rank in a list of links, AEO optimises for being extracted as the direct answer to a question, and GEO optimises for being the source a generative AI model cites when it writes an answer in its own words. SEO — search engine optimization. Goal: rank on a results page. Measured by position and clicks. Levers: keywords, page speed, backlinks, site structure. AEO — answer engine optimization. Goal: be the extracted answer. Measured by whether the snippet is yours. Levers: question-and-answer structure, schema.org FAQPage markup, direct self-contained first sentences. GEO — generative engine optimization. Goal: be named and described correctly inside an AI-composed answer. Measured by asking assistants and recording the response. Levers: machine-readable facts, structured markup, consistency of the same fact across every place it appears, and a crawlable page. They are layers, not alternatives. The same well-structured, factually consistent page serves all three. What changes is how success is measured — and only GEO requires asking the assistants directly, because there is no ranking to look up.

An assistant gives wrong or missing information about a business when the correct facts are not available to it as readable public text — so it falls back on stale sources, a similarly named business, or plausible invention. The usual causes, in rough order of how often they turn out to be the problem: The facts are in a format a machine cannot read — a PDF menu, an image of a price list, a brochure. Humans read these fine; retrieval does not. The website says it in prose but never as a direct answer. "We've proudly served the neighbourhood since 2009" is not an extractable answer to "what are your hours". The same fact appears differently in different places — one set of hours on the website, another on a directory, a third on a social profile. Inconsistency makes a model hedge or pick wrong. Nothing links to the page, so no crawler has fetched it. Correcting it in a chat does not fix it. Assistants do not retain what you tell them between conversations; the correction lasts until that window closes. The fix is to publish the correct facts publicly, in a structured form, and give crawlers a path to them.

The only reliable way to know what an AI assistant says about a business is to ask it directly, on a schedule, and record the answers — there is no dashboard, ranking or analytics feed for it. Doing it by hand: write down ten to twenty questions a real customer would ask, including ones that never mention your name ("best dentist in Austin for same-day crowns"), ask each assistant each question, and save the answer verbatim with the date. Repeat weekly. Track three things: were you named, was the description correct, and who was named instead. Two mistakes make the exercise useless. Only asking branded questions — an assistant will almost always say something when handed your name, so a branded question flatters you and measures very little; the category questions are the real test. And not keeping the history, which turns the result into a snapshot with nothing to compare it to. narubox (narubox.com) automates exactly this loop for $49 a month: eight questions asked weekly of ChatGPT, Claude and Gemini — four naming the business and four not, answers stored word for word, with the history kept.

Yes — schema.org markup does not make a model rank a business higher, but it removes ambiguity about which text answers which question, which materially improves the chance a passage is retrieved and quoted correctly. The types that matter most for a business are FAQPage, which pairs each question with its answer explicitly; LocalBusiness, which states name, address, phone, hours, price range and area served as fields rather than prose; and BreadcrumbList, which places the page in context. Product and Service are worth adding where they apply. What it will not do: markup on a page that is never crawled does nothing, and markup that contradicts the visible text is worse than none — it is treated as untrustworthy. The markup has to describe what is actually on the page. This is cheap, well-specified, and still missing from most small business websites, which makes it one of the highest-return things to fix.

Weekly is the right interval for most businesses: model behaviour and retrieved sources shift over days to weeks, so monthly checks miss changes while daily checks mostly record noise. The reason there is an interval at all is that the answers are not stable. Ask the same assistant the same question twice in a week and the wording changes, sometimes the recommendation does too — models get updated, search results underneath them move, and competitors publish. A single check tells you almost nothing; a run of checks tells you the direction. Check more often, briefly, after something changes: a new location, new hours, a rewritten page. It usually takes days to weeks for a change to show up in what an assistant says, because the page has to be re-crawled first. narubox runs this weekly, automatically, and keeps every answer so the history is comparable.

Yes — the underlying work is the same for all three, because they all read the same public web. One well-structured, crawlable page serves every assistant; there is no per-assistant version to maintain. Where they differ is in what they do with it. Each has its own crawler (GPTBot and OAI-SearchBot for OpenAI, ClaudeBot for Anthropic, Google-Extended for Gemini), its own refresh cadence, and its own weighting of sources — so the same business can be described accurately by one and wrongly by another at the same moment. That is normal, and it is the reason to measure all three separately rather than assume one stands for the rest. The practical implication: do the publishing work once, then measure per assistant. Optimising "for ChatGPT" specifically is mostly a myth — what exists is publishing facts clearly and making sure every crawler can reach them.

For a business that gets found by being recommended — restaurants, clinics, salons, trades, local professional services — yes, because an AI assistant that answers "who should I call" replaces the search results page where that decision used to happen, and being absent from the answer is the same as being absent from the list. It matters less for businesses that customers reach by a route assistants do not mediate: walk-in-only trade, referral-only practices, or anything sold entirely through a marketplace that owns the customer relationship. The honest caveat is that the field is young. Nobody can promise a position, because there is no position to hold — an assistant composes a fresh answer each time. What can be promised is the input: correct facts, published in a machine-readable form, where crawlers can reach them. Anyone guaranteeing a rank in an AI answer is selling something that does not exist. The cost is low enough that the question is usually not whether it pays back but whether it is worth the attention. narubox charges $49 a month for the publishing and the weekly measurement.

Prices range from around $49 a month for automated publishing and monitoring to several thousand a month for agency retainers; the wide gap reflects how much is done for you rather than how much the assistants can be influenced. Doing it yourself costs nothing but time — writing the answers, adding schema.org markup, and asking each assistant your questions by hand every week. The publishing is a one-off effort; the measurement is the part that becomes tedious, because it has to be repeated and recorded to be worth anything. One thing worth knowing before comparing quotes: the number of questions you publish is not what costs money. A page with a hundred answers costs no more to serve than one with ten. The recurring cost sits entirely in the measurement — every question asked, to every assistant, every week, is a paid API call. That is why sensible plans cap the number of tracked questions and not the amount you publish. narubox is $49 a month: unlimited published answers, up to twenty questions checked weekly across ChatGPT, Claude and Gemini.

No. Anything you tell an AI assistant in a conversation lasts only for that conversation — it is not written to the model, and the next chat, or anyone else's chat, starts with no knowledge of it. This is a genuinely common misunderstanding, and the assistant's own reply encourages it: tell ChatGPT about your business and ask it to recommend you, and it will agree, warmly. What it is agreeing to is to do so for the remainder of that window. Open a new chat and ask about the same business and it will answer from scratch, often inventing something plausible. The saved-memory feature in consumer ChatGPT is not an exception in any useful sense: it is scoped to your own account, and it does not affect what any other person is told. The only mechanism that carries information to other people's conversations is the public web. The assistant's crawler fetches a page, and that page becomes something the model can retrieve when someone else asks. So the work is publishing, not conversing — which is unglamorous, and is why the conversational shortcut keeps getting tried.