Weekly AI Marketing & Web Update

Workspace Gets More Agentic, Measurement Gets Smarter & AI Creative Gets Easier to Control

← All posts

· 12 min read · by Laura Ferruggia

This week is less about another giant model launch and more about AI getting embedded deeper into the systems marketing and web teams already use.

Google is turning Workspace into a more genuinely cross-app agentic environment while also making first-party measurement infrastructure more important—and more measurable. OpenAI’s new image model improves one of the biggest weaknesses in AI creative production: reliably iterating on an existing asset without accidentally rebuilding half the damn thing. Claude Code is starting to make AI development workflows testable instead of purely vibe-based, and there’s also an ACF security release worth handling now rather than eventually.

The larger pattern is pretty clear: we’re moving from “our team uses AI” toward “our workflows are designed around AI, measured, standardized and tested.”

1. Gemini is getting substantially more useful inside Google Workspace

On September 9, Google announced new agentic capabilities across Gmail, Drive, Docs, Slides and Chat.

The important part isn’t that Gemini can write a document or summarize an email. We’ve had enough “AI can draft an email” demos to last several geological eras.

The new piece is cross-app orchestration.

Gemini can gather context from relevant emails, files and conversations and then create structured deliverables in other Workspace applications. Google’s examples include turning a Google Chat discussion into a presentation, creating a structured spreadsheet from Drive content and converting source material into Docs or branded Slides while the user stays in the application where the work started.

Why it matters

For teams already working heavily in Google Workspace, this can remove the annoying context-assembly step that happens before the actual work begins.

A workflow that currently looks like:

find the email → find the Drive folder → open three docs → collect the relevant context → start the brief

can increasingly become:

select the relevant Workspace context → ask Gemini to assemble the first deliverable

That’s a much more meaningful productivity improvement than another writing assistant.

It also reinforces the idea that we probably shouldn’t expect one AI platform to own every workflow. Gemini has an obvious structural advantage when the work already lives in Google Workspace. Claude may be stronger for other reasoning, development or creative tasks, while ChatGPT has its own strengths in research, computer use and multimodal work.

The goal is increasingly to figure out which tool belongs at which point in the workflow.

What to try

Pick one repetitive workflow that starts with information scattered across Gmail and Drive and ends with a deliverable.

A straightforward test would be:

client/project emails + Drive files → project brief or status presentation

Then compare the AI-assisted workflow against the existing one. Did it actually eliminate steps and context switching, or did it just produce another draft that needed babysitting?

Agency implication

There is eventually a real consulting opportunity around Workspace-native AI workflow design: helping clients identify repetitive information flows and redesign them around the AI capabilities already included in tools their teams use every day.

Source: Google Workspace — Less switching, more flow: 5 new agentic capabilities

2. Google is making first-party measurement more important—and easier to prove

On September 10, Google announced several additions to its measurement stack.

Data Manager is being integrated directly into Google Analytics and Display & Video 360, Enhanced Conversions are expanding into GA and DV360, and Google says its Data Manager API is now universal and based on the IAB Tech Lab Event and Conversions API standard.

One of the more interesting additions is the new Data Strength Uplift Metric in Google Ads.

The metric is designed to estimate how many additional conversions a stronger first-party data implementation has recovered. Google is effectively trying to make the business impact of better tagging and first-party measurement infrastructure more visible inside the advertising platform itself.

Google also updated Meridian, its open-source marketing mix modeling platform. New agentic functionality can help audit data quality, resolve issues and guide model construction. Meridian can now incorporate branded search/query volume as an input, and Google’s GeoX causal geo-experiment library is now generally available worldwide.

Why it matters

As advertising platforms automate more bidding, targeting, creative selection and campaign execution, agencies need to ask where human expertise becomes more valuable rather than less.

Measurement infrastructure is one of those areas.

The platform can optimize only against the signals it receives.

That makes these capabilities more strategically important as campaign execution becomes increasingly automated:

What to try

Audit major paid-media accounts and ask:

Agency implication

There’s a stronger service here than simply “GA4 setup.”

A modern measurement offering could look more like:

measurement audit → first-party data architecture → Enhanced Conversions → CRM/offline integration → validation → incrementality testing

That becomes harder to commoditize than routine campaign configuration.

One caveat: Google’s conversion-uplift percentages are Google’s own vendor data. The new infrastructure is interesting; the headline performance claims shouldn’t automatically become universal expectations.

Source: Google — Drive profitable growth with new data and measurement tools

3. ChatGPT Images 2.5 improves the part of AI image generation that actually matters: iteration

OpenAI released ChatGPT Images 2.5 on September 8.

The important improvement isn’t that ChatGPT can edit images. It already could.

The bigger change is that the underlying model is better at preserving an existing subject, composition and visual direction while making targeted changes across multiple rounds of editing. OpenAI says reference-image fidelity has improved, multi-turn edits are more consistent and generation latency is up to 50% lower than Images 2.0.

That addresses one of the most irritating problems in AI creative workflows:

“Please change this one thing.”

AI:

“Excellent. I changed that, redesigned the room, replaced the product, moved the person’s face three inches to the left and summoned a new lamp.”

Images 2.5 is designed to make that iteration loop more stable.

OpenAI also introduced Sketch, which lets users draw a rough visual reference directly in ChatGPT, along with templates for common formats and prompt sharing. The model is also better at complex layouts, transparent backgrounds and preserving requested visual styles.

Why it matters

AI-generated imagery becomes much more useful for real marketing work when we can iterate toward an approved result instead of repeatedly starting over.

Potential uses include:

Sketch is particularly interesting because some visual ideas are dramatically easier to point at than describe.

Instead of writing a tortured paragraph explaining where everything belongs, we can effectively say:

“No. The product goes here.”

…and draw it.

