All issues
Discovery Digest · September 14, 2026
Issue 16. The week AI embedded in ads, identity, and the search page
TL;DR
This week, two platforms claimed territory that previously belonged to brands and advertisers. Google completed the mandatory migration of all Campaign-level Broad Match and ACA Search campaigns into AI Max, removing opt-out options entirely, and expanded generative UI from AI Mode into standard AI Overviews, turning the search results page into a functional tool environment. OpenAI launched GPT Image 2.5 as a two-model production system with 50% lower latency and doubled pricing, and went live with Sign in with ChatGPT, moving from AI assistant to identity provider alongside Google and Apple. Meta released Muse Spark 1.3, its most efficient agentic model to date, completing comparable engineering tasks with 20% fewer tool calls and 25% fewer tokens.
01 · Google Search
Google forces all Broad Match and ACA Search campaigns into AI Max throughout September
- What
- From 1 to 30 September 2026, Google is automatically migrating every Search campaign running Campaign-level Broad Match or Automatically Created Assets to AI Max. Campaigns are upgraded in place using what Google describes as equivalent AI Max settings, with existing brand inclusions and exclusions carried over. For ACA campaigns, text customisation defaults on after migration; for Broad Match campaigns it defaults off. No opt-out exists. Google stopped allowing new Campaign-level Broad Match and legacy ACA campaign creation from 3 August 2026. Dynamic Search Ads follow in a separate migration scheduled for February 2027.
- When
- Migration window opened 1 September 2026. Announced via the Google Ads Help Centre and reported by Search Engine Land on 18 August 2026.
- How it shifts discovery
- AI Max moves control of query expansion, ad copy generation, and eventually landing-page selection from the advertiser to Google's models. The framing of the migration as using 'equivalent settings' understates the practical shift: campaigns previously bounded by explicit keyword lists or hand-authored asset sets are now governed by an AI that can expand query matching beyond those parameters. For brand safety and competitive exclusion, the audit question is whether the brand inclusions and exclusions that carried over are specific enough to constrain AI Max's query expansion, because 'equivalent' is Google's interpretation. Paid search teams that have operated with granular control over match types and ad copy are now operating inside a system where those controls are mediated by a model. The key governance question is not what changed at migration but what the model will decide in the next billing cycle without the advertiser's input.
- Questions to ask
- Every Campaign-level Broad Match and ACA campaign in our account has now been migrated to AI Max. Have we audited the brand inclusions, exclusions, and negative keyword lists that carried over to confirm they are specific enough to constrain AI Max's query expansion, and has a named owner reviewed the first post-migration performance data?
- AI Max controls query matching, ad copy, and eventual landing-page selection in ways the previous campaign types did not. For campaigns running in regulated categories, sensitive verticals, or markets with strict brand guidelines, have we tested whether AI Max's decisions stay within our compliance boundaries or whether additional restrictions need to be applied?
- Dynamic Search Ads migration to AI Max is scheduled for February 2027. For teams currently relying on DSA campaigns, does that timeline appear in our 2027 planning, and is there a named owner responsible for auditing DSA performance under current settings before the migration window opens?
- Sources
02 · OpenAI
OpenAI launches GPT Image 2.5 with two purpose-built models and 50% lower latency
- What
- OpenAI released GPT Image 2.5 on 8 September 2026 as two distinct API models: gpt-image-2.5-flare, optimised for fast everyday production at up to 50% lower latency than GPT Image 2, and gpt-image-2.5-sunburst, built for editing precision on work where quality matters more than speed. Both models support image generation, multi-turn editing, and multi-image reference inputs. New capabilities across both include a sketch drawing tool, comment-based image editing, and output up to 4K resolution. API pricing doubles the GPT Image 2 rate: $5 per million text input tokens, $8 per million image input tokens, and $30 per million image output tokens.
- When
- Announced on openai.com, Mon 8 September 2026. Reported across developer and AI publications on 8 to 12 September 2026.
- How it shifts discovery
- The two-model architecture is the commercially significant decision. Splitting fast iteration (Flare) and editing precision (Sunburst) into separate API models means teams building production image workflows now need to decide which stage of their pipeline each model serves, rather than choosing a single model and accepting its trade-offs. The doubled pricing versus GPT Image 2 sets an explicit cost tier: teams that benchmarked GPT Image 2 for budgeted production workloads need to remodel unit economics before migrating. The sketch tool and comment-based editing move the interface closer to collaborative creative review workflows, a pattern familiar to design teams. For brands running advertising creative through Advantage+ or third-party tools that call GPT Image 2, the quality ceiling and the cost floor both moved this week. Neither change is automatic: both require a conscious decision before the next campaign cycle.
