To write this week’s Member Brief for you, dear futurists, I built an Android app.

That’s not something that I thought I would write when the week began, nor is it something I thought I would ever write, to be honest.

Specifically, I built an app that converts any text on my phone into the NATO phonetic alphabet because I am tired of yelling "T as in Tom, no, T, Tango" at airline agents who can't hear me over their headsets after waiting on hold for two hours.

It’s a simple, personal piece of software I need that never previously existed. And until very recently, it’s the kind of thing that would have required me to learn Java, configure Android Studio, wrestle with the “Gradle build system” (whatever the hotel-echo-lima-lima that is), and invest more than a weekend of my free time.

Instead, I described it to Gemini using the newly released Antigravity desktop coding agent, which was released at Google I/O 2026 last week. About 20 minutes later, I had a working application, and I didn't write a single line of code.

The future, as advertised at Google I/O 2026, is here. It is real. And it works.

But it also cost me the equivalent of twelve dollars in API output tokens.

Not for the finished app, mind you. The twelve dollars were for the unsolicited part of the project. 

Gemini, in its enthusiasm, did not just build me an Android app. It built me an entire fake phone simulator in Chrome to run the app in, complete with a home screen, fake apps, and pretend autofill suggestions, because I did not have an Android emulator installed on my laptop.

I did not ask for any of that.

It used roughly four million output tokens to construct a piece of UI scaffolding that I will throw away the moment I deploy the actual Android Package Kit. In mere minutes, I spent the equivalent of a birthday drink at Starbucks on a whim that has no actual utility.

What follows is the breakdown that most recap videos won't give you (or the subsequent crash-outs from AI influencers).

Google left the most important parts off the slides: the brutal math of running these models at scale for the frontier labs, the rigid discipline for the businesses that will actually take full advantage of them, and the fact that most of your everyday merchants are currently lightyears away from the technical and data standards required to take advantage of this optimistic future.

For Future Commerce Plus members planning their 2026 and 2027 commerce roadmaps, the most important thing to understand about Google I/O is this: the future is here, it is wonderful, but the cost of admission will be steep.

What Google Actually Announced

The full list of announcements, in plain commerce-leader terms, was:

• Gemini 3.5 Flash and Gemini Omni. Flash is the new frontier multimodal model. Omni is the "world model" for video and creative generation. Both are proprietary and gated behind paid tiers.
• Gemini Spark. A 24/7 personal agent running on Gemini 3.5 and the Antigravity harness, Gemini Spark continues working when your laptop is closed, and integrates with Workspace, Canva, OpenTable, and Instacart. US Google AI Ultra subscribers get first dibs.
• Universal Commerce Protocol expansion. Use cases have expanded to include hotel booking and local food delivery, and major players such as Amazon, Meta, Microsoft, Salesforce, and Stripe have joined the technical council. UCP will roll out to Canada and Australia in the coming months, and the UK will come after.
• Universal Cart. An agentic hub for products added from Search, Gemini, YouTube, and Gmail, Universal Cart offers background price tracking, compatibility flagging, in-stock alerts, and Wallet integration. US consumers will get to experience the tech on Search and Gemini this summer.
• Agent Payments Protocol (AP2). An open, shared protocol that provides tamper-proof digital mandates for agent transactions within consumer-set constraints. AP2 will roll out gradually to Google products over the next year, starting with Gemini Spark this fall.
• AI Performance Insights (Merchant Center). This share-of-voice metric will show how your products perform on AI surfaces compared to competitor brands, and will even be broken down by funnel stage.
• Conversational attributes for product feeds. Six new optional attributes will determine your legibility for LLMs: Q&A, additional variants, popularity rank, related products, document links, and item groups.

We spoke with three Google executives live on-site at Google I/O to get more details on these announcements. You can listen to them on the podcast.

Now, the real substance for those in Commerce.

In a Think Retail spotlight session at GML, Ashish Gupta, VP and GM of Merchant Solutions, laid out the four foundations of agentic commerce. We are going to call this the Commerce 4 because they encompass the most useful and actionable insights from the entire week for commerce executives trying to figure out where to spend (and what to do) in 2026.

Foundation 1: Supercharged product feeds. 

Without rich, accurate, and conversational product data, Google's AI cannot match shoppers to your products. Augustina Sartori, Ulta's Head of Agentic Commerce, described this as moving to "identity-first product feeds" that contain materially more information than legacy feeds.

Foundation 2: First-party data activation. 

Loyalty data, purchase history, and preference data must be directly connected to Google's AI systems for personalization to work. Ulta noted that 47 million loyalty members drive 95% of the retailer’s sales, and the agentic strategy depends entirely on extending that loyalty context across both on-site and off-site experiences.

