Using Convergence.ai's “Proxy” for Efficient Meta Data Creation
Finding new ways to use artificial intelligence for mundane marketing tasks is never not… absolutely invigorating.
Especially when you experience a personal workflow breakthrough which significantly improves your efficiency and ability to scale.
Combined across tasks and the various AI agents you employ, you may add hours to your workday. And that eventuality is nothing to sniff at. And I have a new suggestion.
With the addition of Convergence.ai’s Proxy tool’s latest update, it’s officially joined the ranks of CUAs (Computer Using Agent) like Project Mariner and Azure. And, brother, you need to stop what you’re doing and go take a hard look.
The right AI application can make for a staggering amount of precious reclaimed time. Just this very day, I discovered how Convergence.ai's “Proxy” tool can transform the tedious process of writing HTML meta titles and descriptions into a streamlined, efficient and easily replicable workflow. Even for something as fundamental (yet still 2025 mission critical) as meta data creation / revision.
My experience was so resounding that I want to share some tool and task-specific tips. I sat down this morning intending to tighten up a few aspects of our infant agency site. Not to spend more time than I’d like to admit writing a marginally obsessive new blog post. That’s a kind of testimonial before we even get started. So I should probably get started.
Convergence.ai Proxy Review
Whether you're an online marketing veteran, or just starting to explore how AI can grant ridiculous efficiency gains when applied properly, this post will help you employ Proxy for quick and accurate meta data authoring, but by the end of next week I’m confident I’ll have it doing laundry.
Because titles and descriptions are the tip of the iceberg. If the kind folks at Convergence are inclined to let me continue to explore possibilities using their Pro-level subscription “on the arm” - I definitely wouldn’t complain. Now, here come the brass tacks:
Table of Contents
What is Convergence.ai?
They’re a London-based AI startup that has developed "Proxy," an advanced AI agent designed to learn how to consolidate tasks into time-saving and scalable workflows.
Founded by machine learning engineers Marvin Purtorab (CEO) and Andy Toulis (CTO) in April 2024, the company aims to redefine how we interact with software by creating personalized AI assistants that can free users from repetitive tasks.
What differentiates Convergence.ai (somewhat) is their focus on creating AI that acquires new skills like humans do, through what they call Large Meta Learning Models (LMLMs). These models are trained to learn independently and retain information through long-term memory, allowing them to adapt to new tasks without requiring extensive retraining.
I’m talking to you, Perplexity.ai.
I’d like to know more about whether this learning occurs only at a macro level, based on everyone’s prompts and usage, or at a micro, thread level. I hope it’s the latter. I’ve been dying for a version of a Perplexity.ai “Space” which incorporates directives given in the information settings or an uploaded file - but also retains knowledge naturally and at the Space level.
The company's mission is clear: "to let people everywhere reclaim their time using the power of machine learning." They believe we're at a pivotal moment in history where AI can reshape our world and enhance human capabilities in unprecedented ways.
I like their angle. As I frequently say, “AI is not a solution - it’s an enhancement.”
Act accordingly, and you’ll have an immediate advantage over everyone else trying to find efficiency gains and practical business applications. On that note, you may get something out of another article I wrote recently which dissects the difference between an AI assistant and an AI agent.
The Story Behind Convergence.ai
The founders of Convergence.ai, Marvin Purtorab and Andy Toulis, have an impressive background in machine learning. They first met while working at Shopify on recommender systems and AI assistants before joining Cohere, an AI platform for enterprises.
They had "been playing with agents for, like, a few years, like three or four years now," but realized the timing wasn't right until recently. This insight led them to leave Cohere and establish Convergence.ai, bringing together specialists from leading AI organizations including Google DeepMind, Meta, and PolyAI.
The company secured a remarkable $12 million in pre-seed funding led by Balderton Capital, with participation from Salesforce Ventures and Shopify Ventures. This substantial early-stage investment highlights the industry's confidence in their vision and technical approach.
What's particularly impressive is how quickly the team moved from concept to product. In just three months, they progressed from an idea to launching their beta product, demonstrating exceptional execution capabilities.
How Proxy Stands Out from Other AI Tools
Convergence.ai's Proxy differentiates itself from competitors like OpenAI and Anthropic in several key ways:
Versatility and Adaptability: Unlike most AI agents designed for specific workflows, Proxy works across various tasks and domains. It acquires skills like a human would through long-term memory and continual learning.
