In short:Project Mariner and Google agents: what merchants must prepare now — December 2024. Google presents Project Mariner. An AI agent based on Gemini 2.0 that navigates the web alone — fills forms, clicks buttons, complètes purchas
2024Project Mariner launched by Google
3major players (Google, OpenAI, Anthropic) with navigation agents
2026-27Estimated public deployment window
Project Mariner: what’s really happening
December 2024. Google presents Project Mariner. An AI agent based on Gemini 2.0 that navigates the web alone — fills forms, clicks buttons, complètes purchases.
Not a lab demo. A signal of strategic direction in active deployment.
OpenAI is rolling out similar capabilities with Operator. Anthropic is testing autonomous navigation agents. The ecosystem converges toward the same point: agents that act on the web in place of humans.
For an e-commerce merchant, the question isn’t « will this arrive? » The question is « is my site readable to these agents? »
2024Project Mariner launched by Google
3major players (Google, OpenAI, Anthropic) with navigation agents
2026-27Estimated public deployment window
How agents navigate e-commerce sites
An autonomous agent like Project Mariner doesn’t navigate like you or I do.
Humans: see images, read titles, interpret design.
Autonomous agent: analyzes HTML DOM, reads button labels, extracts structured data, follows links with explicit aria-label attributes.
A standard e-commerce site is designed for human eyes. Not for an AI agent reading the underlying code.
The friction points identified on current e-commerce sites:
« Add to cart » buttons without aria-label attribute including product name
Prices displayed in CSS only — not in the DOM text
Variant selectors (size, color) without accessible labels
Payment processes with CAPTCHAs and anti-bot obstacles blocking legitimate agents
Product data in images — text not extractable
Accessible catalogs vs. opaque catalogs
In the economy of autonomous agents, catalogs will divide into two catégories:
Accessible catalogs. Semantic HTML, complete structured data, accessible labels on all interactive elements, linear and decomposable checkout process. These catalogs will be navigable by agents — which means they can be recommended and purchased through agent interfaces.
Opaque catalogs. Data in images or undocumented frontend API calls, buttons without semantics, variants handled by inaccessible JavaScript, indiscriminate anti-bot obstacles. These catalogs will be invisible to agents — and thus absent from AI recommendations.
The distinction isn’t cosmetic. It’s the difference between existing or disappearing in the agent economy.
The 6 technical prerequisites
Prerequisite 1 — Semantic HTML and accessible labels. Every button, every link must carry a precise aria-label. "Add to cart" tells an agent nothing. "Add Bodum 350 ml Coffee Maker to cart" tells it everything. On WooCommerce, the WooCommerce Accessibility plugin does the job.
Prerequisite 2 — Complete Schema.org Product (the 12 fields). Agents read the Schema before clicking. Complete Schema pre-qualifies the product. The agent knows what it's looking for without visiting every product page.
Prerequisite 3 — Public product API or extended Open Graph. Expose your product data via REST API or extended Open Graph. The agent retrieves price and stock without crawling HTML. WooCommerce exposes a REST API by default — verify it returns real-time data.
Prerequisite 4 — Linear, labeled checkout process. Cart → address → shipping → payment → confirmation. Each step carries a clear label. Surprise modals, step skips, CAPTCHAs break the agent.
Prerequisite 5 — Agent access policy in robots.txt or llms.txt. Declare who can access. Google-Extended (Gemini / Mariner), OpenAI-GPTBot, Anthropic-claude-bot. Blocking all bots also blocks legitimate agents.
Prerequisite 6 — Real-time availability in metadata. Agents decide now. A product shown available in Schema but out of stock at checkout creates an error the agent handles poorly. Sync real-time stock with Schema and API.
Attention point: Anti-bot solutions (Cloudflare Bot Fight Mode, certain WAF configurations) by default block legitimate shopping agents. Verify your security rules include a whitelist for GoogleBot, GPTBot, ClaudeBot, and PerplexityBot. These agents need to crawl your catalog to recommend you.
What arrives and when
2024 — Demonstration
Project Mariner presented. Navigation agents in private beta. Usage limited to Google testers.
2025 — Selective deployment
Agents available for certain professional uses. First agent purchases documented. Low but growing volume.
2026 — Transition
Progressive public deployment. The first "agent-ready" merchants capture agent traffic.
2027 — Standardization
Autonomous shopping agents become a standard channel. Non-compliant merchants are structurally absent.