Foundation 3: AI-powered ad campaigns. 

Vidhya Srinivasan, VP/GM of Advertising and Commerce at Google, was about as direct as you can possibly be about this on the Google Marketing Live stage: "The only way you can win in this age of AI is with AI."

Google is steering merchants toward AI Max, PMax, and Demand Gen campaigns optimized through Gemini. The conversational, multi-sentence query has broken the keyword-based campaign architecture, and Google is offering AI-driven targeting as the replacement.

Foundation 4: UCP integration. 

Direct integration with the Universal Commerce Protocol (UCP) via Merchant Center, including the API and payments work required to support cart transfers, multi-item checkout, and seller-of-record continuity.

Only one of these four foundations generates revenue for Google directly through ads. The other three are infrastructure investments that primarily benefit the merchant by making them more findable within Google's AI surfaces. Whether that findability translates into incremental revenue depends entirely on consumer adoption of the new surfaces. Here, at least, the third-party data is starting to confirm what Google is claiming on stage.

‍

‍

“Pay No Attention to the Model Pricing Behind the Curtain”

Gemini 3.5 powers nearly every consumer-facing capability launched at I/O: Gemini Spark, the Universal Cart, AI Mode upgrades in Search, Gemini Omni's anything-to-anything model, and Antigravity 2.0 for developers.

Google CEO Sundar Pichai spent a meaningful portion of his keynote on token volume:

• Two years ago: 9.7 trillion tokens per month across Gemini surfaces
• Last year at I/O: 480 trillion tokens per month
• Today: 3.2 quadrillion tokens per month
• API throughput: 19 billion tokens per minute, with 375 customers each processing over a trillion
• Internal usage: from 500 billion tokens per day in March to over 3 trillion per day by mid-May

Google presented this as a triumph. On the engineering side, it is.

Here is what was not on the slide: Gemini 3.5 Flash is priced at $1.50 per million input tokens and $9 per million output tokens. That output cost is roughly three times higher than the previous-generation Flash model, which Google is sunsetting. A faster, more capable model that costs three times more, while quietly retiring the cheaper option.

According to Ramp's enterprise spending data cited by research firm Artefact, the average cost per million tokens across major providers fell from roughly $10 to $2.50 in a single year. Epoch AI's research suggests inference costs are falling dramatically year over year when accounting for both pricing and efficiency improvements.

At $1.50/$9 for API usage, Gemini 3.5 Flash is actually competitive with Anthropic's Claude Opus 4.7 at $5/$25 and sits in the middle of the pack compared with most US enterprise offerings.

“The real cost story is consumption, not unit price,” says veteran retail reporter Nicole Silberstein. “Enterprises that built 2024 budgets around legacy token rates are finding that agentic workflows at 2026 adoption levels consume multiples of what those spreadsheets projected.”

Meanwhile, Chinese providers are explicitly pricing to win this exact arbitrage.

DeepSeek cut prices on V4-Pro this spring in a move analysts read as a direct shot at OpenAI and Anthropic, and Qwen and Kimi are open-source alternatives. Grok sits between the US and Chinese tiers (and developers might actually use Grok now that they’re acquiring Cursor).

Bottom line: the real future of our AI consumption will be in workflow design, not Google's (or anyone else’s) price tag.

Enterprises that build agentic systems with disciplined token budgets and a mix of model tiers will most likely run profitably on any major US provider. Enterprises that let agents run unconstrained will hit budget walls regardless of which logo is on the invoice, and those are the ones who will quietly start routing workloads through DeepSeek or Qwen in 2026 and 2027.

If I were Google, I wouldn't address the bear case from the stage either. If your unit cost drops by 50% and your unit consumption increases by 500%, your total cost increases by 250%.

Artificial Analysis and Simon Willison have both observed that for multi-step autonomous tasks, Gemini 3.5 Flash can actually be more expensive end-to-end than older Pro-tier models due to "token bloat" from heavy reasoning.

It seems like the market is already pricing this in.

Uber's COO said this week it is getting harder to justify AI spend internally, despite continued belief in the technology. Developers brag on social media about generating 189 million tokens through DeepSeek for $2.31. Theo Browne went viral, pointing out (read: crashing out) that Gemini 3.5 Flash scores roughly half of what GPT-5.5 scores on independent intelligence benchmarks while costing meaningfully more to run a comparable workload.

Sundar told the audience that companies are "blowing through their annual token budgets," and that if top Google Cloud customers shifted 80% of their workloads from competing frontier models to 3.5 Flash, they would “save over a billion dollars annually.”

