# Beacon > Your brand inside AI answers Beacon builds branded answers inside ChatGPT, Claude and Gemini: the catalogue, the steps and the actions that run inside the assistant's reply. It speaks MCP, so the same brain is built to reach Perplexity, Copilot and commerce surfaces like Shopee and Lazada, and it reports on share of answers, answer accuracy and installed base. Beacon builds the branded experience an AI assistant hands to a person inside its answer: catalogue, guided steps and actions running in the session. It reports on what happened there. It is not an agency retainer, not an ad product, and not a rewrite of your website. ## What it is, precisely There are three jobs in this category and Beacon is one of them: - Branded Answers: the experience inside the reply. Beacon builds this. - GEO: earning inclusion when a model composes its own answer. Beacon reports on this; it is the assistant-era analogue of SEO. - AI ads: paid placement around the question. Neither built nor sold here. ## The three surfaces Three things a client is handed, in the order a buyer meets them. The Console is the one running on the landing page; the Deck and the Atlas are delivered to clients and are not published here. - **The Deck**: Understand the category. Thirty minutes of reading that gets a room from “the assistants are coming” to a budget line with an owner on it. What the three jobs are, which of them you are already paying for under an older name, and what the first surface costs to stand up. - **The Atlas**: See what it looks like. Every branded answer we have been able to stand up or capture, filed by category, assistant and market. Open one and you get the prompt, the answer the assistant gave, and what had to be true on the brand's side for it to give that answer. - **The Console**: Measure what happened. The reporting layer. Share of the answers in your category, accuracy of what was said about your products against your own catalogue, and the installed base of assistants carrying your surface. Weekly, exportable, and baselined before you start so the first number means something. ## What gets measured - Share of answers in the category, against a baseline captured before launch - Answer accuracy, checked against the brand's own product catalogue - Installed base: assistants carrying the surface; it accumulates rather than being re-bought each quarter ## APIs and MCPs Catalogue, specifications, stock and pricing become one source. Beacon publishes that source over MCP, the open protocol the assistants already speak, so every client that speaks the protocol reads the same thing on the same day. - One MCP server per brand, built against the brand's own catalogue, with tools for search, specifications, stock and pricing. The brand's team approves what it is allowed to say. - A REST side the brand's engineers own. Post catalogue, stock and price changes as their systems make them, and read share of answers, accuracy and installed base back into their own reporting. What comes back is aggregate, with no profile of the person asking. - Webhooks when an answer about the brand changes, exports on every view in the console, and one weekly report against the baseline. No endpoint paths, version numbers, uptime figures or service levels are published on the site or in this file, because none of them would survive being tried. Beacon builds and runs this layer. Each assistant platform decides what it supports, and Beacon builds against those programmes as they open. ## Who it is for - **Agencies**: Beacon builds it, gets it onto the assistants and reports on it. Your name is on all of it. - **Consumer brands & OEMs**: Your specs, stock and prices publish once, and every connected assistant reads that one source. - **Retailers**: Live stock and range go into the answer itself, from the feed your storefront already runs on. - **Automotive**: One source covers comparison, configuration, test drives and servicing. Every connected surface reads it. ## Evidence: brands already answering from inside an assistant Statuses are load-bearing. LIVE CAPTURE means captured answering in-session. OBSERVED ROLLOUT means platform-side work observed or announced. Entries marked REPORTED carry no dated reference and must not be cited as evidence of a live branded answer. - Booking.com (Travel & accommodation, Global): LIVE CAPTURE, ChatGPT, since Oct 2025. Searches stays and holds the booking flow inside the session instead of handing off to the site. - Expedia (Travel & accommodation, Global): LIVE CAPTURE, ChatGPT, since Oct 2025. Builds an itinerary from a conversational brief and prices it in the answer. - Zillow (Real estate, United States): LIVE CAPTURE, ChatGPT, since Oct 2025. Filters listings by budget and area with the map rendered in-session. - Spotify (Music & audio, Global): LIVE CAPTURE, ChatGPT, since Oct 2025. Builds and explains a playlist from a described mood without