Prepare infrastructure before the wave
Public deployment of autonomous agents is 12-18 months away. You have that window to make your catalog readable.
Good news: the technical prerequisites for shopping agents are the same as best practices for accessibility and structured data. These are investments that already improve your SEO and user expérience. Today.
The infrastructure you build for Project Mariner improves your site for all current users. ROI is immediate. Competitive advantage on agents is deferred — advantage on humans is already there.
How Project Mariner works technically — what tests reveal
Project Mariner is not an improved chatbot. It's a web navigation agent that uses a multimodal model to "see" and "interact with" a browser interface the way a human user would. The difference from previous shopping assistants? Fundamental.
The multimodal architecture
Mariner uses Gemini in vision mode. It takes screenshots of the browser, analyzes them pixel by pixel, identifiés interactive elements — buttons, form fields, menus — and generates actions: clicks, text input, scrolling, navigation.
Not an API that queries your catalog. An agent that actually navigates your site. Like a shopper. At machine speed. It sees what your human visitor sees. It makes decisions based on what it sees.
83% of shopping tasks tested in Project Mariner public demos completed successfully — Google DeepMind data, December 2024
What beta tests reveal about navigation behavior
Early feedback from developers with preview access reveals three key behaviors.
Mariner prioritizes visible structure. Clear menu, accessible category filters, functional search bar: it spots them immediately and uses them first. Opaque navigation or obscure URLs slow it down. Sometimes significantly.
It reads customer reviews. Really. In public demos, it was seen browsing review summaries before adding a product to cart. Review density and quality directly influence its recommendations.
It detects friction. Long checkout, mandatory account creation, CAPTCHAs: each obstacle increases failure rate. Mariner isn't yet able to solve all CAPTCHAs. For now.
Documented current limitations
Heavy JavaScript sites (SPAs without SSR) cause rendering problems. Mariner sees a page loading. It doesn't always know to wait for asynchronously loaded elements.
Dynamic prices and real-time availability require correct page refresh. If the agent sees one price displayed and another appears at checkout (undisclosed shipping, tax added to cart), confidence drops.
Key point: Mariner evaluates your site like a demanding user with perfect memory. Any inconsistency between what's displayed and what's billed is detected. It can interrupt the purchase flow.
The 3 catalog types advantaged by Google agents (and why)
Not all catalogs are equal to agents. Agents optimize for task completion — they go where the task complètes easiest and with most confidence. Three catalog types have a structural advantage.
Type 1 — Catalogs with complete structured data
Schema.org Product with all properties filled: price, availability, SKU, description, reviews, multiple images. Mariner can read this data directly from the DOM — it doesn't need to "see" it in the interface, it can extract it directly.
A catalog with 100% of product pages enriched in Schema Product has a processing speed advantage of 40-60% on agents — estimate based on automated navigation benchmarks published by the Chrome Automation team.
40-60% processing speed gain for an agent on a catalog with complete Schema Product vs a catalog with no structured data
Type 2 — Catalogs with predictive navigation
An agent looking for "waterproof trail shoes size 42" navigates faster on a site whose URLs, filters, and catégories reflect these attributes. Predictive navigation: the agent guesses where to find what it wants without testing ten paths.
Descriptive URLs (/trail-shoes/waterproof/). Working facet filters. Breadcrumbs with BreadcrumbList schema. Clear category titles, no ambiguity. Technical SEO fundamentals — their impact on agents is multiplied.
Type 3 — Catalogs with transparent checkout
Agents optimize to finish the purchase. Checkout in 3 labeled steps. Total visible at each step — product price, shipping, tax. Immediate order confirmation: the agent validates the task with certainty.
Opaque checkout — fees added at final step, confirmation wait times, third-party redirects — creates uncertainty. That uncertainty produces abandons or weaker recommendations.
Agent-ready preparation checklist: 12 points
This checklist compiles automated navigation criteria documented in Mariner specifications, crossed with feedback from private beta teams.
