All issues
Discovery Digest · 18 September 2026
Issue 18. ChatGPT becomes a regulated search engine, the agent loop turns free, and voice goes dark on your dashboards
TL;DR
**The EU has formally classed ChatGPT as a very large online search engine, OpenAI has made the agent orchestration layer free, and voice answers just became cheap enough to deploy at scale with no link for you to measure.** Three of this week's releases push discovery into places your analytics cannot see: hosted agents that browse without a person watching, vertical assistants inside walled workspaces, and full-duplex phone conversations at five cents a minute. Meanwhile the ChatGPT Data agent hands warehouse queries to anyone who can type a sentence, which makes your semantic layer the most important marketing asset you own.
01 · Policy and infrastructure
The EU is funding a rebuild of the internet your marketing stack sits on
- What
- The European Commission's Next Generation Internet (NGI) initiative pays grass-root developers to build open-source components across every layer of the web, and the policy page setting out the plan was last generated on 16 September 2026. The stated mission is "to reimagine and re-engineer the Internet of tomorrow, reflecting fundamental human values", with the target described as "an interoperable platform ecosystem that embodies the values that Europe holds dear: openness, inclusivity, transparency, privacy, cooperation, and protection of data". Funding reaches developers through what the Commission calls a "cascade funding mechanism", and it covers open hardware, networking and transport, firmware and operating systems, electronic identities and middleware, decentralised ledgers, developer tooling, and over-the-top and vertical applications. The Commission is explicit that NGI projects play a significant role in enabling EU legislation, GDPR included.
- When
- Policy page last generated 16 September 2026, with funded work catalogued on an ongoing basis in the NGI innovations directory.
- How it shifts discovery
- This is not a product launch, so it will not show up in your reporting next quarter. It matters because the identity, consent and transport layers being funded here are the defaults your measurement, personalisation and ad stack quietly depend on. If electronic identity and privacy-by-default middleware ship as European norms, the assumptions behind cross-site tracking and logged-out attribution weaken further. What I'd do: assign one person to watch the electronic identity and middleware strands, and pressure-test your measurement plan against a world where user-controlled identity is the default rather than the exception.
- Questions to ask
- Which parts of our measurement stack break if user-controlled identity becomes the European default?
- Do we have anyone tracking EU technology funding as an input to our roadmap, or only EU enforcement?
- If open-source identity and consent middleware becomes standard, what is our first-party data plan?
- Sources
02 · AI interfaces
Google AI Edge Eloquent puts a rewrite layer between what you say and what an agent hears
- What
- Google has published Google AI Edge Eloquent, a voice dictation app for iOS and macOS that polishes spoken text and copies the result to your clipboard, sold as "Premium AI voice dictation without subscription". The page carries no launch date. The worked example Google publishes is the interesting part: filler words go, a self-correction is resolved so Tuesday becomes Wednesday, three proper nouns are rewritten (Surreal to Suril, Amy to Ami, Morgan to Mogan), and the clause "as he might still be traveling" is deleted. That clause was the reason for the change, not filler. The line that matters most sits in the FAQ, where Google says it is "evaluating other platforms including desktop for tasks like dictating docs, code and prompting to AI agents".
- When
- Published on the site on 16 September 2026. Google's product page lists no launch date.
- How it shifts discovery
- Voice as an input layer for agent instruction means a model-rewritten version of your words is what reaches the agent. That is a new place for meaning to go missing, and it is invisible in every analytics tool you own. Polishing models trade readability for reasoning, so the clause explaining why you asked for something is exactly the kind of thing that gets cut. What I'd do: if anyone on your team dictates briefs, prompts or client instructions, require a read-back before send, and keep the raw capture where the stakes are high.
- Questions to ask
- Who on our team dictates briefs or prompts, and does anyone check the polished output against intent?
- Where would a deleted reason clause cause real damage: client instructions, legal wording, or agent tasks?
- Do we log the raw input anywhere, or only the rewritten version?
- Sources
03 · AI search regulation
ChatGPT is now a Very Large Online Search Engine under EU law
- What
- On 31 August 2026 the European Commission designated ChatGPT as a Very Large Online Search Engine (VLOSE) under the Digital Services Act, and designated Reddit and Roblox as Very Large Online Platforms (VLOPs). All three declared at least 45 million average monthly users in the EU, which is the trigger threshold. The Commission describes ChatGPT as a hybrid service that responds to prompts including by searching the web, and on that basis classes it as an online search engine. Reddit and Roblox fall the other side of the line because they let users disseminate third-party content to the public. ChatGPT is co-supervised by Coimisiún na Meán in Ireland, while Reddit and Roblox sit with the Authority for Consumers and Markets in the Netherlands. That brings the total to 28 designated platforms and search engines.