Learning Capability: Proxy uses Large Meta Learning Models (LMLMs) that can learn and improve during real-time use. As Purtorab explains, "Imagine an agent that keeps growing and learning whenever you teach them something new”.
Memory Integration: Memory is a core component of Proxy's architecture, allowing it to remember tasks for the long term and use these memories to handle new situations - similar to human learning.
Interface Approach: Proxy is built to reason like a human on the same online interfaces we use, making it easier to teach the agent as you would teach a colleague or friend9.
Continuous Improvement: Unlike traditional models limited by their initial training data, Proxy can seek new information and develop continuously using feedback from users without requiring large training datasets.
According to James Wise, Partner at Balderton Capital, "Few people have the experience and skill that Marvin and Andy have, which makes them well-placed to take on the complex technical challenge of a product like Proxy".
Industry analysts have noted that Convergence.ai's approach could potentially disrupt the market dominated by giants like OpenAI and Anthropic, with reports suggesting Proxy offers superior performance in certain tasks.
As mentioned earlier, with the addition of Proxy’s latest retooling (more on that below) it can now be classified as a CUA (Computer-Using Agent). The way you can actually watch it’s process as it interacts with the web is truly something to see.
Convergence.ai Cost
Convergence offers a Pro Plan that starts at $20 per month which includes unlimited sessions, has a maximum of 20 sessions running in parallel, and access to 20 automations. It’s aimed at advanced users. There is also an Enterprise Plan with custom pricing, as they make an effort to help you integrate Proxy into your business.
If you want to test the product without having to sign up for a paid plan, Convergence has a generous free version offering basic features through which new users can get a taste of the core functionalities.
Personally, I much prefer tools offering a freemium version instead of (or in addition to) a short, free trial. I’m more likely to warm up to, find customized ways to use, adopt, and eventually pay for AI agents which don’t make you rush to find time to kick the tires more comprehensively. I don’t expect to be let into the VIP lounge if I’m wearing sneakers, but don’t throw me out the back door of the club, either.
Here’s a more comprehensive list of your option:
Free Plan: $0/month – 5 free sessions per day, 1 automation.
Pro Plan: $20/month – Unlimited sessions, 20 automations, up to 20 parallel sessions.
Enterprise Plan: Custom pricing – Tailored features for large organizations.
The site includes a Template Studio which makes it easier to create reusable automations, and the Template Hub where the user community can share their creations. They’ve also recently introduced Proxy Lite, which is an open source VLM.
And only a week or so ago, they announced a major Proxy update, available to Pro and Enterprise, users called Deep Work. There isn’t a lot of press available online yet, but I saw one user describe DW as, “goal-oriented, not instruction-based“, and it seems positioned as their answer to options like OpenAI’s Operator.
If you’d like to learn more about it, the video above is your best bet, as the host works through some of the new functionality while sharing his screen. You have my attention, Convergence folks. DW may be the straw which breaks my wallet’s back.
My Experience Using Proxy for Meta Data Creation
I recently used Proxy to refresh the meta data for our entire website (this one). While I had experimented with Proxy before, this was my first time using it for a substantial, project-focused task.
The process began with me logging into https://proxy.convergence.ai and crafting a prompt that would guide Proxy in creating meta titles and descriptions. I used an updated version of a self-authored prompt I knew from experience tends to provide decent results with other AI tools, though typically requires multiple iterations:
"You are a professional website meta title and meta description writer with 30 years of experience doing so across verticals, with a deep knowledge of meta data and crafting convincing title and description best practices. Comprehensively rewrite the meta data on every page of https://www.fastfrigate.com. Give me three title/description suggestions for each page. Limit meta titles to 80 characters and meta descriptions to 180 characters. Use the content of each page to write accurate and engaging meta data unique to the content and intent of each individual URL, keeping best practices for calls to action and high click-through-rate potential in mind. The audience are business owners and marketing professionals wanting to learn how to increase traffic to their websites, and other aspects of online marketing."
What impressed me most was how Proxy handled this task with minimal guidance. I’m comfortable saying: It (almost) got it right the first time.
While tools like Perplexity.ai and ChatGPT typically require multiple human interventions to achieve optimal results, Proxy delivered higher-quality suggestions in it’s first pass.
Step-by-Step Guide to Using Proxy for Meta Data
If you want to replicate my early success using Proxy for meta data creation, here's a detailed walkthrough of today’s process.
