How We Built Hyper, the Best AI for Meta and Google Ads (Featured on Supabase)

Over the last twelve months, across more than $100 million in connected ad spend, the accounts Hyper runs have seen a consistent 54% lift in ROAS and 20% to 27% lower CPMs, about 24% on average. We built the system that does this on Supabase, who featured the story. This is how it works, and why we think it's the best AI for Meta and Google ads.

The result

Most AI marketing tools tell you what they'd do. Hyper does it. It researches the brand, builds the campaign, launches it into Meta and Google, watches it, and changes what isn't working. The point of all of that is performance, so let's start there.

Across the accounts we run, over twelve months, ROAS came up about 54% and CPMs came down 20% to 27%, about 24% on average, versus each account's prior baseline. Those are blended averages across a wide mix of businesses, not a cherry-picked best case, and they hold because the same few things drive them every time.

What drives the lift

The 54% isn't one trick. It's a handful of jobs, done well and done constantly.

Creative testing and a real feedback loop. Meta now rewards a steady supply of genuinely different creative, and the win rate on any single ad is low, so the teams that win are the ones that test the most. Hyper's agents watch for creative fatigue, kill what's tired, and keep fresh variations flowing, which is most of where the lift comes from.

It watches the whole internet for patterns. The agents are constantly scanning where your market actually is: Instagram and TikTok for what's trending in your niche, Reddit for what your customers are really talking about, and the Meta Ad Library for which ads your competitors are running. That running read of the internet is the raw material for everything downstream.

Creative that performs in production, handed to your team on a silver platter. The agents don't just generate ads in a vacuum, and they don't just analyze. They make creative built to perform once it's live on Meta and Google, then they hand your team the brief: here are the ten references trending in your niche right now, here are your top competitors and what's working for them, and here are start frames and concepts with actual graphics to generate from. Pulled from that constant scan plus what's performing across our own anonymized network, and pointed at the people who build the creative, it's what makes each batch better instead of just bigger.

Audience targeting and customer segmentation. Advantage+ and Andromeda took a lot of the manual audience controls away and replaced them with a black box. Hyper works with that, segmenting customers and shaping targeting signals to counteract the parts of the automation that quietly waste spend, which is a big piece of the CPM reduction.

Execution, so none of it sits in a slide. A recommendation you have to action by hand is half a recommendation. The agents take care of the execution, so the insight actually gets shipped.

Execution

Here's the difference in one line: instead of setting up the campaign and moving all the settings and levers yourself, you give it to Hyper and it does it.

It builds and launches across Meta and Google from a chat. It pulls the brand's products, writes and generates the creative, sets the audiences, budgets, and bids, syncs the campaign to the ad manager, then reads performance back out and optimizes. The full build-launch-optimize loop runs in one place.

That end-to-end execution is the thing most of the category doesn't actually do, and it's the thing the performance numbers depend on.

The harder problem behind it is shipping correct work at scale. Most AI marketing platforms sit on a frontier model and stop there, so when you push back and ask "are you sure, what about this?", they tend to flip-flop instead of holding a position. Agents on Hyper are built to ship correct work at scale: frontier models, plus our own proprietary models and datasets and defined workflows that are dynamic and constantly being optimized. The result is an agent that holds the right call when you challenge it, and changes course only when it should.

We spent twelve months building that depth into the Meta and Google integrations specifically, through a custom API, and we expose it to users as an MCP, a CLI, and an SDK. That parameter-level coverage is the layer most tools are missing.

Autonomy

People ask how hands-off it is. The honest answer is: as much as you want. The agents handle the repetitive execution on their own, the daily checks, the pacing, the fatigue swaps, the reporting. The decisions that deserve a human, like signing off on a new angle or a budget jump, come to you with the context already gathered. You stay in control of strategy; the agent does the work.

Speed and cost

Two numbers matter here, and neither of them is the inflated kind.

You can test about 20x faster, because the agent can spin up, launch, and read variations without a person touching the levers between each one. More tests, found faster, is most of how the win rate compounds.