- When
- Designated 31 August 2026, press release updated 1 September 2026. Compliance obligations bite four months from notification, so January 2027.
- How it shifts discovery
- A regulator has now said in writing that an assistant answering prompts by searching the web belongs in the same legal category as a search engine. That is the argument settled for procurement, board decks and anyone internally still treating AI assistants as a side channel. Systemic risk assessment and researcher data access also mean more disclosure about how answers are produced, which is the closest thing to an AI search transparency regime we have. What I'd do: stop justifying AI search work as experimental. Rewrite your channel definitions so assistants sit in the search line, and diarise January 2027 as the point where compliance documentation becomes a source of evidence you can read.
- Questions to ask
- Does our internal reporting still treat ChatGPT as social or other, rather than search?
- Which DSA transparency outputs due by January 2027 could we mine for insight into how answers are assembled?
- If EU regulators treat assistants as search engines, does our measurement and forecasting reflect that?
- Sources
04 · AI agents
Claude Cowork folds into chat, and Claude now decides whether your sentence is a job
- What
- Anthropic is merging Claude Cowork, its separate agentic workspace, into ordinary Claude chat for Pro and Max subscribers. Chats, projects, artifacts, connectors and skills carry over intact, and Claude Docs, Claude Slides and Claude Design shipped the same day alongside a Salesforce plugin and a set of small-business plugins. Three behaviours change. Claude now classifies the task rather than the user choosing between chat and Cowork, so the same typed sentence can produce a one-paragraph reply or a multi-step run across many sources. Tasks run asynchronously, so work continues after the laptop closes. And Claude interrupts with clarifying questions mid-task instead of failing silently on an ambiguous brief.
- When
- Announced and published on the site on 16 September 2026, with the merge rolling out to Pro and Max plans over the coming weeks.
- How it shifts discovery
- The asynchronous part breaks the assumption behind almost every AI visibility tool on the market: that a model interaction is a request and a response, bounded in time. If a task keeps reading sources after the session closes, your idea of when and how your content was consulted stops matching reality. Better clarifying behaviour also means people delegate bigger jobs, which means more of your content gets read by software mid-task rather than by a person mid-session. What I'd do: ask your visibility vendor how they handle asynchronous, multi-step runs, and treat any tool that only samples single prompts as a partial view.
- Questions to ask
- Does our AI visibility tooling assume one prompt equals one bounded response?
- If Claude classifies intent for the user, how do we test what a job-shaped prompt surfaces about our category versus a question-shaped one?
- Which of our content formats survive being read by an agent working through a multi-step task?
- Sources
05 · AI agents
OpenAI's Agents API makes the agent loop a free hosted product
- What
- OpenAI put the Agents API into public beta on 10 September 2026, a managed service for building and running cloud agents powered by the Codex harness behind its coding agent. You write the tools and the instructions, OpenAI runs the loop: durable sessions, streaming progress, connections to custom tools and MCP servers. Four capabilities carry the weight: long-running sessions that span far more than a single request, context compaction that summarises earlier work so long jobs do not run out of room, subagent coordination so one agent can hand work to others, and sandboxed execution either in an OpenAI-hosted sandbox or a developer-nominated environment. There is no separate orchestration fee. You pay only for the models and tools each session consumes.
- When
- Public beta from 10 September 2026.
- How it shifts discovery
- The boring parts, retry logic, crash-surviving state, context that does not overflow, were the barrier to entry, and the barrier is what kept the number of production agents small. Price that work at zero and hand it to anyone with an API key, and the population of agents pointed at the open web grows quickly. That lands in your logs as non-human traffic that reads, compares and acts without a person watching, and in your funnel as sessions that behave nothing like a visit. What I'd do: get raw server logs in front of your analytics owner this month, segment agent traffic separately from bots and humans, and check whether your forms, pricing pages and product data are legible to software that never sees your layout.
- Questions to ask
- Can we currently separate agent traffic from human traffic and from crawlers in our logs?
- Are our key conversion paths usable by software that does not render the page the way a person does?
- If agent volume doubles, which of our reported metrics become misleading first?
