AI Search News September 2026: GPT-6 Arrives, ChatGPT Ads Hit $1B & Google's Opt-Out Goes Global

Key takeaways

OpenAI shipped GPT-6 twice in three weeks. GPT-6 Astra launched September 3 with "AGI era" framing from OpenAI itself and immediate safety scrutiny over its architecture; cheaper, faster siblings GPT-6 Sol and Luna followed September 22.
ChatGPT Ads crossed $1 billion in annualised revenue in under 200 days — announced August 31 alongside self-serve expansion to India, Europe, the Middle East, and North Africa, then a further expansion to Southeast Asia and Taiwan on September 23.
Google's publisher opt-out for AI Overviews and AI Mode went live for every website worldwide on August 31, months ahead of the UK regulator's own deadline — and on September 1, Google confirmed it's offering that same global opt-out as a move to resolve a separate EU antitrust probe.
Google sent publishers two contradictory signals in the same window. A new "Preferred Source" button lets readers pin favourite publishers across Search, Discover, and News — while Google Discover is separately testing a "Dive Deeper" feature that answers a story with an AI-generated overview instead of sending the click to a publisher at all.
Google's fourth spam update of 2026 rolled out in the back half of September, alongside a report that AI Overviews are citing more external links than before — but that those links increasingly route readers into AI Mode rather than out to the source website.
The UK's competition regulator kept pressing Google on search choice, ordering it to make switching the default search engine easier on Android and Chrome — a separate front from the AI Overviews opt-out fight covered in our Google AI Overviews statistics post.
Anthropic shipped Claude Fable 5.1 and Mythos 5.1 on September 1 — and by month's end, Australia's competition regulator had summoned both Anthropic's and OpenAI's CEOs to testify in an AI probe.
Perplexity retired its original Sonar chat-completions API on September 27, moving developers onto its newer Agent API — a quiet but real signal of where Perplexity is putting its engineering effort.

If June 2026 was the month AI search's infrastructure hardened — opt-out toggles, ad platforms, agent runtimes — September was the month that infrastructure started actively pulling in different directions at once. Google built a real publisher opt-out and extended it worldwide, then almost simultaneously tested a Discover feature that answers questions with AI summaries instead of sending clicks anywhere. OpenAI shipped its most capable model yet with genuine internal safety alarm attached, then immediately turned around and crossed $1 billion in ad revenue. Regulators on three continents pressed forward at once — the EU on content licensing, the UK on search-engine choice, Australia on model-maker accountability — while the underlying products kept shipping on their own schedule regardless.

Here's each story in turn.

OpenAI ships GPT-6 — twice — with real scrutiny attached

September's single biggest AI story wasn't search-specific, but it will shape every AI search product downstream of it. On September 3, OpenAI released GPT-6 Astra as a limited preview, calling it "the most intelligent and aligned model in the world" and — per chief global affairs officer framing that ran through OpenAI's own messaging — the start of "the AGI era." Astra sets new marks on computer use, browsing, software engineering, and cybersecurity, with a 1.05 million token context window. It shipped with unusual candour about risk: OpenAI's own system card disclosed that Astra's capabilities triggered internal security measures, and outside researchers flagged concern over a "recurrent depth" architecture technique reported to make the model harder to interpret and control. Reuters, the Guardian, Axios, CNBC, and Wired all covered the launch the same day; NBC News specifically covered the internal security trigger.

Three weeks later, on September 22, OpenAI expanded the family with GPT-6 Sol and Luna — smaller, cheaper models aimed at the lower end of the cost-intelligence curve, alongside a prompt-caching improvement for GPT-6 announced the same day. For anyone tracking visibility in ChatGPT, a full-family model refresh this large typically means citation patterns and answer structure shift too — worth re-running your prompt tracking in the weeks after a release like this rather than assuming last month's results still hold.

ChatGPT Ads crosses $1 billion — and keeps expanding

OpenAI's ad business hit a real milestone in early September: on August 31, OpenAI announced ChatGPT Ads had reached $1 billion in annualised revenue run rate in under 200 days since launch, now used by "tens of thousands of advertisers" with more than 50 technology and measurement partners. The same announcement opened self-serve access via Ads Manager to India, Europe, the Middle East, and North Africa, pushing ChatGPT Ads past 40 countries total. OpenAI cited early results — one ecommerce advertiser reporting 3x return on ad spend over 28 days, a partner reporting 80%+ of ad-driven ChatGPT traffic came from new customers.

The expansion didn't stop there: on September 23, OpenAI confirmed ChatGPT Ads had expanded to Southeast Asia and Taiwan. Combined with June's expansion into Japan and South Korea, OpenAI's ad platform has now gone from a US-only experiment to a genuinely global product in about seven months. As we noted after June's expansion: ranking organically in ChatGPT and buying placement in ChatGPT are becoming two different disciplines, and September made that split even clearer — CPC and outcome-optimized bidding now account for the majority of ad campaigns, with product feeds and custom audiences maturing well past pilot-stage features.