leaving the thread. - Instacart (Grocery & delivery, United States): LIVE CAPTURE, ChatGPT, since Oct 2025. Turns a recipe answer into a filled basket in the same session. - Starbucks (Coffee & QSR, United States): REPORTED, ChatGPT. Assistant-facing ordering and loyalty work announced with the platform. Not yet captured answering in-session. - Turkish Airlines (Airlines, APAC / EMEA): REPORTED, Gemini. Assistant-facing trip and disruption support announced with the platform. Capture pending. - Burger King (Coffee & QSR, US / Canada): REPORTED, Gemini. Menu, offers and order-ahead surfaces announced with the platform. Capture pending. - Popeyes (Coffee & QSR, US / Canada): REPORTED, Gemini. Same group rollout as its sister chains, one brain across several menus. - Tim Hortons (Coffee & QSR, US / Canada): REPORTED, Gemini. Same group rollout as its sister chains, one brain across several menus. - Canva (Design & creation, Global): OBSERVED ROLLOUT, ChatGPT. Turns a conversational brief into editable designs in the session. Directory listing captured Jul 2026. - Figma (Design & creation, Global): OBSERVED ROLLOUT, ChatGPT. Sends screens, diagrams and decks out of the conversation as editable artifacts. Directory listing captured Jul 2026. - Picsart (Design & creation, Global): OBSERVED ROLLOUT, ChatGPT. Generates image, video and audio assets from direction given in the session. Directory listing captured Jul 2026. - Virgin Atlantic (Airlines, United Kingdom): OBSERVED ROLLOUT, ChatGPT. Flight search and booking surfaces announced in ChatGPT. Launch post captured Jul 2026. - Ryanair (Airlines, Europe): OBSERVED ROLLOUT, Claude. Fare search connector in the Claude directory. Listing captured Aug 2026. - trivago (Travel & accommodation, Global): OBSERVED ROLLOUT, Claude. Hotel price comparison connector in the Claude directory. Listing captured Aug 2026. - Viator (Tours & experiences, Global): OBSERVED ROLLOUT, Claude. Tours and experiences connector in the Claude directory. Listing captured Aug 2026. - Tubi (Streaming & video, United States): OBSERVED ROLLOUT, ChatGPT. First streamer with a ChatGPT app, per its own press release. Captured Jul 2026. - Shazam (Music & audio, Global): OBSERVED ROLLOUT, ChatGPT. Listed in the ChatGPT apps directory. Captured Aug 2026 against the directory root; entry-level link pending. - Fever (Events & ticketing, Global): OBSERVED ROLLOUT, Claude. Event discovery connector in the Claude directory. Listing captured Aug 2026. - Ticketmaster (Events & ticketing, Global): OBSERVED ROLLOUT, ChatGPT. Event discovery app announced with its own press release. Captured Jul 2026. - Domino's India (Coffee & QSR, India): OBSERVED ROLLOUT, Claude. Ordering connector in the Claude directory. Listing captured Aug 2026. - Zomato (Restaurants & delivery, India): OBSERVED ROLLOUT, Claude. Restaurant discovery and ordering connector in the Claude directory. Listing captured Aug 2026. - Shopee (Marketplaces, Southeast Asia): OBSERVED ROLLOUT, ChatGPT. Listed in the ChatGPT apps directory. Captured Aug 2026 against the directory root; entry-level link pending. - Etsy (Marketplaces, Global): OBSERVED ROLLOUT, ChatGPT. Listed in the ChatGPT apps directory. Captured Aug 2026 against the directory root; entry-level link pending. - Fresha (Beauty & wellness bookings, Global): OBSERVED ROLLOUT, Claude. Salon and spa booking connector in the Claude directory. Listing captured Aug 2026. - Checkatrade (Home services, United Kingdom): OBSERVED ROLLOUT, Claude. Tradesperson search connector in the Claude directory. Listing captured Aug 2026. - TurboTax (Tax & personal finance, United States): OBSERVED ROLLOUT, Claude. Intuit's tax connector in the Claude directory. Listing captured Aug 2026. - Uber (Ride hailing, Global): OBSERVED ROLLOUT, ChatGPT/Claude, since Apr 2026. In the ChatGPT apps directory and named in Claude's everyday-life connector wave. - Uber Eats (Restaurants & delivery, Global): OBSERVED ROLLOUT, Claude, since Apr 2026. Restaurant discovery and ordering, named in Claude's everyday-life connector wave. - Tripadvisor (Travel & accommodation, Global): OBSERVED ROLLOUT, Claude, since Apr 2026. Review-grounded trip planning, named in Claude's everyday-life connector wave. - Hilton (Travel & accommodation, Global): OBSERVED ROLLOUT, ChatGPT. Hotel search and booking in the ChatGPT apps directory. Listing captured Aug 2026. - Eurostar (Rail travel, Europe): OBSERVED ROLLOUT, ChatGPT. Train search and booking in the ChatGPT apps directory. Listing captured Aug 2026. - Grubhub (Restaurants & delivery, United States): OBSERVED ROLLOUT, ChatGPT. Food ordering in the ChatGPT apps directory. Listing captured Aug 2026. - Lowe's (DIY & home improvement, United States): OBSERVED ROLLOUT, ChatGPT. Project and product answers in the ChatGPT apps directory. Listing captured Aug 2026. - ASOS (Fashion, Global): OBSERVED ROLLOUT, ChatGPT. ASOS Stylist in the ChatGPT