- Sources
06 · AI search visibility
ChatGPT for Financial Services shows the template for every data-heavy vertical
- What
- OpenAI launched ChatGPT for Financial Services on 10 September 2026, a version of ChatGPT Work bundling the GPT-6 Astra model with licensed data from Daloopa, PitchBook, LSEG News, Crunchbase and Quartr. Morgan Stanley and Evercore were design partners, and the named use cases are valuation analysis, leveraged buyout modelling, buyer screening, earnings analysis and pitchbook preparation. It ships with role-based access, encryption and audit-log exports, the minimum a regulated institution needs before anything reaches a desk. The moat here is made of contracts and process rather than parameters: licensed data, a workflow shaped around one job, and controls a risk function will approve.
- When
- Launched 10 September 2026.
- How it shifts discovery
- If a bank's analysts research inside that product, the question "which vendor should we use" gets asked there, against that licensed data, and answered from those sources. Your visibility problem moves from the open web into a walled workspace no tool measures. Tracking a single platform already gives you the wrong number, because answer engines quote wildly different sources, and vertical assistants add surfaces nobody covers at all. Expect the same pattern in legal, healthcare and industrial data next. What I'd do: identify which licensed data providers your buyers' assistants are likely to read, and work out whether your product information, pricing and category data exist inside those datasets rather than only on your site.
- Questions to ask
- Which licensed data providers in our vertical would an enterprise assistant bundle, and are we represented in them?
- How would we ever know if we lost a deal inside a closed workspace we cannot measure?
- Are we optimising only for the open web when our buyers increasingly research inside vendor-bundled data?
- Sources
07 · AI search and measurement
GPT-Live-1 makes voice cheap, and voice has no link to measure
- What
- OpenAI shipped GPT-Live-1 in the API on 10 September 2026 at $0.05 per minute, its first model that listens and speaks at the same time. Full duplex means it handles interruption mid-sentence, registers an acknowledgement like "mm-hm" without treating it as a new instruction, and sits through silence or background noise without narrating. It ships 12 voice options, supports custom voices and is designed to work over telephony, with named use cases in reservations, order updates and customer service. The architecture explains the price: GPT-Live-1 is the front-end voice layer and delegates hard reasoning to a separate backend model such as GPT-6 Astra or Codex, so you pay five cents a minute to hold the conversation and pay separately for the thinking.
- When
- Available in the API from 10 September 2026.
- How it shifts discovery
- Splitting the cheap conversational layer from the expensive reasoning layer is what makes voice agents economic, because most of a call is greeting, confirmation and closing. A ten-minute call costs fifty cents to hold, which turns voice from an innovation project into a finance decision. The measurement problem is that every AI visibility model assumes a screen: citations are links, links get clicked, visits get counted. A voice answer has no link, so if a caller asks an assistant to compare two suppliers and book one, your brand can win or lose with no trace in any analytics tool. What I'd do: start capturing the voice channel deliberately, through call reason coding, unique numbers, and post-conversion surveys that ask how the customer found you. Assume a growing share of demand arrives with no referrer at all.
- Questions to ask
- Do we ask new customers how they found us in a way that can capture a voice assistant answer?
- Is our product, pricing and availability data structured enough to be read aloud correctly by an agent?
- How much of our direct or unattributed traffic growth might already be assistant-driven?
- Sources
08 · Analytics governance
The ChatGPT Data agent removes the query, not the responsibility
- What
- OpenAI introduced the Data agent for ChatGPT Work on 9 September 2026, a plugin that connects to approved company data sources, investigates why a business metric moved, and builds interactive dashboards colleagues can share. It reaches warehouses and databases including Amazon Redshift, Google BigQuery, Snowflake, Databricks, ClickHouse, MongoDB and Datadog, pulls files from Google Drive and SharePoint, and works alongside existing dashboards in Power BI, Tableau, Sigma, Omni, Oracle BI and ThoughtSpot. Administrators install it and control access through Workspace settings, where individual data-source plugins can be enabled, configured and restricted.
- When
- Introduced 9 September 2026.
- How it shifts discovery
- Self-service analytics has failed for twenty years because the tool was never the bottleneck. The bottleneck is knowing which table is authoritative, which metric definition is current, and which joins quietly double-count. A natural language agent removes the syntax barrier completely and the semantics barrier not at all, so wrong numbers stop being slow and rare and become fast and plentiful, wrapped in a shareable dashboard. Attribution arguments get louder, and whoever brings the chart sets the frame. What I'd do: find out who holds the Workspace switch for data-source plugins, then write down your canonical tables, current metric definitions and known double-count traps before the agent is switched on. The agent will be exactly as trustworthy as the semantic layer underneath it.
- Questions to ask
- Who in our organisation controls which data-source plugins are enabled, and have they spoken to data governance?
- Is our metric dictionary written down and current, or does it live in three analysts' heads?
