# Pulse ยท August 28, 2026 > Google splits site-reputation enforcement by EEA location, AI Mode adds hotel booking, OpenAI ships WebMCP in ChatGPT, and Databricks previews chart retrieval. Canonical: https://brandonlazovic.dev/pulse/2026-08-28/ Author: Brandon Lazovic Published: 2026-08-28 ## Google stops enforcing site reputation abuse penalties for searchers in the EEA Status: Confirmed | Topics: organic-search-core Google will stop applying manual actions under its site reputation abuse policy to search results shown to users in the European Economic Area starting August 30, 2026, following discussions with the European Commission. Manual actions still register in Search Console and continue to affect rankings for users outside the EEA, but affected site sections inside the EEA may instead be separated algorithmically so they rank independently over time. In plain terms: A manual action is a human reviewer at Google penalizing a specific page or section for breaking the rules, separate from an automatic ranking change. From August 30, that penalty simply won't apply to what EEA searchers see, even though the same page still gets flagged for everyone else. Why it matters: This is Google's first location-based split in spam-policy enforcement, so EEA-based publishers now get a materially softer manual-action consequence than sites serving the rest of the world for the identical violation. Our take: This is Google's first acknowledged split in spam-policy enforcement by searcher location, and it came from DMA pressure, not a search-quality rationale. Expect this precedent to get tested. Any publisher facing a manual action now has a concrete argument for why enforcement should differ by region, and other regulators may treat the carve-out as a template. - [Google Search Central Blog: Update to the Site Reputation Policy](https://developers.google.com/search/blog/2026/08/update-site-reputation-policy) ## AI Mode adds hotel booking, flight price tracking and loyalty pricing in Search Status: Confirmed | Topics: ai-overviews-ai-mode, agentic-commerce Google added three travel capabilities to AI Mode in Search on August 27, 2026. Flight price-tracking alerts now arrive by email, and redemption pricing in points or miles gets pulled from airline and hotel loyalty partners including Alaska Airlines, American Airlines, Hilton and Wyndham. In-conversation hotel booking runs through partners such as Booking.com, Expedia and Marriott, paid via Google Pay, covering the US in English so far while the other two features reach more than 180 countries. In plain terms: Merchant of record just means the company that legally sells you the room or ticket and processes the payment. Booking.com, Expedia and the hotel chains keep that role. Google is only hosting the conversation where the booking happens, letting them complete the sale inside a Search conversation instead of on their own websites. Why it matters: Travel bookings can now complete inside AI Mode itself through partner merchants, which pulls the last click of a purchase onto Google's own surface even for queries that start as ordinary search. Our take: I argued in 'Your product feed doesn't know if you can deliver it' that the record deciding what an agent can actually book usually sits with an intermediary, not the merchant. This AI Mode rollout confirms the pattern again. Booking.com, Expedia and the hotel chains remain the merchants of record, while Google owns the conversation where the booking happens, so a hotel or airline optimizing only its own site's schema markup is optimizing a step that increasingly is not where the sale closes. - [Google: 3 new ways to plan and book travel in Search](https://blog.google/products-and-platforms/products/search/book-travel-ai-mode/) ## OpenAI ships WebMCP site tools in ChatGPT's built-in browser Status: Confirmed | Topics: llm-models-agents, chatgpt-assistants OpenAI's ChatGPT documentation confirms its built-in desktop browser now supports WebMCP, an experimental open standard that lets a website register JavaScript functions as discoverable 'site tools' with a name, description and input schema, instead of leaving an agent to guess at the UI. Selecting Site Tools in the address bar lists what a page offers, and each tool invocation gets a safety review before running. OpenAI is also running a WebMCP hackathon with Google Chrome, Shopify and Cloudflare as partners. In plain terms: Instead of an AI assistant clicking around your website's buttons and forms and hoping it lands in the right place, WebMCP lets your site directly hand the assistant a menu of actions it can call, the same idea as an API but built for a page a person is actively browsing. Why it matters: A browser-native tool-registration standard shifts the integration work from an agent scraping a page's markup to a site publishing its own callable actions, a new surface an SEO practitioner now has to audit alongside schema markup. Our take: I argued in 'An MCP server is a supply-chain dependency with reach into your data' that a tool's name and description are untrusted text a model reads before deciding what to do. OpenAI's own WebMCP documentation states the identical principle for site tools, warning that a tool's claim to only read data is not proof of what it does. The trust boundary just moved from an installed MCP server to any website you happen to browse with an agent, a much larger surface to reason about. - [OpenAI / ChatGPT Learn: Site tools (WebMCP)](https://learn.chatgpt.com/docs/webmcp) - [OpenAI: The WebMCP Challenge](https://openai.com/webmcp-challenge/) ## Databricks previews structured chart extraction to fix agent retrieval errors on charts Status: Confirmed | Topics: ai-data-stack, llm-models-agents Databricks is adding chart-aware retrieval to its data pipeline. A forthcoming update to ai_parse_document will convert a chart's visual data into structured JSON that sits alongside surrounding text in a retrieval index. In Databricks' own benchmark, a frontier agent reading a chart as an image took 50 seconds and miscounted its local maxima as 17, while the structured-JSON approach through Genie answered correctly (18) and scored 75.9% on the ViDoRe V3 benchmark. Databricks says the feature 'will be available soon.' In plain terms: A chart image alone forces an AI to guess at the numbers behind the lines and bars. Structured extraction turns a chart into a small table of the underlying data values the AI can read exactly, the same way it already reads text. Why it matters: If chart-reading errors like this are common across agent products, any page whose key numbers live only inside a chart image, not the surrounding text, risks being read wrong by an AI answer engine today, before any fix ships. Our take: I wrote in 'What RAG actually does' that retrieval quality depends on what actually sits in the index, not just the model reading it. This result is a clean demonstration: the same chart produced a wrong answer as a raster image and a right one once its data got extracted into structured JSON. It is also a narrow benchmark, two datasets and one vendor's own numbers, so treat the accuracy gain as a lab result until it ships and gets tested against a page you do not control. - [Databricks Blog: Enhancing Agent Retrieval with Structured Chart Extraction](https://www.databricks.com/blog/enhancing-agent-retrieval-structured-chart-extraction)