Structured data (4 points)
Schema Product on 100% of product pages with: name, description, image, offers (price, availability, currency), aggregateRating (if you have reviews)
Schema BreadcrumbList on all category and product pages
Schema Organization on homepage with address, contactPoint, sameAs
Complete Open Graph on all indexable pages (og:title, og:image, og:price:amount)
Navigation and architecture (4 points)
Descriptive URLs: each URL segment reflects the catalog's thematic hierarchy
Functional facet filters for key attributes (size, color, material, price)
Primary navigation in pure HTML, no JavaScript dependency for menu rendering
Internal search with auto-complete and relevant results — agents use search as a primary entry on large catalogs
78% of automated agents use the search bar as primary entry on catalogs over 500 items — internal tests 2024
Checkout and trust (4 points)
Transparent total price: shipping calculated and displayed before payment page
Guest checkout accessible: zero mandatory account creation, zero friction
HTTPS on entire funnel: valid certificate from cart to confirmation
Immediate order confirmation: stable URL. No temporary page that disappears
What merchants in betas learned — real expérience
Early feedback from merchants in Project Mariner beta — and on OpenAI Operator or Google Shopping Actions — all say the same thing.
Load speed is critical — more than classic SEO
An agent navigates and waits. Your page takes 4 seconds to load? The agent waits 4 seconds. Across 10 pages in a comparison journey, that's 40 seconds of waiting — and a much higher probability of abandon or timeout than a patient human.
Beta merchants report that sites with LCP under 2.5 seconds have agent completion rates 2.4x higher than sites over 4 seconds.
2.4x better agent completion rate for sites with LCP < 2.5s vs LCP > 4s — compiled beta merchant feedback 2025
Counter-intuitive to classic SEO: very short descriptions but structured in attributes (key attributes in list, main benefits in max 3 points) are better handled by agents than long narrative descriptions.
The agent doesn't need to read 600 words to understand a jacket is waterproof, lightweight, and available in 5 colors. Those 3 attributes in HTML list are enough. Long description stays useful for SEO and human persuasion — but for the agent it's noise to process.
Recent reviews weigh more than overall rating
A product with 12 reviews from the last 3 months generates more agent confidence than a product with 200 reviews, the last from 18 months ago. The agent interprets review freshness as a signal that the product is still sold, still appreciated, still relevant.
Direct action: set up systematic post-purchase review collection. An 8% collection rate (8 reviews per 100 orders) is achievable with a simple J+14 email. On 500 orders per month, that's 40 fresh reviews monthly.
2026-2027 perspective: Google agents aren't yet deployed at scale for e-commerce transactions. The preparation window is open. Merchants who structure their catalog now will lead when deployment accelerates — exactly like those who invested in technical SEO in 2018-2019 dominated organic traffic in 2022-2024.
Frequently asked questions
Can Project Mariner complete purchases on any e-commerce site?
Not yet reliably. In beta, Project Mariner works correctly on sites with clean semantic HTML, linear checkout processes, and no aggressive CAPTCHAs. On sites with complex UX (heavy JavaScript, modals, opaque multi-step processes), the agent fails or abandons. Your site's technical compliance directly determines whether Project Mariner can use it.
Do I need to create a dedicated API for shopping agents?
Not necessarily. WooCommerce already exposes a native REST API covering most needs (catalog, price, stock, orders). Shopify too. Priority isn't creating a new API, but ensuring your existing API returns real-time data, product endpoints are publicly accessible (no catalog access restrictions), and Schema.org data on pages is consistent with API responses.
Will autonomous agents replace humans for recurring purchases?
For low-decision recurring purchases (consumables, replenishment, subscriptions), yes — within 2-3 years. The buyer configures the agent once ("restock my office supplies when stock hits 20%") and the agent executes. For high-decision purchases (equipment, premium, new products), the agent will play a search and recommendation role, but final decision stays human longer. Immediate impact is on B2B recurring and B2C consumable purchases.
How do I know if an agent tried to navigate my site?
Check your server logs. Agents like GoogleBot-Extended, GPTBot, ClaudeBot, and PerplexityBot identify themselves in the User-Agent string. Filter your Apache or Nginx logs for these identifiers. If these bots access your product pages and cart, agents are trying to navigate your catalog. Cloudflare Analytics and tools like Botify also show these accesses in dedicated reports.
What's the difference between Project Mariner and other AI shopping agents like ChatGPT Shopping?
ChatGPT Shopping recommends products in a conversational interface — the buyer then clicks a link and complètes purchase themselves on the merchant site. Project Mariner goes further: the agent navigates the site, selects the product, fills the form, and complètes the order autonomously. The distinction is fundamental for merchants: ChatGPT Shopping = referral traffic. Project Mariner = autonomous transaction with no visible traffic in GA4.
"Agent-Ready" audit of your catalog
I test your site's navigability by autonomous agents, identify technical friction points, and provide you with a compliance roadmap before large-scale deployment.
