When the agent recommends: what MCP changes for brands and campaigns

When AI assistants recommend products, context beats cookies, verified brand facts matter and labelling must be clear. What brands need now and how visibility in ChatGPT, Claude, Gemini and Perplexity can be measured.

From the click to the answer For decades, digital advertising ran along a simple chain: someone searches, sees a list of results and ads, clicks and lands on a website. Marketing learned to measure, optimise and buy along that chain. Keywords, bids, landing pages, conversion pixels. That chain is gaining a new link. More and more people ask a language model instead of using a search engine: "Which mineral water goes with fondue?", "Which agency in Zurich knows Meridian?", "What do I need for an autumn hike?". The answer is not a list of results but a paragraph with two or three names. If you are in it, you get asked. If you are not, you do not exist in that moment. There is a second step. Assistants such as Claude or ChatGPT no longer draw only on what they learned in training; they fetch information through tools. The protocol an agent uses to call such tools is the Model Context Protocol (MCP). The current specification is version 2026-07-28, and it made the core stateless so servers can scale more easily. For marketing this means an agent can search a product catalogue, check prices, look up inventory and even create campaigns during a conversation, without a person ever opening a website. The large platforms have responded. Meta runs its own MCP server for ads, through which an agent can manage campaigns, ad sets and catalogues and pull reports. That is the execution layer, and it is quickly becoming standard equipment. The more interesting question for brands sits one level earlier: why does an agent recommend a particular product at all? What a brand has to provide for agents to recommend it An agent recommends what it can find, understand and back up. That leads to three requirements that are related to classic search engine optimisation, but not the same. 1. Structured product data. An agent does not read hero banners. It needs name, price, availability, attributes, category and where to buy, in a form it can process. The more complete the attributes, the more likely a product fits a specific question. "Water filter" is a weak signal; "water filter, countertop, no electricity, for limescale and chlorine, replacement cartridge every three months" is a strong one. 2. Verified brand facts. Language models confuse brands, invent founding years or assign retailers incorrectly. If you do not provide your facts yourself, you leave …

Author

Nexbid Marketing

Published

2026-11-08

Topics

Agentic, Ad Tech, Privacy, Research