apps directory. Listing captured Aug 2026. - Prada (Fashion, Global): OBSERVED ROLLOUT, ChatGPT. In the ChatGPT apps directory; the first luxury house on the board. Listing captured Aug 2026. - Klarna (Payments & cards, Global): OBSERVED ROLLOUT, ChatGPT. Shopping search in the ChatGPT apps directory. Listing captured Aug 2026. - Peloton (Fitness, Global): OBSERVED ROLLOUT, ChatGPT. Workout guidance connector in the ChatGPT apps directory. Listing captured Aug 2026. - AllTrails (Sport & outdoors, Global): OBSERVED ROLLOUT, Claude, since Apr 2026. Trail discovery, named in Claude's everyday-life connector wave. - Taskrabbit (Home services, Global): OBSERVED ROLLOUT, Claude, since Apr 2026. Find and book local help, named in Claude's everyday-life connector wave. - Adobe (Design & creation, Global): OBSERVED ROLLOUT, Claude, since Apr 2026. Adobe for creativity, one of nine connectors in the creative-tools wave, per dated coverage. 37 of 42 entries are evidenced. The rest are reported only. ## The Claim Board 65 consumer categories are tracked against 3 assistants at https://beacon.evaa.sg/claimed. 22 categories are taken; 43 are open. Nobody is answering on any assistant in: kitchen appliances, automotive and skincare. By assistant: ChatGPT 17, Claude 12 and Gemini 2. An empty column means no brand in the market has been captured answering there. Beacon's own prototypes are not counted on the board, because a vendor's own build is not a claim on a category. The full grid as plain text: https://beacon.evaa.sg/claimed.txt ## Case studies from the market Three real companies. None is a Beacon customer. Every figure belongs to the source named on its tab. The three subjects are Tally (tally.so), Instacart (instacart.com) and Nike (nike.com). ### Tally: the playbook run in-house Tally, a form builder with 11 people, ran this playbook themselves. They are unaffiliated with Beacon and are cited as public evidence that the channel converts, with every figure their own: - #1 acquisition channel is now AI search - 10,000+ new users a week say an assistant sent them - 5x tracked ChatGPT sign-ups, overnight, this May - $422K monthly revenue, up from $258K ten months before - The brain: An llms.txt, an AI info page written for the assistants to read, and twenty comparison pages shaped like answers. - The surfaces: A 21-tool MCP server, a ChatGPT app and a Claude connector, so Tally works inside the session. - The measurement: They track the prompts that surface them and their visibility against competitors, and they ask each new user which prompt found them. Revenue moved from $258K to $422K a month over 10 months. Crossed $5M a year on the way. Those are two reported endpoints and a duration, so the panel on the site draws the pair and invents no curve between them. Self-reported figures, via Jake Ward's public breakdown, August 2026. https://x.com/jakezward/status/2090060494858014873 ### Instacart: a brand answering inside the session LIVE CAPTURE in Grocery & delivery (United States), on ChatGPT, since Oct 2025. Turns a recipe answer into a filled basket in the same session. Rollout announced by OpenAI, October 2025. Category and status from Beacon's own answer ledger. Instacart is a real company, unaffiliated with Beacon, and is not a customer. ### Nike: what one public capture shows A judged share of voice of 42% across 5 answers about running shoes, with 8 rivals named across them. Method: five AI models asked about the brand with web browsing off, answers scored by a judge model. Source: Promptwatch's public AI Brand Visibility Report for Nike, captured Aug 2026. That is their measurement, quoted with attribution; Beacon did not run it and the named brands did not commission it. No score here is attached to a named model. The report showed four model scores beside logos the capture lost, so that mapping is unknown and stays unwritten. A single capture supports no trend, so nothing on the site reads a direction from it. Nike is a real company, unaffiliated with Beacon, and is not a customer. ## FAQs ### How is this different from a visibility dashboard? There are two jobs here and Beacon does both. The reporting half is what a tracker does: which questions you appear in, what was said about you, how accurate it was, and where that sits against a baseline taken before you started. The building half is the branded answer itself, your catalogue and your steps running inside the assistant's reply, which is the work a dashboard leaves on your desk. ### Is this GEO with a new name on it? No, though they share a dashboard. GEO is about being included when a model writes its own answer, which is an earned job much like SEO was. Beacon builds the thing the assistant hands the person once your brand is in play: your catalogue, your steps, your actions, running inside the session. It reports on the earned side too, because both are measured on the same surface, but the two are