Updated on October 7, 2026 to add GPT-6.1 Sol, which was released this day but missed while the news harvest was offline.
Confirmed
Amazon Blocks Meta's Muse Shopping Agent Over Undisclosed Browsing and Credential Storage
Amazon began blocking Meta's Muse shopping assistant from Amazon.com on the night of September 20, 2026, after Meta declined to voluntarily exclude the site. Amazon said Muse doesn't identify itself while browsing and appears to capture and store customer credentials. Spokesperson Lara Hendrickson added that agentic purchasing tools must 'operate openly and respect service provider decisions.' Amazon's robots.txt, per Search Engine Journal, names more than 99 blocked crawlers but never covered Muse, which carries no declared user-agent to list.
In plain termsrobots.txt is a text file a site publishes to tell automated visitors which pages they're allowed to enter, and it only works on visitors that say who they are, so it had no way to stop an assistant built not to.
Why it mattersRobots.txt only binds crawlers willing to name themselves, so retailers are now policing agentic shoppers through account terms and behavioral detection instead of the crawl file.
Our takeI made a version of this point in 'Blocking AI crawlers did not stop AI citations': a control only reaches the traffic it's pointed at, and Amazon's own robots.txt already lists 99-plus crawlers it does bind. Muse sidesteps the file entirely by carrying no declared identity, the sharper failure mode behind the same mismatch.
Shopify Extends WebMCP Into Checkout, Letting Browser Agents Submit Orders
Shopify shipped WebMCP tools for its checkout experience on September 28, 2026, letting browser-based agents read checkout details and update contact information, shipping, payment method, and discount codes on eligible stores. Agents can submit an order only after the buyer confirms through Shop Pay; they cannot change line items or add new card details. Support is limited to standard three-page Shop Pay checkouts on Chromium browsers, excluding B2B, embedded, and mobile SDK flows.
In plain termsWebMCP is a set of hooks Shopify built into its checkout pages so a browser-based AI assistant can read and fill in the same fields a shopper would, then place the order once the shopper approves it through Shop Pay.
Why it mattersMerchants running Shop Pay checkout now have a live agent-submission path to audit, since Shopify states plainly that a Web Bot Auth signature is not a substitute for the buyer's own confirmation.
Our takeI've written that an MCP server is really a credential boundary dressed up as a protocol, in 'An MCP Server Is a Supply-Chain Dependency With Reach Into Your Data.' Shopify draws that boundary at the most consequential point possible, payment, and is explicit that a bot-authentication signature doesn't replace the buyer's own Shop Pay confirmation.
Google Appeals EU Order to Share Search History Data, Citing Privacy Harm
Google filed an appeal with the EU's General Court on September 29, 2026, challenging two Digital Markets Act orders: one requiring it to share anonymized search-and-click data with rival search engines and AI chatbots starting January 2027, the other requiring Android changes by July 2027 to open Gemini's assistant slot to competitors. Google's Oliver Bethell said the mandate lacks 'sufficient anonymisation' and would cause 'irreversible harm to user privacy.'
In plain termsThe Digital Markets Act is the EU law forcing the largest tech platforms to open parts of their systems to competitors, and this specific order would make Google hand rival search engines a stream of anonymized data about what people searched for and clicked.
Why it mattersIf the appeal fails, competing search engines and AI assistants gain a standing feed of anonymized Google query data starting January 2027, which resets the baseline for anyone benchmarking share against Google.
Our takeI covered Google's own admission to Reuters that EU compliance already produced its worst quality drop in 29 years, in 'Google's results are now geographically forked, and Google said so.' This appeal is the same fight one deadline later: Google contesting the next compliance requirement before it has to absorb it, rather than after.
Google Rolls Out AI Mode Information Monitoring to All Search Users Globally
Google's VP of Product for Search, Robby Stein, announced on September 28, 2026, that AI Mode's information-monitoring capability, previously limited to Ultra and Pro subscribers, is now available to every Google user globally. Users tell AI Mode what to watch, and Search continuously checks sites, forums, social posts, real-time data, and its 60-billion-item Shopping Graph for changes, then surfaces updates. Stein's examples include new restaurants, local events, and back-in-stock or price-drop alerts.
In plain termsInstead of searching once and getting one answer, a user can now ask AI Mode to keep watching a topic, like a price or a local event, and Google will check back and tell them when something changes.
Why it mattersA standing monitoring task inside AI Mode gives Google a new recurring reason to re-query a page after the first crawl, so freshness and stock or price accuracy now matter beyond the one-time ranking moment.
Our takeThis turns AI Mode from a query-and-answer tool into a standing subscription to a topic, which raises the value of structured, frequently-updated data, like stock status, event dates, or price, over static page copy. Pages that update those fields promptly are the ones the monitoring loop keeps re-checking; pages that don't simply stop getting re-polled.