What to try

Take one real marketing asset through several rounds of revision.

Measure whether Images 2.5 can preserve the parts we’ve already approved while changing only what we ask it to change.

Separately, try:

rough Sketch → generated concept → refinement → Figma

The useful question isn’t whether the first image looks cool.

It’s whether the tool can reliably help move an idea toward a usable production asset.

Agency implication

If image iteration becomes faster and more reliable, it becomes easier to include custom visual assets, article imagery, campaign variations and lightweight creative experimentation inside content and marketing packages that previously depended more heavily on stock imagery or separate design requests.

That doesn’t replace designers.

It gives designers and marketers a much faster pile of raw material to work with.

Source: OpenAI — Introducing ChatGPT Images 2.5

4. Claude Code can now evaluate AI development workflows instead of relying entirely on vibes

Claude Code 2.1.269, released September 11, added claude plugin eval.

The feature lets developers run a plugin’s evaluation suite against Claude Code and receive scored, reproducible results, including JSON and HTML reports.

That sounds nerdier than it actually is.

Most AI development workflows still rely heavily on:

“We’ve used Claude for this a bunch of times and it normally works.”

That’s useful, but it isn’t exactly rigorous QA.

Evaluations create a path toward defining expected behavior and repeatedly checking whether an AI-assisted workflow still produces acceptable results.

Why it matters for web teams

Imagine reusable Claude Code workflows for:

If those become shared plugins, skills or standardized workflows, we can eventually test their behavior rather than judging everything subjectively.

That’s an important distinction.

Having Claude Code isn’t much of a competitive advantage when everyone can buy Claude Code.

Having reliable, tested internal workflows built around Claude Code can be.

What to try

Pick one development workflow we already repeat and ask:

What would “correct” output actually look like?

Then investigate whether those expectations could become a simple evaluation suite.

We don’t need an elaborate internal AI testing laboratory next Tuesday.

One real workflow is enough to understand whether the approach has value.

Agency implication

Over time, reusable and evaluated AI development workflows can become internal agency IP.

The model is commoditized.

The process doesn’t have to be.

Source: Anthropic — Claude Code releases

5. If you manage WordPress sites using ACF, update it

Advanced Custom Fields released ACF and ACF Pro 6.8.10 on September 10 with several security fixes, and the ACF team recommends upgrading as soon as possible.

The fixes touch multiple areas, including:

For teams running custom WordPress builds—particularly ACF-heavy ones—this is worth handling now rather than waiting for the next leisurely plugin cleanup.

What to do

Confirm managed sites are on ACF 6.8.10 or newer, using the normal backup, staging and validation process.

There is no groundbreaking AI strategy hiding in this section.

Sometimes the correct digital-transformation initiative is:

install the damn security patch.

Source: Advanced Custom Fields — ACF 6.8.10 Security Release

6. Anthropic’s latest threat report is a useful reminder to know where AI prompts actually go

Anthropic’s September threat-intelligence report raises an issue that matters as teams start using more model aggregators, routing services and third-party AI interfaces.

Anthropic says it found third-party routing and proxy services relaying user conversations to frontier models, in some cases retaining or reselling those conversations without users understanding the full data path. The material Anthropic observed included company information, internal work and, in some cases, active credentials.

This is Anthropic’s own threat research, so it should be read as such. But the underlying operational lesson is sound regardless of provider.

Why it matters

The interface someone is using may not necessarily be the company actually processing the request.

Model routers can be convenient, particularly when they provide access to many models through one subscription.

They also introduce another organization into the data chain.

That matters when prompts contain:

What to do

Know which AI tools are approved for sensitive work and understand who actually processes the data.

Direct business relationships with established AI vendors are generally easier to assess than an unknown intermediary promising every frontier model on Earth for $11.99/month and a dream.

And regardless of the platform: don’t paste live credentials into AI prompts.

Agency implication

AI governance is becoming part of operational competence.

Clients increasingly need rules around approved providers, sensitive-data handling, permissions and agent access—not just advice about which model writes the nicest paragraph.

Source: Anthropic — September 2026 Threat Intelligence Report

What we should actually do this week

  1. Test Gemini on a real cross-Workspace workflow. Give it a project with relevant Gmail, Drive and Chat context and see whether it can produce a useful brief, document or presentation while genuinely removing manual steps.
  2. Audit first-party measurement on important paid-media accounts. Look at Enhanced Conversions, CRM events, offline conversions and other useful first-party signals that aren’t currently informing optimization.
  3. Test ChatGPT Images 2.5 on a real marketing asset. Don’t judge it from one pretty generation. Push an asset through multiple rounds of editing and see whether it can preserve approved elements while reliably handling revisions.
  4. Experiment with one testable Claude Code workflow. Pick something repetitive and define what successful output should look like. Explore whether claude plugin eval could turn that into a repeatable quality check.
  5. Update ACF to 6.8.10+ across managed sites.
  6. Review where sensitive AI work is being routed. Make sure proprietary or client material stays inside approved platforms, particularly when third-party model routers are involved.

The bigger pattern

A lot of AI adoption has focused on what individual people can generate faster.

This week’s developments point toward the next stage.

Gemini is starting to orchestrate work across the productivity systems where information already lives. Google’s advertising ecosystem is putting more weight on the quality of the data feeding its automation. ChatGPT Images is getting better at preserving and iterating on creative work instead of merely generating first drafts. Claude Code is gaining ways to test repeatable agent behavior.

The competitive question becomes less:

“Does our team use AI?”

and more:

“Which workflows should AI participate in, how do we know they’re working, and how do we make those workflows repeatable?”

That’s a much more interesting problem—and a much more defensible capability for a marketing or web team to build.

Sources / further reading

AI Marketing Marketing Technology WordPress AI Workflows