Observed
Google Tests AI-Generated Summaries Inside AI Overview Citation Cards
Search Engine Roundtable reported on October 23, 2026 that Google is testing AI-generated summary text inside the citation cards that sit beside AI Overview answers. SEO consultant Brodie Clark posted a screenshot showing card text that does not match the wording on the cited page, writing that Google is summarizing based on what it judges most relevant to the query rather than quoting it directly. The feature was not present on every query Barry Schwartz tested.
In plain termsA citation card is the small preview box inside an AI Overview that names and links the page an answer drew from; testing it with AI-written text means that preview may already be a paraphrase rather than a quote from the page.
Why it mattersA citation card that already paraphrases the source page means the text searchers compare against your copy is Google's summary of it, not your actual wording, which narrows how much control a publisher has over the first impression a citation makes.
Our takeWorth tracking against Search Console's citation reporting if Google ships this broadly, since a rewritten card means the link and the summary text say different things about what actually represented a brand in that answer. The gap stays invisible until someone compares the live summary to the source line by line.
Observed
Google's Universal Cart Expands From AI Mode Into Core Search Results
SEO consultant Glenn Gabe reported firsthand on October 17 that Universal Cart, the cross-retailer cart Google launched inside AI Mode and Gemini on October 15, now also appears in core Google Search results. A cart icon sits at the top of the results page, and product listings carry their own 'Add to cart' buttons that check out through the Universal Commerce Protocol without leaving Search. Google has not separately confirmed the core-Search rollout.
In plain termsUniversal Cart is a single shopping cart that holds items from multiple online stores and checks out through Google's own system instead of each store's checkout page; this update puts that cart inside regular Google Search results, not just the AI chat-style modes.
Why it mattersA live checkout surface now sits inside core Search rather than only the AI Mode and Gemini surfaces, so the share of transactions it pulls away from a retailer's own cart is a bigger question for mainstream Search traffic than it was two weeks ago.
Our takeI wrote in 'UCP Spec Breaking Changes' that the protocol had already shipped three backward-incompatible releases in eight months while merchants were still integrating it. Universal Cart moving into core Search multiplies the transaction volume riding on that same unstable spec, so a breaking change now risks checkout failures at mainstream Search scale, not just inside an AI Mode pilot.
Confirmed
Google Launches Business Home, A Unified Portal For Profile, Ads, Merchant Center And YouTube
Google's own support documentation confirms Business Home, described as 'a growth partner built with Gemini models,' that connects a business's Google Business Profile, Ads, Merchant Center and YouTube accounts into one dashboard. Barry Schwartz got invitation access and reported it October 22, 2026; the portal adds AI content and photo tools, review-response management, and a stated plan to flag demand surges and suggest seasonal offers. Google calls it a closed beta limited to select categories in the US and UK.
In plain termsBusiness Home is a new single dashboard where a business manages its free Google listing, paid ads, product feed and YouTube channel together instead of switching between four separate tools.
Why it mattersA single dashboard spanning Profile, Ads, Merchant Center and YouTube is Google consolidating the tools a multi-location or e-commerce business already juggles separately, so a beta invitation is worth accepting now to learn the new surface before it becomes the default.
Our takeThe Gemini branding and the demand-surge detection are the parts to watch once this leaves beta, since a tool that already sits across Ads, Merchant Center and Profile data has everything it needs to start making budget and offer suggestions on a business's behalf rather than just reporting on them.
Observed
Google Tests Redesigned Performance Reports For Business Profiles
Search Engine Roundtable reported on October 23, 2026 that Google is testing a redesigned performance-reporting section inside Google Business Profile, based on screenshots from two users, Jaikishan Raturi and Syed Amir. The test overview consolidates engagement, review, search-trend and audience metrics into one dashboard and adds a recommendations panel. Barry Schwartz called it 'a very big improvement' over the current report but had not seen it in his own account, and Google has not confirmed the test's scope.
In plain termsThe Performance tab in a Google Business Profile is where an owner sees how many people viewed, called or got directions to their listing; this test redesigns that tab into a single, more detailed overview with Google's own suggestions attached.