Google Tests Moving AI Overview Citation Cards From the Side Rail to the Bottom
Search Engine Roundtable reports a live Google test, spotted by Glenn Gabe on September 28-29, 2026, that moves AI Overview citation cards from the right-hand rail to a collapsed block beneath the answer, expandable through a 'show more' control. Inline link icons within the answer text stay unchanged. Google has tested variations of this layout before, in March 2026 and September 2025, and has not confirmed a rollout timeline or scope.
In plain termsA citation card is the small preview box AI Overviews shows for each source it cites, and this test buries those boxes below the answer text instead of showing them right beside it, so a reader has to click 'show more' to see them.
Why it mattersMoving citation cards below the fold of the answer changes how much effort it takes a reader to reach a cited source, which is the interaction AI-visibility tracking is actually trying to measure.
Our takeGoogle has run this same layout test twice before without shipping it broadly, so treat this as a signal to watch citation-card click-through in server logs if it reaches your traffic, not as a change worth building a fix around yet.
Databricks Ships Full-Text and Vector Search Extensions for Postgres, GA on AWS and Azure
Databricks announced general availability, on AWS and Azure, of two Postgres extensions built into its Lakebase database: lakebase_vector for approximate nearest-neighbor search and lakebase_text for BM25 full-text search. Databricks says the combination runs semantic, keyword, and hybrid search inside Postgres itself, replacing separate search infrastructure, at twice the throughput and a quarter the cost of cloud Postgres with pgvector. The post frames the launch around AI agents, which it says trigger thousands of concurrent retrieval requests per workflow.
In plain termsRAG, retrieval-augmented generation, is the pattern where an AI system looks up relevant text before answering, and this release lets that lookup happen in one query against the same Postgres database holding the data, instead of a separate search system.
Why it mattersSearch and retrieval infrastructure is consolidating into the database layer itself, so a RAG pipeline built on a separate vector store is now competing against a cheaper option living inside Postgres.
Our takeDatabricks is pitching this against dedicated vector databases on cost and operational simplicity, not raw capability, so the real test is whether lakebase_text's BM25 implementation holds up against a purpose-built search engine at scale, not the benchmark numbers in the launch post.
Snowflake Adds Claude Sonnet 5.5 to Cortex AI, Its Fastest Anthropic Model Yet
Snowflake announced Claude Sonnet 5.5 is now available through Cortex Inference on Snowflake account endpoints, in a post dated September 28, 2026. Snowflake cites Anthropic's own benchmark of 70.6% on Terminal-Bench 4.0, close to seven times Sonnet 5's score, and says the model generates output more than 30% faster. Snowflake positions it for bug fixes, code review, document creation, and chart analysis, with Cortex Agents and SQL-based Cortex AI Functions integration listed as coming soon.
In plain termsCortex is Snowflake's built-in layer for running AI models against data that's already stored there, so a company can use Claude without copying its data out to a separate AI provider first.
Why it mattersRunning Claude inside Cortex keeps enterprise query data from leaving Snowflake's governance boundary, which is the actual reason a company picks a warehouse-native model over calling the Anthropic API directly.
Our takeThe benchmark Snowflake leads with, a sevenfold jump on Terminal-Bench, is Anthropic's number for the model generally, not anything Snowflake measured inside Cortex itself, so it says more about Claude Sonnet 5.5 than about how it performs against warehouse-scale queries.
AI model launches, model API price changes and major assistant features, kept separate from the search news above. Every model story Pulse has covered is collected under LLM Models & Agents, and the monthly model comparison shows how they stack up.
Confirmed
OpenAI Releases GPT-6.1 Sol, Near-Astra Performance at One-Fifth of Astra's API Price
OpenAI released GPT-6.1 Sol on September 29, 2026, describing it as near-Astra intelligence for coding, computer use and professional work at one-fifth of GPT-6 Astra's standard API prices. It costs $2 per million input tokens and $10 per million output tokens, the same list price as GPT-6 Sol, with a 1,050,000-token context window. On the independent Terminal-Bench 4.0 coding leaderboard it scores 58.2%, against 49.4% for GPT-6 Sol.
In plain termsA token is a piece of a word and the price is what a developer pays per million of them, so GPT-6.1 Sol buys a higher coding score at the same cost as GPT-6 Sol.
Why it mattersTeams paying for GPT-6 Astra can now test whether a model at a fifth of the price clears their own quality bar, which is the comparison OpenAI's own documentation tells developers to run.
Our takeI argued in 'Sonnet 5 brings near-Opus agents at a fraction of the Opus price' that the price drop, not the top score, is what pulls autonomous agents into economic reach. GPT-6.1 Sol repeats that pattern one tier up: its full Terminal-Bench 4.0 run cost $600, against $3,300 for GPT-6 Astra at the same 58.2% score.