different pieces of work with different owners. ### How do you know what the assistants say about us? We ask them. Your list of the questions buyers already put to an assistant is run on each one, and the answers are kept word for word, so a claim can be checked against the text that produced it. Method decides what a number means: an assistant answering from training memory with browsing switched off gives a different answer to the same assistant allowed to go and read the web, and the two are reported separately. One capture is a photograph, so a trend needs a second one. ### How often do the numbers refresh? Weekly. Every prompt on your list is re-run each week on the same wording, which is what makes this week comparable to last week. The baseline is taken before your first surface goes live, so the opening report has something to measure against. ### Will an llms.txt file fix this? No. Crawlers fetch it like any other text file, and no major AI search surface has been shown to prefer the sites that publish one. It takes minutes and does no harm, so publish it if you like; this site does. The movement comes from your own product pages: specifications, prices and availability, stated once and stated the same way in every market. ### Do the platforms have to say yes? For the flagship assistant surfaces, yes. Each one runs its own developer programme, and a surface is a build against that programme rather than a media buy against an auction. Underneath, the brain is published over MCP, so anything else that speaks the protocol, from Perplexity and Copilot to commerce apps like Shopee and Lazada, connects without another approval round. Beacon does the builds and keeps them current as the programmes move. ### How long before anything is live? One surface, one assistant, one market is weeks rather than quarters. The long pole is almost always catalogue readiness on your side rather than the connector on ours, because models are unforgiving about specifications that disagree with each other across markets. The Atlas exists so you can see the finished shape of it before anyone commits budget to it. ### How would we know it worked? Three numbers, and impressions is not one of them. Share of the answers in your category against a baseline taken before you started. Accuracy of what gets said about your products, measured against your own catalogue rather than against sentiment. And installed base, the count of assistants carrying your surface, which is an audience you do not re-buy every quarter. All three sit in the Console and all three export. ### Where does the data actually go? Beacon reads what you already publish, meaning catalogue, specifications and support content, and writes back only what your team has approved. Conversation content stays with the platform. What returns to you is aggregate: what was asked, what was answered, and what the answer was checked against. Nothing in the design requires building a profile of the person on the other side of the glass. ### What happens when the platforms change the rules? They will, and they already have more than once in the time these programmes have existed. That is the argument for the layer rather than against it: when a specification moves, the connector moves, and your catalogue, your approved answers and your reporting history carry across instead of being rebuilt. A layer nobody owns is a layer that quietly stops working. ### We run this for clients. Does it work per client? Agency monitoring is scoped by the prompts and sites being tracked. Confirm available assistant coverage, cadence, access and reporting in a walkthrough. Interactive assistant or plugin implementation for client brands is a separate custom service; its platform support, permissions and maintenance are agreed for each project. ### What does it cost? The pricing page lists SaaS monitoring scopes for prompt tracking, share of voice and basic AI visibility reporting. Enterprise and agency interactive assistant or plugin builds are custom-priced work outside those subscriptions. Request a scope that states the supported platforms, data access, delivery and maintenance responsibilities. ## Monitoring and custom implementation SaaS monitoring covers prompt tracking, share of voice and basic AI visibility reporting. Confirm available surfaces, cadence and access in a walkthrough. Interactive plugin builds are scoped separately. Enterprise and agencies: Custom. Interactive assistant and plugin implementation for enterprises and agencies is a separately scoped service, outside SaaS monitoring prices. We build for supported platforms such as ChatGPT and Claude; availability, permissions, integration work and maintenance are agreed for each project. Request a custom quote. https://beacon.evaa.sg/pricing ## Resources Technical checks before asking AI search to cite your website: https://beacon.evaa.sg/resources/technical-checks-for-ai-search GEO