Why it mattersA consolidated metrics view with built-in recommendations changes what a client expects to see when they log into their own profile, so knowing this test exists before a client asks about it beats explaining an unfamiliar dashboard cold.
Our takeWorth distinguishing from Business Home in client conversations, since Schwartz notes the two overlap: one is a new standalone portal, the other a redesigned tab inside the profile a business already has, and a client who reads about one will ask about the other.
Confirmed
Anthropic Donates Agent Skills Standard To The Agentic AI Foundation
Anthropic announced on October 22, 2026 that it is transferring Agent Skills, the open standard it published in December 2025 for packaging an AI agent's specialized knowledge and workflows, to the vendor-neutral Agentic AI Foundation under the Linux Foundation. The donation includes the standard's name, logo, GitHub organization and agentskills.io domain. Anthropic reports over 70 platforms have adopted it, including ChatGPT, Cursor, GitHub Copilot and VS Code, with more than 1 million published skills and nearly 280 million installations.
In plain termsAgent Skills is the format that lets an AI agent load a packaged set of instructions and reference files for a specific job on demand, the same underlying idea whether a vendor calls it a Skill, a Gem, or a loop; handing its governance to an outside foundation means no single company can unilaterally change the spec everyone already built against.
Why it mattersA spec with 70 platform integrations moving to neutral governance instead of staying Anthropic's to unilaterally change is the signal that Agent Skills has become infrastructure worth a multi-year integration, the same test MCP passed before it got the same treatment.
Our takeI wrote in 'Skills, Gems, and Loops' that this packaged-procedure pattern is the same idea across Claude Skills, Google Gems and custom agent loops, with vendors mostly differing on naming. A convergent pattern used by 70 platforms needs one neutral spec instead of five incompatible ones, which is exactly the problem donating it to the AAIF is meant to solve.
Model & AI releases
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 Takes The Decisions API To General Availability, Built On GPT-6 Luna
OpenAI announced on October 22, 2026 that the Decisions API, built on GPT-6 Luna, is now generally available for low-latency, bounded-answer decisions such as classification and routing. OpenAI reports responses up to 10x faster than calling Luna directly through the Responses API. Since its beta, OpenAI added native GPT-Live integration, prompt caching, US and EU inference residency, native JSON inputs, and raised the question limit per request from 200 to 400.
In plain termsThe Decisions API is a stripped-down way to ask a model a narrow, multiple-choice-style question, such as which category a document belongs in, so it answers faster and cheaper than sending the same question through the model's full conversational API.
Why it mattersA dedicated low-cost endpoint for bounded decisions is the model-choice lever for routing, tagging or triage workloads a team is currently running on a full chat-completion call, worth re-testing against whatever model is doing that job today on cost alone.
Our takeThe jump from 200 to 400 questions per request and the added GPT-Live integration both point at the same target: real-time voice and agent workflows that need many small decisions per second rather than one big one, which is a different cost model than most teams are currently budgeting against.
Confirmed
OlmoEarth v1.3 Ships Precomputed Satellite Embeddings For The Whole Planet
The Allen Institute for AI published OlmoEarth v1.3 on Hugging Face on October 22, 2026, a satellite embedding model that turns a year of Sentinel-1 radar and Sentinel-2/Landsat multispectral imagery into 128-dimensional vectors for every 10-meter pixel on Earth, now precomputed for the full planet at roughly 150 TB per year. AllenAI reports state-of-the-art results on 20 of 27 benchmarks, including a new tropical-farming crop-type test, while cutting compute 53% versus v1.2 to about $2,720 in rented H100 time.
In plain termsAn embedding model does not answer questions directly; it converts raw data, here a year of satellite images, into a compact numerical fingerprint that a simpler, cheaper model can then use to classify what's in each spot on Earth.
Why it mattersA planet-scale embedding layer any team can query instead of training its own remote-sensing model is the model-choice decision for anyone doing geospatial classification work, like crop monitoring or land-use change, that currently justifies a custom model.
Our takeA 53% compute cut on the same benchmark set is the number worth testing against, more than the state-of-the-art claim itself, since it determines whether re-running embeddings on a new region is a weekend job or a line item.