In plain English: Flash is the budget option that prevents you from going bankrupt running Claude Opus. If you’re somewhat shy about tripling the costs of your frontier model API costs, this is then a defensible comparison to have to argue from the main stage of your developer conference.

Since we began writing this story, a new Lite model has been introduced, which may lessen the burden here.

The Search Behavior Shift

One of the most repeated stats at Google Marketing Live was that AI Mode queries are three times longer than traditional search queries.

Ashish Gupta, VP & GM of Merchant Shopping at Google, added that “one in six” AI Mode queries now include audio or an image. Liz Reid, who leads Search at Google, said AI Mode queries have more than doubled every quarter, and AI Mode now has over a billion monthly users.

The refrain across both events is that consumer search behavior has fundamentally changed in about 12 months.

We do not take Google's stage numbers at face value, especially when they conveniently support the products they sell. But on this specific claim, the independent data is unusually corroborative.

RealityMine's April 2026 panel data, drawn from consumers in the US and UK, show average monthly Google searches per person up 26% in the UK and 37% in the US compared to May 2025. Average search length is up 8% year over year in both markets. The share of searches over 30 characters has grown 24% in the UK and 17% in the US, with searches over 40 characters growing at a similar clip.

SOCi's research finds that LLM queries are nearly six times longer than traditional search queries on average, and full sentences are replacing the traditional “four-word” search (e.g., coffee shop near me).

Trakkr.ai measured the gap directly across competing platforms and found an average AI-native query length of 7.2 words versus 2.8 words on traditional Google search, a 2.6x ratio broadly consistent with Google's 3x claim.

Brands that built their entire SEO and SEM strategy around one- to three-word keyword targeting now operate with infrastructure designed for a search paradigm that is actively shrinking.

Which brings us to an announcement Google did not put on the main stage but should have.

Google's Chrome Lighthouse, the tool that grades website performance and indirectly informs SEO ranking, just added an experimental "Agentic Browsing" audit category. The new audit checks for the presence of an llms.txt file, which is a machine-readable summary of a website's structure and primary content designed to help autonomous agents navigate and extract data efficiently. Sites that lack one or that return errors on the file will be flagged.

Per Search Engine Land coverage from earlier this month, Google currently says that the absence of an llms.txt file will not affect search rankings or Lighthouse scores, though history suggests this position has a shelf life. When Google added robots.txt audits to Lighthouse years ago, the same "won't affect your score" language became a meaningful ranking signal within roughly two years. The pattern is consistent enough that treating the current "no impact" framing as a stable state would be a mistake.

That means organizations that are currently blocking AI crawlers (GPTBot, ClaudeBot, PerplexityBot) to protect content from being scraped find themselves in a tricky position with Google, which is now asking them to publish a machine-readable site summary explicitly for AI agents to consume.

This dynamic extends far beyond a content team decision and squarely into the territory of the CMO and general counsel's involvement, and it needs to happen before the holiday, not after.

The bigger structural shift in Search itself is generative UI. Robby Stein's I/O demo showed Gemini 3.5 Flash and Antigravity building custom interactive layouts on the fly per question.

Robby's shopping example was telling: “If you're shopping, we give you products.” Search results are no longer just about predictable SERPs with consistent positions. Soon, we’ll expect dynamically composed UIs, a product card might appear alongside an interactive widget, a historical pricing chart, or whatever else Search decides to generate for that query.

Findability optimization now has to account for the site experience being non-deterministic for each question.

While I find this exciting (largely because it is a fulfillment of a long-term prediction of Future Commerce), it is the ultimate “works on my machine” hellscape.

The Clean Data Problem

Here is the question we asked Suresh Ganapathi, Google's Senior Director of Product Management for Consumer Shopping, in our interview at I/O: “Does being on Shopify accelerate a merchant's ability to benefit from [Universal Cart]?”

It’s a question that still needs a solid answer.

We asked Ashish Gupta similar questions. The answer, paraphrased fairly, was that merchants of all stripes can integrate via the open UCP spec or through Merchant Center, and that Google is working on simplified onboarding tooling.

Both answers are technically accurate.

But the operational reality is that the prerequisite for benefiting from the bright and shiny Universal Cart-shop-anywhere future is something many merchants do not have: clean, structured, conversationally enriched product data, connected to loyalty data.

One does not simply “walk into Mordor,” and one does not simply “do” structured and enriched data. To have that data, you must first have the organizational discipline to maintain it as the model and protocol evolve.

Gartner's most recent survey of 248 data management leaders found that 63% of organizations either do not have or are unsure they have the right data management practices to support AI. Gartner separately predicts that through 2026, organizations will abandon 60% of AI projects unsupported by AI-ready data.