- Which joins in our marketing data would quietly double-count if someone asked a plain English question?
- Sources
09 · AI risk
Anthropic's threat report: agents now rewrite malware until detection stops
- What
- Anthropic published its most detailed AI threat intelligence report to date on 1 September 2026, covering operations detected and disrupted between December 2025 and August 2026 across seven harm areas and roughly 40 tracked groups. The harm areas are cyber operations, influence operations, surveillance, scams and fraud, biological misuse, conventional weapons development, and distillation (training a rival model on another model's outputs). One case describes a suspected Russian espionage actor whose AI agents watched security products to see when its malware was detected, then rebuilt that malware repeatedly until it stopped being caught. Anthropic also states that its newer models can no longer be assumed to sit safely below the threshold for meaningful bioweapons assistance.
- When
- Published 1 September 2026, covering December 2025 to August 2026.
- How it shifts discovery
- Signature-based detection assumes a new variant costs the attacker time and skill, and that cost is what gives a signature its shelf life. An agent that reworks code against your defences on its own removes the cost, so it removes the shelf life. Behaviour-based detection holds up better, because what an attacker is trying to do is far harder to vary than the code they use to do it. Read the disruption claim carefully too: a report of disrupted operations is, by construction, a report of the ones that were found. What I'd do: if you own web properties, forward this to whoever owns detection and ask one question, how much of our coverage depends on known file signatures. Then check your own agent-facing surfaces, forms, APIs and login flows for abuse assumptions written when attacks were manual.
- Questions to ask
- How much of our detection depends on known signatures rather than observed behaviour?
- Do our public forms, APIs and account flows assume a human attacker working manually?
- Who owns the relationship between marketing-owned web properties and the security team?
- Sources
10 · AI adoption data
One billion ChatGPT users, disclosed in an engineering post about storage
- What
- OpenAI confirmed it now serves over one billion ChatGPT users, and the number appeared in an engineering post about scaling online storage rather than a keynote. The post describes Habitat, an internal storage platform that started at DevDay 2023 as a Python library talking to a single database and now runs as a globally distributed system, with OpenAI's own summary putting the load at 22 million requests per second serving over one billion users. Coverage of the same post cites a larger current figure, close to 70 million requests per second across roughly 40 regions and more than 500 petabytes of data, so the gap is worth flagging rather than resolving. One detail deserves attention on its own: two engineers rewrote Habitat from Python to Rust in Q2 2026 using Codex and GPT-5.5, reportedly six times more CPU efficient and fifteen times more memory efficient, now carrying 95% of production traffic. An earlier OpenAI engineering post described 800 million users, so you can watch the number move across comparable documents.
- When
- Post published 25 August 2026.
- How it shifts discovery
- This is a better class of evidence than almost every AI adoption figure in circulation, because the user count appears as a constraint, the reason the team had to rebuild something, rather than as a boast. Nobody inflates the size of a problem they had to solve. But reach is not usage, preference or intent: a billion users tells you nothing about how often people ask questions in your category, or whether they act on the answer. What I'd do: use this to stop the internal argument about whether AI assistants are big enough to matter, then refuse to size the channel from it. Size it from your own data instead: unattributed traffic, brand search shifts, and how customers say they found you. Also steal the reading test, a number in a document written for another purpose beats the same number in a document written to persuade you.
- Questions to ask
- Are we sizing the AI channel from vendor reach figures rather than our own demand data?
- What is our best first-party estimate of assistant-influenced demand, and how confident are we in it?
- What does two engineers rewriting a storage layer in a quarter imply for our own engineering and content throughput assumptions?
- Sources
Key takeaways
What to walk away with this week
- 01
The EU now treats ChatGPT as a search engine in law, with compliance due by January 2027. Move assistants into your search reporting line and stop labelling the work experimental.
- 02
With the OpenAI Agents API charging no orchestration fee, expect far more software reading your site without a person watching. Segment agent traffic in your logs now.
- 03
Voice at $0.05 a minute creates a discovery surface with no link, no referrer and no click. Capture it with call reason coding and how-did-you-find-us questions instead.
- 04
Claude's merge makes agent work asynchronous, which breaks the request-and-response assumption behind most AI visibility tooling. Ask your vendor how they handle multi-step runs.
- 05
The ChatGPT Data agent removes the SQL barrier but not the semantics barrier. Document canonical tables and metric definitions before it is enabled in your workspace.
- 06
Vertical assistants bundled with licensed data move buyer research into workspaces nobody can measure. Check whether your product data exists in the datasets those assistants read.
