Editor's note
Today's digest reflects AI moving from experimental to everyday infrastructure. Google and Microsoft are baking AI into advertising, transcription, and search measurement as defaults rather than options. Google's Search Console now surfaces AI visibility as a first-class metric, signaling that LLMO sits beside traditional SEO on the priority list. Meanwhile, the US-EU regulatory split means businesses operating globally must navigate two incompatible frameworks simultaneously — making AI tool selection a compliance decision as much as a capability one.
AIGoogle DeepMind
Google Launches Gemini 3.8 Flash — Higher Performance, but Prices Double in January
Google released Gemini 3.8 Flash on September 2 at the same price as its predecessor — $0.75 per million input tokens and $3.75 per million output tokens. The model surpasses Gemini 3.7 Flash across every benchmark Google published, and supports a 1-million-token context window with 64K output. However, both input and output pricing will double on January 1, 2027, and Google itself recommends staying on 3.7 Flash for efficiency-first workloads.
Context
Gemini 3.8 Flash is a derivative of 3.7 Flash rather than a new base model — it achieves performance gains by running more reasoning steps on the same weights, which means it consumes more thinking tokens per request. For document-heavy workflows such as long-form analysis and coding agents, it completes more than three times as many tasks as 3.7 Flash. A companion model, Gemini 3.8 Flash Cyber, was released simultaneously for cybersecurity-specific tasks including vulnerability detection and automated patching.
HaLVision's take
The most actionable takeaway is the pricing timeline, not the benchmark headline. The January 1 price doubling means applications running stably on 3.7 Flash should not migrate without a clear performance need. If you run document analysis or agentic pipelines with large context windows, evaluate whether 3.8 Flash's throughput gains justify the higher per-token cost before pricing changes — and lock in current rates where possible.
AIMicrosoft AI
Microsoft Launches MAI-Transcribe-2 — 60 Languages, $0.10/Hour, Industry-Leading Accuracy
Microsoft AI released MAI-Transcribe-2 on September 3, a speech recognition model priced at $0.10 per audio hour — a 72% reduction from the $0.36 charged by its predecessor five months ago. The model tops the FLEURS benchmark across all 60 supported languages with an average word error rate of 5.2%, and is ten times faster than OpenAI's GPT-Transcribe. Standard features include speaker diarization, word-level timestamps, and keyword biasing for domain-specific terminology.
Context
The speech recognition market is contested by OpenAI's Whisper family, Google's Chirp, and ElevenLabs ScribeV2. MAI-Transcribe-2 claims the top FLEURS position across all 60 languages while undercutting rivals on price. Speaker diarization and word-level timestamps are included as defaults rather than paid add-ons, which meaningfully lowers the barrier for business use cases like interview transcription, meeting notes automation, and multilingual subtitling. Microsoft released the model under its Microsoft AI brand rather than Azure Cognitive Services, signaling a developer-facing API-first strategy.
HaLVision's take
For SMBs currently transcribing meetings or customer interviews by hand, MAI-Transcribe-2's $0.10/hour pricing with speaker diarization included is worth a direct test. Japanese is among the 60 supported languages, but accuracy on less-resourced languages should be validated with sample audio before committing to production use. Clients already on the Azure ecosystem will find API integration straightforward; starting with a narrow meeting-notes automation pilot is the lowest-risk entry point.
WebGoogle Ads Blog
Google Auto-Migrates Search Campaigns to AI Max — No Rollback Option
Starting September 1, Google is progressively migrating Search campaigns that use Campaign-level Broad Match or Automatically Created Assets to AI Max for Search. Existing brand inclusions and exclusions carry over automatically, but no rollback is available after migration. Google claims an average 7% increase in conversions at a comparable CPA or ROAS, though independent testing across more than 250 retail campaigns has recorded a 35% ROAS drop in some cases.
Context
AI Max for Search uses AI to control search term expansion, ad text customization, and final URL expansion simultaneously, designed to work alongside Smart Bidding. The auto-upgrade rewrites campaign settings without explicit advertiser opt-in — which matters most for campaigns with strict brand messaging requirements or fixed landing page rules. Individual levers are retained: text customization and URL expansion can each be disabled, and negative keywords carry over. The Dynamic Search Ads forced migration, originally scheduled for September, was extended to February 2027.
HaLVision's take
If you manage Google Ads for clients, open the interface now and check whether migration notices have appeared on affected campaigns. For e-commerce clients where ROAS is the primary success metric, monitor performance weekly against pre-migration baselines — a 35% swing is material. If AI-generated ad copy conflicts with brand guidelines, disable text customization in the AI Max settings. The February 2027 DSA deadline also means now is the right time to begin planning the transition to Performance Max.
WebGoogle Search Central
Google Search Console AI Visibility Reports Now Available Worldwide
Google rolled out its Generative AI Performance reports in Search Console to all sites globally on August 31. The reports show impressions within AI Overviews, AI Mode, and Discover's generative AI features, broken down by page, country, device, and time period. The feature launched for a limited set of sites in June and has now reached global availability. Click data is not included in the current version.
Context
The inclusion of a dedicated AI performance section in Search Console signals that Google is treating AI-surface impressions as a first-class SEO metric alongside traditional organic rankings. As AI Overviews appear increasingly at the top of search results, which pages get cited within AI features — not just which pages rank — is becoming a distinct optimization problem. The absence of click data limits direct traffic attribution, but impression data provides a usable proxy for measuring how well a site's content is structured for AI referencing — what practitioners are calling LLMO.
HaLVision's take
If you have not opened Search Console this week, check now whether the Generative AI performance section has appeared in the left-hand menu. Low impression counts point to opportunities: structured data markup, clear heading hierarchies, authoritative authorship signals, and direct answers to question-intent queries are all LLMO-relevant improvements. Since click data is absent, use impression performance to identify which content patterns AI features prefer, then replicate those patterns across other pages on the site.
AIAl Jazeera
US and EU Diverge on AI Governance — US Pushes Looser Rules at G20 as EU Begins High-Risk Audits
On September 2, the United States hosted a G20 ministerial in Chapel Hill, North Carolina, and argued against AI-specific regulations, urging G20 members to avoid imposing binding rules. The EU AI Act's full set of high-risk AI obligations became enforceable on August 2, with the EU AI Office now actively auditing high-risk systems. In the US, federal preemption legislation remains stalled in committee, leaving state-level laws — led by California — to fill the regulatory gap.
Context
The EU AI Act classifies AI systems by risk level and imposes documentation, risk management, human oversight, and data-governance requirements on systems deemed high-risk — including those used for hiring, credit scoring, and medical diagnosis. The US, after scaling back prior executive orders on AI, has adopted a light-touch federal posture and is actively discouraging other nations from adopting binding AI rules. For companies operating AI products globally, this split means navigating two incompatible compliance frameworks simultaneously: EU obligations that mandate human oversight and documentation, versus a US environment where voluntary frameworks predominate.
HaLVision's take
This regulatory divergence is a macro story for now, but it has real implications for vendor selection. When evaluating AI tools for clients — especially those with European customers or partners — ask vendors explicitly how their products address EU AI Act obligations. In the medium term, the cost of maintaining dual compliance stacks will favor AI vendors that invest in global certifications. For SMB clients, choosing tools from vendors actively pursuing EU compliance is low-cost future-proofing.