Google's opt-out goes global — and becomes leverage against the EU

The UK's Competition and Markets Authority forced Google to build a way for publishers to opt out of AI Overviews and AI Mode specifically back in June (see our AI Overviews statistics post for the full history). What happened next belongs in September's recap: per a footnote quietly added to Google's own announcement, Google rolled that same Search Console opt-out control out to every website worldwide on August 31, 2026 — more than three months ahead of the CMA's own December 2026 deadline for the feature. Once opted out, a site stops receiving any traffic or impressions from AI Overviews, AI Mode, or AI Overviews in Discover, and the toggle is explicitly excluded from affecting ordinary search ranking.

Then, on September 1, the story took a genuinely new turn: reporting confirmed Google is offering that same global opt-out specifically as a proposed remedy to resolve a separate EU antitrust probe — the European Commission's December 2025 investigation into whether Google uses publisher content unfairly to power AI Overviews. A fix built to satisfy one regulator in one country became, within three months, Google's opening move to satisfy a second regulator across an entire continent.

Google sends publishers two contradictory signals

The same month Google extended its opt-out worldwide, it also shipped a tool that gives publishers more control over their relationship with readers — and separately started testing a feature that could make that relationship less necessary at all.

On the "more control" side: Google launched a "Preferred Source" button that publishers can embed directly on their pages. One click marks that site as a reader's preferred source across Search, Discover, and Google News, and immediately returns the reader to where they left off on the publisher's page — Google ships the button in 16 languages with ready-made graphics and a full Search Central implementation guide.

On the other side: Search Engine Roundtable reported in late September that Google Discover is testing a "Dive Deeper" feature that, per its own description, "does not take you to a publisher's article" when clicked — instead surfacing an AI-generated topic overview with community reaction and links to some other articles. A secondary "Dive deeper" tap inside that overview does lead to publisher links, but the first click doesn't. As one widely shared write-up put it: if Google's own AI-generated summary gives a reader full context up front, there's less reason for them to click through to any publisher at all — the same zero-click dynamic that's driven every AI Overviews click-through study for two years, now showing up inside Discover specifically.

September's spam update lands, and AI Overview links increasingly route to AI Mode

Google shipped its fourth spam update of 2026 in the back half of September, its latest algorithmic pass against low-quality and manipulative content — the same ranking signals that feed both classic blue-link results and what AI Overviews draws from. Google Search Console also picked up a multimodal search type filter in its Performance report the same week, letting site owners separate multimodal query performance from standard search for the first time.

More consequential for anyone tracking AI Overviews visibility: a report circulating in late September found AI Overviews are now citing meaningfully more external links than earlier in the year — a positive signal on the surface. But a second, closely related finding complicates it: those AI Overview links are increasingly routing users into AI Mode rather than out to the cited website. More citations don't automatically mean more traffic if the click path terminates inside another Google surface instead of on your page — worth factoring into how you read your own AI Overviews citation counts going forward.

The UK keeps pressuring Google on search choice

Separate from the AI Overviews opt-out fight, the UK's CMA opened a new front in September: it's ordering Google to make it easier for Android and Chrome users to switch their default search engine away from Google. It's a reminder that the CMA's scrutiny of Google covers general search-engine competition, not just AI-answer features specifically — and that both fronts are moving on their own regulatory timelines at once.

Anthropic ships new flagship models, then faces regulators in Australia

Anthropic's September opened with a product release: Claude Fable 5.1 and Claude Mythos 5.1 launched September 1, becoming the current flagship models atop Claude's lineup (see our Claude statistics post for the full model history and revenue context). It closed with a regulatory story instead — by September 27, Australia's competition authority had called both Anthropic's and OpenAI's CEOs to appear as part of its ongoing AI industry probe, a sign that scrutiny of the major model providers is no longer confined to the US, UK, and EU.

Perplexity retires its original Sonar API

A smaller but real infrastructure shift: Perplexity confirmed it is retiring the original Sonar chat-completions API on September 27, 2026, moving developers onto its newer Agent API as the primary surface for that work going forward. It's not a consumer-facing change, but it's consistent with the direction Perplexity spent all of June signalling — investment in agent infrastructure (Computer, SPACE, Microsoft 365 integration) rather than the original search-API product line it launched with.

What this means if you're tracking AI search visibility

September's news doesn't point in one clean direction the way some months do — it's genuinely two stories running at once. A few practical implications:

If you publish content, the "Preferred Source" button is worth embedding — it's a low-effort way to capture returning readers directly. But don't let it distract from the more important finding: Google is actively testing features (Discover's "Dive Deeper") that reduce your traffic even while other Google teams build tools that could increase it. Read both signals, not just the friendlier one.
If you're tracking AI Overviews citations as a proxy for traffic, re-check whether the citation links you're earning route to your site directly or through AI Mode first — a rising citation count doesn't necessarily mean a rising click count anymore.
If you track ChatGPT visibility, expect some drift in citation patterns following the GPT-6 Astra and Sol/Luna releases — re-run your prompt tracking rather than assuming August's baseline still holds through a two-model-family refresh.
If you sell products through ChatGPT, the ad platform is no longer a US-only experiment — it's live in 40+ countries and crossed $1 billion in run rate in under seven months. Worth testing now if you haven't, before the CPC market matures further.