First Published: 27 October, 2026
In February 2025, when I was speaking at the Marketing Association's Brainy Breakfast, I described a customer who never visits your website. They tell an AI agent what they want, and the agent does the research, comparison, and purchase for them. A number of people looked unconvinced, and there were plenty of questions. Over a year later, two of the largest companies on earth have created easily accessible AI agents that do exactly this.
On the 8th of September, Meta released Muse. The online chatter has been crazy, and the App Store and Play Store have listed Muse as their top download. [4] Meta's own description is interesting: "It doesn't just answer questions; it actually does the work." [1] Muse opens a browser, fills in a form, negotiates on your behalf, and checks out with a single-use card from Link by Stripe, pausing for approval before a purchase. [1][2] At Meta Connect, it gained Best Buy, Gap, Sephora, Walmart, and Wayfair as retail partners, the entire Shopify catalogue, and a business model. Mark Zuckerberg expects Meta to "profit by taking a small fee from transactions". [3] Interesting fact, the agent's physical avatar is called Jolly. [3] Sensor Tower counts 10.2 million US downloads in Muse's first 40 days, a milestone only 11 other apps have reached, the most recent one being Threads. [4]
On the 29th of September, OpenAI launched Dots. A dot is an always on agent inside of ChatGPT that runs on GPT-6 Astra. It's got its own cloud computer and browser, connects to more than 4,000 apps, and keeps working long after you close the chat. The first one is included in Pro and Business Premium plans. [5][6] Behind them sit Instinct, a text message agent reportedly valued at US$10 billion within six weeks of launch, and Google's agentic checkout inside Gemini and AI Mode. [7][8]
Muse is currently only available in the US and Canada, but it's coming our way. [4] Dots need a paid plan. Currently, nobody in New Zealand is losing a sale to a dot this week. However, the reason to act now is because the agents are coming faster than websites are being made ready for them. In this article, I'll detail some of the fixes that you need to implement that don't take very long at all.
How an agent reads your website
Meta has published a lot of information about this. Its crawler documentation says, "Allowing Meta-WebIndexer in your robots.txt file helps us cite and link to your content in Meta AI's responses." [9] In its safety write-up, it says the sub-agent that drives the browser "sees an accessibility tree snapshot of the page", not the underlying code. [2] What this means to you and me is the agent reads your site the way a screen reader does. Unlabelled buttons, a size chart that is only an image, a price that only appears via a script, and the agent is stuck. One person tested a Shopify store where six of the seven sizes were visibly sold out. The accessibility tree showed seven identical unmarked buttons. [10]
Early tracking by Centium finds Muse searches the web on nearly every prompt, reads roughly 25 results, and credits about one source per answer. It favours specialist publications and review sites. [11] This squares with Similarweb's recent finding that only 6.8% of ChatGPT answers carried a citation in May 2026, while recommended brands saw two to four times as many subsequent visits as competitors that were not. [12] Your website is being read far more often than it's being credited for, and this is what decides what purchasers buy.
Then looking at Amazon. On the 20th of September, it blocked Muse from shopping on its site with a pop-up telling the agent that "Continued access by an unauthorized AI agent violates Amazon's Conditions of Use". [13] It's apparent that the largest retailer on earth would rather be the agent than host one. Most New Zealand businesses are not so lucky and don't have this option.
Being mentioned is not a recommendation
In July, we asked 76 New Zealanders to put three questions into ChatGPT, Google AI Mode and Copilot on their own accounts. From the 888 answers we got, Electric Kiwi was mentioned in 97% of power answers and named first in 25%. Meridian was named in 70% and first in only 1%. [14] When a human reads the answer, a mention still earns you something. When an agent executes, it's only the first pick that matters. The same study found Canstar named in 66% of power answers and Consumer NZ in 63%. [14] An AI agent doesn't trust what you say about yourself. It relies on what the world says about you, third-party data. And in a market this size, the world is a handful of institutions.
What to audit, what to change and what to measure
Audit. Read your robots.txt and bot protection rules. Allow meta-webindexer and OpenAI's OAI-SearchBot and ChatGPT-User, unless you've decided otherwise for a particular reason. [9] Your bot protection vendor should tell you whether it can tell a customer agent from a scraper. If you sell on Shopify, open Sales channels and then Agentic. You may already be in Meta's connector tier without knowing. [10] Then open Chrome DevTools on your checkout and read the accessibility tree yourself.
Change. Make sure you put price, stock, shipping cost, delivery times, and returns in visible text on the page, not only in the schema or behind scripts. Structured data is not in the accessibility tree. [10] Use real buttons with real labels and offer guest checkout. Earn the proof the agent will go looking for. Think specialist reviews, awards, Consumer NZ, Canstar and other relevant websites. Complete your Instagram and Facebook business profiles and catalogues, because Meta said its AI will cite content shared on Instagram, Facebook and Threads "with credit back to the content creators". [15] If you have a public API, apply for a Muse connector. Meta says it wants Muse to work with small businesses too. [3]
Measure. Attribution is important too. Meta says that when Muse browses, "it will appear as your activity", so an agent visit looks much like a human in your analytics, and a connector order never touches your pages at all. [10] Reconcile agent orders from your platform's order records. Move the dashboards from share of mention to share of recommendation. Watch branded search because Scrunch's panel found first named brands 389% more likely to be searched after an AI recommendation. [16]
For 24 years, we've optimised for a human with a mouse. The next customer has neither, and has no emotion. Make sure your website allows them to understand, access, and purchase what you offer.
Category
.svg "MA-Logo (1)")