Preparation: Before starting I uploaded my personal and frequently curated "no-ai-words.txt" file, to ensure the suggestions wouldn't appear AI-generated. This helps maintain a natural, human tone that resonates with readers and search engines alike, and I use said file all the time, regardless of task or tool. I uploaded it easily through Proxy’s main prompt entry field. Feel free to use the file yourself. Or improve on it, send it back, and I will re-share.
Prompt Creation: I customized a proven and detailed prompt that specified:
The exact task (meta data creation)
The website in question
Character limits (80 for titles, 180 for descriptions). Let me save you some time - I know the 80/180 length is outside of the established SEO best length practice. I like to give AI a little creative wiggle-room. I review and edit, as all humans should, any AI-generation before publishing.
Audience information
CTAs and CTR should be a consideration when generating suggestions.
Desired output format
URL Collection: For future Proxy meta tasks, I'll collect a comprehensive list of URLs from a sitemap, Screaming Frog, or whatever. This would only add about 15 seconds to prompt creation, but ensure complete coverage.
Execution: Submitted my prompt and observed as Proxy navigated through FastFrigate.com, analyzing content and generating meta data suggestions.
Review and Selection: Evaluated the three title/description options Proxy provided for each page, selecting the most compelling iteration. Or, more frequently, combining my favorite elements of the three examples into one kick-ass final version.
Refinement: Made minor adjustments to the selected meta data as needed, ensuring they align with your brand voice and SEO strategy.
Implementation: Applied the finalized meta data to FF’s pages.
Unique Features That Enhanced My Workflow
Several Proxy-specific differentiations made my meta data creation process particularly efficient:
Saved Credentials: I had to put this one at the top of the list. You’ve seen mention of how Proxy’s ability to interact with web pages is a standout. Wouldn’t it be great if it could log into email, Google Sheets or even your SEMRush account without any assistance? It can.
Add your logins to their credential manager, and any time your instruct Proxy to use something specifically, or it decides it want to use a paid or otherwise locked down website at which you have a subscription - it’ll just go right ahead and login as one of its task steps. Very cool, indeed.
Visual Page Recognition: One thing I loved right away was how Proxy displayed right-appearing page images as it worked through each page-specific task. This visual browser context helped me quickly verify that Proxy was indeed analyzing the correct pages.
Contextual Navigation: Proxy didn't just stay on our internal pages, it followed links from our client portfolios to gather additional context. This was impressive - and resulted in more detailed meta recommendations for those.
Integrated Browser: The orange globe button in the prompt entry field opens a browser version of whatever you're going back and forth with Proxy on. This feature allows it to directly interact with web content, giving it a more comprehensive understanding than tools that rely solely on text prompts.
Memory Retention: Unlike other AI tools that treat each interaction as isolated, Proxy remembered my preferences from earlier threads and applied similar styling to later suggestions, creating more consistent meta data across the site. The prompt iteration process was also much shorter than with other tools I use every day.
This browser capability likely helps Proxy better understand page layout, visual elements, and content hierarchy - all crucial for creating effective meta data. It’s also super-awesome for helping Proxy get around CAPTCHA-type bottlenecks, and several times it paused the step it was on and asked me to do so.
Maybe I’ve just been living under a rock, but I’ve never seen anything like that in an AI tool before.
Results and Performance Analysis
My Proxy-powered meta refresh was memorable. In a good way. Or I wouldn’t have spent half my Saturday writing about it. Goose needs a shite walk. And is definitely pissed.
While Proxy didn't include every site page comprehensively - missing some items not in the main navigation and pages linked from dropdown menus (like Fast Frigate's individual client pages and the blog posts under our Resources section) the quality of what it did produce was exceptional.
For each page it did process, Proxy generated three sets of title/description suggestions that required minimal editing. The suggestions demonstrated:
Strong understanding of page content and intent
Appropriate keyword inclusion without 1999-era stuffing
Compelling calls-to-action to encourage clicks
Adherence to character limits while maintaining readability
Consistent brand voice across different page types
The efficiency gain was substantial compared to my previous experiments with other AI tools. What would typically require multiple prompts, refinements, and human intervention was accomplished in (comparatively) one pass with Proxy.
Limitations and Areas for Improvement
Despite its impressive performance, I did identify some limitations in my experience with Proxy:
Incomplete Coverage: As mentioned, Proxy missed some pages not included in the main navigation or linked from dropdown menus. In future projects, providing a comprehensive URL list will address this issue. No big whoop.