Even before factoring in the specific demands of UCP and agentic shopping, a majority of enterprises do not have a defensible answer to the question “Is our data ready?” Layer on the requirement to answer real-time agent queries like "Is this Cat5 ethernet cable compatible with the Wi-Fi mesh system I just added to my cart?" and the readiness gap widens even further.

The work stack will look something like this for businesses that haven’t begun to assess their investments yet:

• Adding the six new conversational attributes at the SKU-level
• Building a supplemental product feed (which Google “strongly suggests” to avoid polluting your primary feed)
• Reconciling agent-aware loyalty rules
• Integrating with Google Wallet for payment methods/perks/merchant offers
• Maintaining tax nexus accuracy across new geographies as UCP rolls out
• Shipping integration, speed, and rate quotes

Each of these is independently solvable. Stacking them, prioritizing them, funding them, and assigning organizational ownership is the actual challenge, as it has always been.

This is the gap between what the keynote shows and what your operations team will do over the next 18 months. 

The upside: for more businesses to get on board with UCP, we’ll need more qualified service agencies to deliver the products and services, which means more Google Cloud and AI qualifications and (most likely, as we have seen in the past) funding from Google to make that happen. 

Where Consumers Will Actually Encounter Universal Cart (Hint: It's Ads)

A theme unifies all of the Google I/O announcements.

Despite the consumer-facing language of Universal Cart, Spark, and AI Mode, the place where most shoppers will first encounter UCP-powered commerce is not on a YouTube creator's video page.

It is in a Google ad.

The GML announcements made this explicit for me, at least:

• UCP-powered direct checkout will appear in Demand Gen campaigns on YouTube
• We will be able to deliver Direct Offers in AI Mode with one-tap purchase in Ads
• Affirm and Klarna BNPL are now available directly into Google Pay at the ad level
• Ask YouTube, the new conversational search experience inside YouTube, features new Ads capabilities

The law of acquisition: if “organic is dead” is true, and the “zero-click phenomenon” is really here, then the brands that get the most lift from agentic commerce in 2026 will be the ones that pair full UCP integration with aggressive spend on AI-powered ads.

The combination gets us as close to a frictionless funnel wet dream as demand gen marketers could ever dream of: a shopper sees an ad, asks a follow-up question in a conversational “brainstorm,” gets a personalized response from Gemini, adds the product to Universal Cart, and checks out in three taps.

That is genuinely powerful and supremely faster than the traditional “funnel.”

The promise of the ‘handoff’ to the merchant is telling, though. At any point, the consumer can hand the cart off to the merchant and migrate it to the merchant’s branded site, indicating a live, persistent connection between UCP and the merchant.

Google's commercial model is not changing; it’s probably increasing its dominance. Ads are still the dog that wags the tail. UCP and Universal Cart make the post-click experience more competitive against Amazon’s watered-down “Buy for Me” offering, but the cost of the click remains unchanged, and the consumer attention you need to compete for is now being mediated by Gemini's intelligence layer rather than the searcher's own.

As Srinivasan put it from the GML stage: “The other way you can get discovered is with Ads.”

In 2026, my Sister in Christ, the only way you can get discovered is with Ads.

What to Do This Quarter

If you are a commerce executive trying to operationalize this Brief, the work breaks into four concrete takeaways:

1. Conduct a clean-data audit before anything else. Inventory the maturity of your product data, your first-party customer data, and the connective tissue between them. If you cannot answer “Is this product compatible with that product?” from your own systems, you are not ready for the Universal Cart to ask the question on your behalf.
   
   
2. Run a token cost model for at least two scenarios. Calculate your current AI spend against the same volume of work shifted to agentic workflows that consume five to ten times as many tokens per task.
   
   Take that number to your CFO (before you take it to the AI vendor’s sales team).
   
   
3. Do not plan on UCP integration in Q3 of 2026. Be present in the conversation, prepare your data, and track consumer adoption through the holiday season. Unless there is a clear, offered incentive to do so, there is no need to rush. If Universal Cart is in genuine consumer use by January, integrate hard in Q1 2027.
   
   
4. Invest in the conversational attributes regardless. They are useful for Google, for competing AI shopping services, and for your own on-site search experience. You probably won’t regret this in the long run.
   
   

As William Gibson might have said, the future is here, but [the token cost] is not evenly distributed. Or something like that.

The merchants who do get to Universal Commerce in the coming months will be those who treat Google's keynote as a technical sales pitch, albeit a great one, while understanding that the future that they buy into will be one of their own making, with careful planning and consideration along the way.

The rest? They’ll be saying a lot worse than “whiskey-tango-foxtrot.”