- Questions to ask
- GPT Image 2.5 pricing is double the GPT Image 2 rate. For every production workflow currently calling GPT Image 2, have we recalculated the cost per asset at the new rate and confirmed whether the quality improvement from Flare or Sunburst justifies the increase, or whether GPT Image 2 remains the right choice for those tasks?
- The two-model structure separates fast iteration (Flare) and precision editing (Sunburst) into distinct API calls. In our image production pipeline, have we mapped which stages require speed and which require precision, so we are calling the correct model at each step rather than defaulting to one for the entire workflow?
- GPT Image 2.5 adds a sketch tool and comment-based editing, moving toward collaborative creative review. For design and content teams currently using separate tools for creative feedback and iteration, does the new interface reduce the number of tools in the workflow, and is there a named owner assessing the consolidation case?
- Sources
- OpenAI, GPT Image 2.5: Flare and Sunburst, openai.com, 8 September 2026
- Kanaries, GPT Image 2.5: How to Use It, Flare vs Sunburst, and API Pricing, September 2026
- AiCybr Blog, OpenAI GPT-Image-2.5 API: Flare vs Sunburst, Pricing, Resolution and Editing, September 2026
03 · Meta
Meta releases Muse Spark 1.3 with 20% fewer tool calls and improvements in agentic coding
- What
- Meta released Muse Spark 1.3 on 2 September 2026, rolling out to the API and Muse Code. The model completes comparable engineering tasks with roughly 20% fewer tool calls and 25% fewer tokens than Muse Spark 1.2. Improvements cover long-running agent task management, multi-workflow coordination, adherence to complex instructions, and user confirmation handling. Resistance to adversarial inputs and prompt injection has been increased. A max-reasoning variant is in development pending additional safety testing. The model is deploying to Meta AI across Instagram, Facebook, and WhatsApp in the weeks following API availability.
- When
- Released Tue 2 September 2026. Reported by Bloomberg and Axios on 2 September 2026.
- How it shifts discovery
- The efficiency gains are the commercially relevant headline, not the benchmark ranking. A model that completes the same agentic tasks with 20% fewer tool calls and 25% fewer tokens reduces cost per task at exactly the layer where agentic workflows accumulate cost: multi-step tool orchestration. For teams evaluating agentic models for coding, content production, or research workflows, Muse Spark 1.3 positions Meta's API as a substantive option alongside GPT-5.6 Terra and Grok 4.6 at current pricing. The deployment pattern is the other signal worth tracking. Meta releases to developer API first, then deploys to its consumer platforms: that means the AI model writing and serving advertising creative, answering product questions, and handling brand interactions across Instagram, Facebook, and WhatsApp is Muse Spark 1.3 within weeks of this release, not a version several iterations behind. For brands with significant Meta advertising and community presence, the model underpinning their AI-mediated interactions just changed.
- Questions to ask
- Muse Spark 1.3 completes agentic tasks with 20% fewer tool calls and 25% fewer tokens than 1.2. For any agentic workflow we currently run on GPT-5.6 Terra or Grok 4.6, have we benchmarked Muse Spark 1.3 against our specific tasks and documented the performance-to-cost comparison before committing another quarter's spend to a more expensive alternative?
- Muse Spark 1.3 is deploying to Meta AI across Instagram, Facebook, and WhatsApp in the weeks after its API release. Have we tested how Muse Spark 1.3 represents our brand, products, and category in Meta AI interactions since the update, and has the quality or framing of those outputs changed materially compared with the Muse Spark 1.2 baseline?
- Meta's API-first deployment pattern means Muse Spark 1.3 will power advertising creative generation and community AI interactions on Meta's platforms within weeks. Has our Meta advertising team been briefed on the underlying model change, and does our brand safety review process cover AI-generated creative output that has been produced by an updated model?
- Sources
04 · Google Search
Google expands generative UI into standard AI Overviews, building interactive tools inside search results
- What
- Google began rolling out generative UI to AI Overviews on 19 August 2026, extending a capability previously available only inside AI Mode. Generative UI builds custom visual layouts, interactive tools, and functional calculators directly inside the AI response for relevant queries. A science query can return an interactive diagram; a personal finance query can surface a working calculator; study queries return practice quizzes with explanations. Study notebooks and quiz generation across multiple subjects and standardised tests are available globally in English at no charge. The feature requires no user action to enable.
- When
- Confirmed by Google on Tue 19 August 2026. Reported by Search Engine Journal and ChromeUnboxed on 19 to 20 August 2026.
- How it shifts discovery
- Generative UI changes the SERP from a list of links into a functional environment. An AI Overview that builds and embeds a working calculator or interactive diagram does not need the user to click anywhere to complete the task. The zero-click outcome is no longer just a text answer absorbing the query: it is a functional tool completing the task. For brands in categories where an interactive tool is the conversion driver (mortgage calculators, tax estimators, dosage guides, product configuration selectors), the implication is that Google may now surface a functionally equivalent tool inside the AI Overview before the user reaches the brand's own page. The strategic response is not to build a better calculator but to ensure the brand is cited or present inside the Overview that contains the interactive output. The expansion to standard AI Overviews from AI Mode is the critical transition: this feature now applies to the search results every user sees by default, not only to AI Mode sessions.