Unexpected Task Closing: I did not like the way Proxy decided to close certain tasks, making them an archive of sorts, when my intent for it was to be one task of several leading up to a goal.
Maybe I just have more to learn, or maybe I should be employing the goal-oriented focus mentioned elsewhere in this post, but at least twice the prompt field became locked as I was in the middle of writing a next step forcing me to start over. I’m left wondering if there’s a time limit and will update this post after learning more.
While each task thread in the left hand nav has a menu with the option to “Repeat Session”, Proxy’s level of research and detail can cause even simpler tasks to take a long time. But I was ready with the follow up prompt, entered it quickly after step 1 was completed, and eventually reached the goal.
Stoppage Between Steps: 16 pages equals 16 steps. Proxy kept stopping after one or two steps before the entire meta rewrite task was completed. Then it would say, “Next, I will proceed to gather content for the 'Work, 'Process', and 'Resources' pages to craft similar meta titles and descriptions.“
To which I would reply: “Great. Do that. What are you waiting for?” For future tasks I will try to be even more specific about wanting it to carry all steps through to completion the first time. Having to sit, watch, and coax it to continue 12 out of the 16 times removed any efficiency gain. This is a pretty common annoyance for many of the AI tools I’ve tried to date, and I’m sure there’s a way to keep it from happening.
No file upload ability: I’m still scratching my head over this omission. I had a GSC Excel export I wanted to include in a task. Where’s the file upload button? To get what I needed I imported it to Google Sheets and gave Proxy the link instead. But until I realize I’m missing something - like maybe you can only attach files as part of an automation - my scalp will remain raw.
Occasional Repetition: For very similar pages, Proxy sometimes created meta descriptions that were too similar to each other, requiring manual differentiation. But that was not a chronic occurrence.
Learning Curve: While powerful, some of Proxy's advanced features require experimentation to fully grasp.
Subscription Limitations: As a newer business, Fast Frigate is currently using the basic subscription level, which does not include all of Proxy's advanced capabilities.
But that’s a “me” problem. And the other three callouts shouldn’t discourage anyone now thinking of taking Proxy for a spin, either.
Update - 3/23/25: Feeding Proxy every full site URL as part of the prompt, which I suggested I’d try above, was not necessary. For my next attempt I merely added the following sentence to give it more context around site structure and its indexing became flawless: “Comprehensively look for all main navigation, client work, and resource section blog pages”. But the pausing between steps for no apparent reason, with me having to tell Proxy to continue, persisted.
Note: If you are a Proxy user, or even developer, and have identified anything I’m doing wrong - I invite you to email me with details I will then update this review with.
Key Takeaways
Efficiency Boost: Proxy dramatically reduced the time required for meta data creation compared to traditional methods and other AI tools.
Quality Output: The generated meta titles and descriptions were high-quality and required minimal editing.
Visual Context Helped: Proxy's ability to see and understand web pages visually lead to a more intuitive experience than text-only AI tools.
Preparation Pays Off: For optimal results, provide Proxy with comprehensive information about your website structure.
Continuous Learning: Proxy's ability to learn from interactions means it gets better with continued use on the same website or similar projects.
Conclusion: Is Proxy Worth It?
In a word, “friggin’ absolutely - yes”.
Based on my experience, Convergence.ai's Proxy represents a significant advancement in AI-assisted marketing-related tasks. Its ability to understand context, learn from interactions, interact with the external web in real time, and produce high-quality meta data with minimal guidance, sets it apart from many competitors.
For digital marketing professionals looking to enhance efficiency without sacrificing quality, Proxy deserves serious consideration.
Again, in case anyone important is listening, I would love to try the pro subscription level. As Fast Frigate is just getting started, it's not yet in the budget. Perhaps, if the good folks at Convergence.ai see and appreciate this post, they'll offer a free month for testing.
But I didn’t spend (way too much) time writing about them to get a free bowl of soup. I genuinely want readers to know that Proxy stands out in a deep sea of competing choices, and I'm confident it has a bright future ahead.
Whether you're managing a single website or overseeing online marketing for multiple clients, tools that enhance efficiency without compromising quality are invaluable.
Convergence.ai's Proxy has (probably) earned a permanent place in my digital marketing toolkit, and I suspect it will do the same for many of you.
