Weekly AI Marketing & Web Update

Claude Code Leaves the Laptop, Search Gets More Visual & Conversational Ads Keep Expanding

← All posts

· 13 min read · by Laura Ferruggia

This week is less about another shiny model launch and more about AI changing where work happens, how people discover businesses, and what marketers can actually measure.

Claude Code can now keep working in the cloud after a developer closes the laptop. Google Search Console can finally isolate visual-search discovery from Lens, Circle to Search and image uploads. Google Ads is preparing to show more of what AI Max actually did, while YouTube is moving conversational AI directly beside ads. ChatGPT Voice is also becoming a real execution interface instead of glorified dictation.

The useful pattern: AI is moving away from the prompt box and deeper into the workflow, the search experience and the customer journey.

I’m keeping the genuinely important dev/security items, but they’re no longer allowed to eat the entire newsletter like a hungry raccoon in a server room.

1. Claude Code cloud sessions are now generally available — and worth testing immediately

On September 23, Anthropic moved Claude Code cloud sessions out of research preview. A cloud session runs Claude Code on Anthropic-managed infrastructure instead of the developer’s machine, so long-running work can continue after the laptop closes. Developers can start a task and later check or steer it from the browser, desktop app, mobile app or terminal.

Cloud sessions are available on Pro, Max and Team plans, plus eligible Enterprise seats. They currently work with GitHub repositories and run inside isolated virtual machines. They still share the account’s broader Claude usage limits, so this is not an infinite-agent loophole invented by benevolent gods.

Anthropic is also offering temporary promotional cloud-session credits to eligible existing subscribers, with claim and expiration dates in early October and November respectively.

Why it matters

For teams already using Claude Code heavily, this changes the natural unit of work. Local Claude Code is ideal for active collaboration. Cloud sessions are better suited to work that benefits from asynchronous delegation:

For WordPress work, obvious experiments include a full custom-theme audit, scanning for deprecated patterns, accessibility remediation across templates, or refactoring repeated ACF/block implementations while nobody babysits the terminal.

What to try

Take a real GitHub-hosted web project and give a cloud session something large enough that we would normally hesitate to hand it because it would monopolize a local session. Give it a tightly defined task, acceptance criteria, tests or validation instructions, and permission to work independently. Then close the damn laptop.

The metric is not “did it write code?” It is “how much useful, mergeable work came back without active supervision?”

Agency implication

Asynchronous agentic development can become a real margin advantage if developers can define, delegate and review multiple parallel jobs instead of spending all of their time producing each change manually.

2. Claude Opus 5.5 makes high-end agentic work cheaper and faster

Anthropic released Claude Opus 5.5 on September 22. This model launch earns a place here because the interesting part is economics, not another benchmark victory lap.

Anthropic says Opus 5.5 costs about 40% less than Opus 5 on typical workloads while delivering output more than 30% faster. Published API pricing drops to $4 per million input tokens and $20 per million output tokens, with cached reads dropping especially sharply. Anthropic also says the model generally uses fewer tokens and fewer tool calls to complete comparable work.

A higher-priced Fast mode is available for jobs where latency matters more than token cost.

Why it matters

Agent economics are not just the per-token price. The real equation is closer to: model price × tokens × tool calls × retries × human cleanup. If a stronger model costs less and completes complicated jobs in fewer steps, previously extravagant long-running workflows become more reasonable.

That matters beyond developers. Any marketing operation involving large research jobs, content transformations, analysis or multi-step agent workflows benefits if the cost per completed task falls.

What to try

Benchmark Opus 5.5 on one ugly real-world task rather than a toy example. Track elapsed time, interventions, token usage, failed approaches and human cleanup. The useful metric is cost per completed task, not whether one model makes a slightly prettier first answer.

Agency implication

As agent costs fall, agencies should be careful not to price AI-assisted work as if efficiency simply means fewer billable hours. The value of the completed outcome does not collapse just because the production method improved.

3. ChatGPT Voice can now use plugins and run Work tasks

On September 23, OpenAI expanded ChatGPT Voice so it can use plugins and connected apps on web, iOS and Android. Voice is also now available inside ChatGPT Work on web and mobile.

That means a user can verbally ask Work to create documents, presentations or spreadsheets, use connected apps, or operate through its browser. If the voice conversation ends while the task is still running, Work can continue in text. Existing permissions and approval requirements still apply.

Why it matters

Voice has mostly lived in the “pleasant interface” category. This starts moving it into the input layer for actual work. Complex agentic tasks do not necessarily need to begin at a desk with a carefully written prompt.

The potential workflow becomes: speak intent → agent gathers context → agent executes → human reviews the artifact.

What to try

Use Voice + Work for something you would normally wait until you were back at a computer to start: prepare a meeting brief from connected sources, turn a conversation into a document, research something and start a presentation, or gather information from email/files and produce a deliverable.

Do not test it as dictation. Test whether speaking is now a viable way to initiate and steer multi-step work.

Agency implication

This matters for adoption because “prompt literacy” matters less when people can simply explain outcomes, constraints and approval boundaries naturally. That is a more durable skill than memorizing prompt formulas.

4. Search Console can now show how people find you through Lens, Circle to Search and other visual searches

On September 24, Google added a new multimodal-search filter to Search Console. It reports when users discover a site through visual-search experiences including Google Lens, Circle to Search on Android, image uploads to Google Search and Chrome’s “Search this image” experience.

Google added multimodal data to both the standard Search performance report and the Generative AI performance report, and the feature is rolling out globally.