- Questions to ask
- Generative UI can build functional calculators, interactive diagrams, and practice tools directly inside AI Overviews for standard queries. For the highest-traffic queries in our category, have we tested what Google now builds generatively for those searches, and does any generative output compete with or displace our own interactive tools or conversion assets?
- The feature is live globally in English and requires no user opt-in. Have we updated our AI Overview monitoring to cover generative UI outputs alongside text answers, and is there a named owner reviewing whether our brand or content appears inside the interactive outputs Google is generating for our category?
- For brands that have invested in proprietary interactive tools (calculators, quizzes, configurators) as lead generation or conversion assets, does a Google-built equivalent appearing in an AI Overview before the user clicks change our investment case for maintaining and distributing those tools independently, and have we quantified that exposure?
- Sources
05 · OpenAI
OpenAI launches Sign in with ChatGPT, positioning the platform as an identity provider for the web
- What
- OpenAI launched Sign in with ChatGPT in beta on 2 August 2026, establishing ChatGPT as an identity provider alongside Google, Apple, and Microsoft. The system lets users create, link, or access accounts on supported external applications using their ChatGPT account. Six named launch partners: Airtable, GitLab, HubSpot, Notion, Supabase, and Vercel. Partner applications receive the user's name, email address, and profile picture at sign-in. OpenAI retains the user's IP address, device details, and a record of every application the user authenticates to as identity provider. The feature is available globally to all authenticated ChatGPT users, including Enterprise accounts.
- When
- Launched in beta Sun 2 August 2026. Reported by RuntimeWire and TechTimes on 2 to 3 August 2026.
- How it shifts discovery
- An identity provider is the system a user trusts to vouch for them across the web. Google and Apple built that position over years by controlling the login button on millions of applications. OpenAI's six launch partners establish the deployment pattern, not the scale. Each partner that integrates Sign in with ChatGPT routes authentication events through OpenAI's identity layer, giving OpenAI visibility into which platforms its user base authenticates to, at what frequency, and from which devices. The data relationship this creates is broader than what partner applications receive: partners get name, email, and profile picture; OpenAI retains the cross-platform authentication graph. For brands building or integrating applications that touch ChatGPT users, the identity layer changes the context that surrounds the user before they arrive: a user who signs in with ChatGPT carries an identity that OpenAI has already cross-referenced with their session history and preferences. The question for every digital team is not whether to support Sign in with ChatGPT but whether the data relationship that comes with it is within your application's privacy architecture.
- Questions to ask
- Sign in with ChatGPT gives OpenAI visibility into every platform its users authenticate to, while partner applications receive only name, email, and profile picture. For any application in our stack that we are considering integrating with Sign in with ChatGPT, have we reviewed what OpenAI retains as identity provider against our data handling obligations and user privacy commitments?
- The six launch partners span project management, developer tools, and productivity software. Which applications in our organisation's technology stack are candidates for Sign in with ChatGPT integration, and is there a named owner responsible for assessing the authentication, data flow, and compliance implications before any integration decision is made?
- OpenAI as an identity provider gains a cross-platform session graph that grows with each new integration. As the number of applications supporting Sign in with ChatGPT expands, how does that position change our assessment of OpenAI's data advantage relative to Google and Apple as identity providers, and does it affect which AI partnerships we prioritise?
- Sources
Key takeaways
What to walk away with this week
- 01
Google's mandatory AI Max migration is now active for all Campaign-level Broad Match and ACA campaigns: the only safe response is a full audit of every brand inclusion, exclusion, and negative keyword list that carried over, before the model makes a buying decision you did not authorise.
- 02
Generative UI in standard AI Overviews turns the SERP into a functional tool layer: if any of your conversion assets is a calculator, configurator, or interactive guide, test what Google builds for those queries now, before your users find a Google-native version first.
- 03
OpenAI is now a model provider, an image production system, and an identity provider: each new layer deepens its data relationship with users and its position between your brand and its audience, so track every OpenAI touchpoint your customers use.
- 04
Meta's Muse Spark 1.3 completes comparable agentic tasks with 20% fewer tool calls and 25% fewer tokens than its predecessor: re-benchmark any workflow currently routed to GPT-5.6 Terra or Grok 4.6 before committing to those rates for another quarter.
- 05
OpenAI's identity layer and Google's generative UI both move platform infrastructure into spaces previously controlled by brands: audit what your team controls in advertising, identity, and interactive content, and what has been automated away this month.
