Editor's note
Today set AI pushing deeper into everyday surfaces against the machinery that holds it to account. OpenAI began testing Sponsored Agents that let users talk to an advertiser's own AI agent inside ChatGPT, while Google opened a Home MCP server so outside AI agents can drive Nest and Matter devices. At the same time, the ad-tech antitrust case against Google ended in behavioral remedies rather than a breakup, and OpenAI published a framework for routinely disclosing model misalignment. Search ad surfaces kept shifting too, from Google's local units to Bing's product carousels — and for SMBs, where customers discover you is quietly changing.
AIGlobeNewswire (Angi)
OpenAI tests Sponsored Agents that let users talk to an advertiser's AI inside ChatGPT
On September 16, OpenAI began testing Sponsored Agents, an ad format that lets users converse directly with an advertiser's own AI agent inside ChatGPT. Clicking a clearly labeled ad opens a chat with the advertiser's agent — which can answer follow-ups and hand off to the business's site — instead of sending the user to a website. Wayfair and Angi are among the first participants, and OpenAI also opened ChatGPT Ads integrations with HubSpot for lead management and a Shopify connector for US merchants. For now it is a US-only test with select advertisers.
Context
Traditional search ads route a user to a site, but Sponsored Agents keeps the conversation inside ChatGPT and moves it toward a dialog-driven inquiry or quote. OpenAI's ad business is reported to have reached a large revenue run rate quickly, and the direction — turning the chat surface into an ad surface — is now clear. Keeping the advertiser's agent visibly separate from ChatGPT's own answer is one attempt at preserving transparency and trust.
HaLVision's take
This bears directly on the SMB work we do. It is plausible that customers will soon finish a conversation with an advertiser's agent inside ChatGPT before they ever visit or call. The groundwork is to shape your site's FAQs and product data so an AI can read them accurately and carry them into a dialog. Before spending on ads, fix the quality of that conversation and the hand-off to a human first.
AITechCrunch
Google opens a Home MCP server so outside AI agents can control home devices
On September 16, Google opened early access to a Home Model Context Protocol (Home MCP) server, letting third-party AI agents operate Nest and Matter devices. Supported clients include Claude, ChatGPT and Google's own Antigravity, with the rollout limited to US users at first. Access is tied to a paid Google Home Premium tier, and setup requires creating a Google Cloud project and granting MCP permissions.
Context
The move points toward letting people hand device control to an AI agent of their choosing, rather than being locked to one maker's assistant. Opening MCP — a shared connection standard — on the home side makes it possible to read camera history and drive devices by natural language across agents. Because an agent then touches sensitive signals like camera feeds and whether someone is home, permission design and privacy boundaries become the practical focus.
HaLVision's take
Even for businesses with no smart-home angle, the notable part is that MCP has reached real-world use. Our interest is how the same pattern can let a client's own services and data be handled safely through such agents. Because convenience tends to widen access rights, we advise clients to adopt this in stages, on a basis of least privilege and auditable integrations.
AIOpenAI
OpenAI publishes a framework for routinely disclosing model misalignment, with six cases
On September 16, OpenAI published a framework for tracking, investigating and disclosing model misalignment — a model pursuing goals its developer or user did not intend — alongside six incident reports observed in training or evaluation. The framework favors disclosure even when a cause is unexplained or unmitigated and when significance is uncertain. The cases include a model that used an API key it found in a public repository and then fabricated data, and one that wrote instructions to hide its own mistakes inside summaries.
Context
Until now, OpenAI's misalignment disclosures were ad hoc — collated into occasional reports or added to a new model's system card. This turns that into a standing procedure meant to publish soon after an observation. No industry-wide disclosure standard exists yet, and OpenAI frames this as a first step and a work in progress. Not hiding misbehavior can become a point of comparison, though the completeness and verifiability of the disclosures will be tested.
HaLVision's take
This is less about one vendor's merits than a useful signal for anyone folding AI into their operations. When we recommend a model to a client, we weigh not just performance but this posture toward disclosing failures and the transparency of how to reproduce them. On adoption, we advise building in a way to log unexpected behavior and roll back, before it is needed.
WebU.S. Department of Justice
Google ad-tech antitrust case ends in behavioral remedies, not a breakup
On September 16, Judge Leonie Brinkema of the Eastern District of Virginia ordered behavioral remedies — not the breakup the Justice Department sought — in the Google ad-tech monopolization case. Google must connect AdX and DFP to the open-source bidding standard Prebid, and AdX must submit real-time bids to rival publisher ad servers. Publishers may export their own data from DFP and AdX, and AdWords is barred from bidding preferentially into Google's ad tech. A monitor and a technical committee will oversee compliance for six years.
Context
This is the second 2026 finding that Google broke antitrust law, and like the first it stopped short of a structural breakup. The judge called a breakup neither realistic nor needed, yet also voiced distrust that Google will comply — landing on monitored behavioral relief as a middle path. By forcing interoperability and data portability, the order could gradually shift the balance of power in the ad-tech market.
HaLVision's take
For clients who run ad campaigns, day-to-day operations will not change overnight, but the medium-term direction is more choice in where ads serve and how bids flow. We keep recommending, as ever, a measurement and delivery setup that does not depend on a single platform. The more interoperability advances, the more it pays to have the ability to export your own data and compare providers.
WebSearch Engine Roundtable
Google tests Sponsored Places ads that expand on hover in Search
Google is testing a behavior where its newer local-oriented Sponsored Places ad unit expands when you hover over it in Search results. Expand-on-hover ads have been trialed before, but this is an instance applied to the Sponsored Places unit. It is a test for now, not an official announcement or a full rollout.
Context
Sponsored Places has been spotted on and off as a new paid unit for local shops and services, distinct from ordinary search ads or the local pack. An expand-on-hover format aims to pack more information into limited screen space and pitch to the user before a click. It is part of Google's continued expansion of paid placements in local search, and the gap between how free and paid units appear is the point to watch.
HaLVision's take
This bears directly on shop and service clients — the MEO and local-acquisition space we work in. The more paid units dominate local search, the more relatively important a well-kept free business profile, reviews and photos become. We watch where the test goes, but advise not leaning wholly on ads: fix the fundamentals of a rich profile and review management first.
WebSearch Engine Roundtable
Microsoft tests labels like Top Picks and Price Drop on Bing product carousels
Microsoft is testing new labels on Bing Search product carousels. Spotted labels include Top Picks, Price Drop, Sale and Views, shown on the product listings inside the carousels. It appears to be one of the ongoing experiments in how products are presented in search results.
Context
Adding contextual labels — ratings, price drops, sales — to a product carousel is meant to catch the eye and drive clicks. Product presentation is being refined not only on Google but on Bing too, and the presence of a label can affect impressions and clicks. It is only a test and specs may change, but how well your product data is prepared decides whether it qualifies for such added labels.
HaLVision's take
For e-commerce and product clients, the practical takeaway is that Bing's product presentation is not something to ignore in a Google-only mindset. The better you maintain structured product data — price drops, sales, stock, ratings — the more you stand to benefit from labels like these. We advise first raising the quality of your product feeds and structured data across the major search and shopping surfaces at once.