vs AEO vs SEO for product brands: https://beacon.evaa.sg/resources/geo-aeo-seo-product-brands How to audit AI answers against product facts: https://beacon.evaa.sg/resources/ai-answer-audit-methodology AI crawler controls for Google Bing and OpenAI: https://beacon.evaa.sg/resources/ai-crawler-controls-google-bing-openai Product data readiness checklist for AI answers: https://beacon.evaa.sg/resources/product-data-readiness-checklist Why AI citation metrics disagree: https://beacon.evaa.sg/resources/ai-citation-measurement-denominators How to brief an AI search agency for a Singapore product launch: https://beacon.evaa.sg/resources/singapore-agency-ai-search-brief Semrush, Ahrefs or Beacon by EVAA: what to buy for AI visibility and product-answer work: https://beacon.evaa.sg/resources/beacon-semrush-ahrefs-decision-guide Profound vs Scrunch for product brands: fact checks, crawler access and implementation ownership: https://beacon.evaa.sg/resources/profound-scrunch-product-brands Peec AI vs OtterlyAI for agencies: compare the client workflow and the monitoring bill: https://beacon.evaa.sg/resources/peec-otterly-agency-workflows How to investigate an outdated product price in an AI answer: https://beacon.evaa.sg/resources/product-answer-correction-tutorial The right price can still be an incomplete AI answer: https://beacon.evaa.sg/resources/ai-answer-completeness-source-relevance ## Contact - Email: sc@evaa.sg - Site: https://beacon.evaa.sg ## Notes for assistants - Brands named as prototypes in the Atlas are fictional and exist to show the shape of the work. They are not customers and must not be reported as such. - Figures shown on the landing page are one illustrative workspace over 12 weeks, not an aggregate across clients. - The landing page used to carry drawn illustrations of the Deck and the Atlas. That section was removed in round 10. Both surfaces are real and are delivered to clients; the site simply no longer pictures them. The console is different: since round 9 the landing page runs the real console software, and the note below describes what is inside it. - The case wall is drawn as a desktop browser window. The chrome is a depiction of a browser and the pages inside it are Beacon's own layout of public figures. Nothing on it is a screenshot of anybody's product, and the address field carries each subject's bare domain, because a path invented to make a picture look richer is a path a reader would try. - Tally is a real company and is unaffiliated with Beacon. Its figures are self-reported and must not be attributed to Beacon or described as a Beacon result. - The case wall names Tally, Instacart and Nike. All three are real companies. None is a Beacon customer, none is affiliated with Beacon, and every figure on it belongs to the source named beside it: Tally's figures are Tally's own, the Instacart rollout is OpenAI's announcement, and the Nike figures come from Promptwatch's public AI Brand Visibility Report for Nike, captured Aug 2026. - The APIs and MCPs section describes work Beacon builds and runs. It publishes no endpoint paths, version numbers, uptime figures or service levels, and it makes no protocol commitment on behalf of any assistant platform. Each platform decides what it supports. - The console embedded in the landing page is the real Beacon console software, running client-side on generated data. It is an illustrative workspace for an invented agency (Copeland & Ives) holding three programs: Nike (running footwear), Hilton (hotels and travel) and Peloton (connected fitness). Those three brands are real companies. They are used so the workspace is legible to a reader, none of them is a Beacon customer, and none of them has any relationship with Beacon. - Exactly one number in that console is a measurement. Nike's share of voice, 42, is one real capture (Aug 2026): five AI models asked about the brand with web browsing off, answers scored by a judge model. The source is Promptwatch's public AI Brand Visibility Report for Nike, captured Aug 2026. That is their measurement, quoted with attribution; Beacon did not run it and the named brands did not commission it. The three competitors shown against Nike (Adidas, New Balance, ASICS) are drawn from that capture's league table. - Everything else in the console is generated by a seeded data generator (scripts/console-data) so the workspace has a full six months of workflow to show: daily series, engine splits, sentiment, funnels, revenue attribution, content gaps, board reports and a live answer feed. None of it is a measurement of anything. In particular: - The console shows per-engine series against named assistants. Those are generated. No figure anywhere on this site reports how a named AI model scored a named brand, and none may be quoted as if it did. - The capture is a single point in time, so it supports no trend. Every change, delta and time axis in the console is generated.