Why it matters

This gives marketers first-party visibility into a search behavior that has been annoyingly difficult to measure. Search is increasingly not “type words → receive links.” People can point a camera at something, circle part of an image, upload a screenshot or ask Google about something visual.

That makes image strategy more important than “did we remember the alt text?” Useful original diagrams, product imagery, screenshots, before/after examples, charts and photography can become discovery assets in their own right.

Because multimodal impressions can also appear in the Generative AI report, we can start seeing where visual discovery and AI search overlap.

What to try

Open Search Console Performance and look for the new Multimodal search type. For sites receiving meaningful impressions, identify which pages are being discovered visually, what kinds of images those pages contain, whether particular formats repeat, and whether those pages also show generative-AI visibility.

Agency implication

Visual-search optimization now belongs inside modern SEO/AEO audits because it is becoming measurable. That does not mean we need to invent a $1,500 “Multimodal GEO Package™.” Civilization has suffered enough.

5. Google Ads is preparing to show marketers more of what AI Max actually did

On September 23, Google announced a new AI Max reporting view designed to show the complete Search-ad journey in one place: search term → creative assets shown → landing page selected.

The feature is intended to help advertisers understand and validate how AI Max assembles campaign experiences. Google says availability is coming later this year, so this is announced rather than broadly deployable today.

Google is also expanding the closed beta of AI Brief, which lets marketers describe their business, audience and messaging to AI Max in natural language, into additional languages.

Why it matters

One of the legitimate frustrations with increasingly automated advertising is: cool, Google got the conversion — what the hell did Google actually do?

When AI controls more query matching, creative selection and landing-page choice, reporting needs to expose those decisions. A unified journey view should make it easier to spot questionable query matching, weird asset combinations, inappropriate landing pages and combinations worth reinforcing.

What to try

Nothing to deploy yet. But teams running AI Max should plan to make this view part of campaign QA as soon as it appears rather than treating AI Max as a black-box performance number.

Agency implication

As campaign mechanics become more automated, strategic oversight and QA of the automation become more valuable. The work is not disappearing; it is moving up a layer.

6. Google is putting conversational AI directly beside YouTube ads

Google’s September Demand Gen Drop, announced September 24, adds Business Agent for YouTube Ads. For eligible video ads using product feeds, viewers can interact with conversational AI alongside the ad and ask questions about the brand or products without leaving YouTube.

Google is currently asking interested advertisers to sign up, so this is an emerging rollout rather than something universally available today.

The same update includes one-click image-ad experiences on YouTube Shorts, simpler one-click landing-page access from Gmail ads, and Affiliate Location Extensions in Google Maps. Google also shared its own internal performance data for Gmail creative; treat the exact uplift as vendor-reported rather than universal truth.

Why it matters

This fits directly with the conversational-ad trend we saw in ChatGPT Sponsored Agents. The funnel is beginning to look more like: ad → AI conversation → answer questions/objections → action, instead of always forcing the user into a landing page before the platform can help them understand the offer.

Google and OpenAI moving in the same direction independently is a much bigger signal than either individual beta.

What to try

For ecommerce or product-heavy clients, start identifying the questions users repeatedly ask before purchase: compatibility, sizing, specifications, shipping, use cases, product differences and availability. Then ask whether the product feed and website actually contain reliable structured answers.

Agency implication

Conversational ad optimization is starting to look like a real cross-platform discipline. Marketing teams may increasingly own not just the ad, but also the knowledge and structured product information that power the conversation after engagement.

Also worth knowing

These did not earn full sections in the rebalanced version, but they are still useful operationally.

What we should actually do this week

  1. Run a real Claude Code cloud session. Give it a substantial GitHub-hosted task with clear acceptance criteria and see how much mergeable work comes back without babysitting.
  2. Check Search Console multimodal reporting. Capture a baseline for sites receiving Lens/Circle-to-Search/image-upload visibility and inspect what kinds of visuals are actually earning discovery.
  3. Prepare for conversational ads. For a product-heavy client, map the pre-purchase questions an AI agent would need to answer and audit whether the site/feed contains reliable answers.
  4. Benchmark Opus 5.5 on a real task. Measure time, interventions, usage and cleanup rather than benchmark vibes.
  5. Try Voice → Work on an actual multi-step workflow. The test is whether spoken intent can reliably initiate useful work with connected apps, not whether the model can chat pleasantly.
  6. Handle the boring-but-important hygiene. Verify WordPress security updates, current coding-agent versions/plugins and GA4 hostname filtering where relevant.

The bigger pattern

The strongest signal this week is that AI is moving away from the prompt box in several directions at once.

Claude Code can work asynchronously without staying attached to a developer’s computer. ChatGPT Work can be initiated by voice. Google can now measure visual discovery in Search Console. Paid platforms are moving conversational agents directly into ad experiences while trying to make their own automation less opaque.

That changes what good AI adoption looks like for marketing and web teams.

The scarce skill is becoming less “knowing how to use AI” and more knowing what work should be delegated, what data and context the system needs, how the output should be measured, and where humans still need to inspect the result.

That is also where agency differentiation gets more interesting. Model access commoditizes quickly. A team that knows how to connect AI to reliable data, useful content, clean measurement, safe development workflows and actual business decisions has something much harder to copy.

Which is considerably less sexy than “10 AMAZING PROMPTS THAT WILL 10X YOUR MARKETING,” but, inconveniently, much more useful.

Sources

AI Marketing AI Development Visual Search Conversational Ads