And on the infrastructure side, because our agents query a real database with SQL instead of calling each ad platform's API live, data retrieval is 10 to 100x cheaper. That's not a marketing figure; it's a literal cost reduction in how the system fetches and reasons over data, and it's what lets us run this at scale.

Twelve months across verticals

This isn't a demo we ran last week. It's what we've been running for twelve months, across a real spread of businesses: restaurants, dentists, courier services, caterers, chiropractors, and other local service businesses, plus ecommerce brands from small shops to large catalogs, and into lead generation and AI search. The metrics hold across that range, which is the part that makes them believable.

The clearest test was a head-to-head. We ran Hyper on 50% of a marketing agency's client social and SEO accounts, and left the other 50% with their human team, same clients, same window. Hyper made all the execution decisions on its half. We wrote up what happened in the agency benchmark case study, and the ecommerce case study walks through the numbers on a single brand. Both sit alongside our original case study, where one team freed up 29 hours a week.

Who uses Hyper

The same loop runs for very different businesses, which is the whole point. A few of the shapes it shows up in:

Marketing agencies

Agencies run Hyper across their entire client book, one operator covering accounts that used to need a team. In our 50/50 benchmark, the Hyper-run half lifted ROAS 61% against the agency's own people.

Ecommerce brands, small to large

From first-product shops to brands spending $2M a month on Meta. The ecommerce case study is a DTC brand at that scale, where 54% more ROAS and 28% lower CPMs is a six-figure monthly swing.

Local service businesses

Restaurants, dentists, courier services, caterers, chiropractors, and other local operators who often have no marketing team at all. For them, Hyper is the team.

In-house marketers and founders

Lean in-house teams and solo founders who need an operator, not another dashboard to learn. One team freed up 29 hours a week and removed the ceiling on how much they could run.

How your data is handled

A note, because it matters and because we want to be clear. Every account on Hyper is anonymized. We do not train models on your data, and we do not hand it to anyone. What we do is use aggregate, anonymized signals from across the network to make the system smarter for everyone, the same way any good ad platform improves from aggregate performance. If you'd rather not be part of that, you can opt out, and nothing changes about how your own agent works.

How we compare to other platforms

Analysis-first tools can often be good at reading an account and telling you what to change. Where they stop is execution, and where they break is consistency. Most run on a frontier model alone and lack the deeper Meta and Google parameters real execution needs. Ryze, for example, runs on lighter models and is missing many of the more complex endpoints, like the granular audience controls that real analysis depends on.

The clearest tell is what happens when you push back. Ask an analysis-first agent "are you sure, what about this?" and it tends to flip-flop, because it has no grounded context to defend a position. In our own evals of consistency under follow-up, general frontier-model agents held the correct call 10% to 20% of the time. Hyper held it about 74%.

The reason is context. Agents on Hyper have 360-degree context across your accounts, creative, and performance, so they know what to do and why, and they don't fold the moment you question them. We wrote about how that context layer works in Super Intelligence for Marketing. Most of the field recommends; Hyper makes the execution decisions, holds them under pressure, and carries them out.

Built on Supabase

The reason any of this works is the data layer, and a lot of it runs on Supabase. Hyper syncs every connected platform into Postgres, a real data pipeline that turns scattered ad and analytics data into one queryable warehouse, then lets the agents write SQL against it instead of calling APIs live. As I told the Supabase team, "agents are great at writing SQL, it's way more efficient, a literal 100x cost reduction."

It's also where security lives. Every customer's data sits in its own isolated database with Row Level Security enforced at the database layer, so one workspace can never see another's, with no fragile middleware in between. That database-level isolation is something every account gets by default, not just the big ones. On top of it we use Postgres as the source of truth, pgvector for agent memory, Realtime for live signals, Edge Functions, and Storage. Supabase thought the build was worth telling themselves: they wrote a full customer story on how we did it, Hyper builds AI marketing agents on Supabase. It's one of the things we're proudest of. Our own deeper writeup is here, and we shared the short version, from fragmented data to unified insights, on LinkedIn.

Where Hyper fits

If you want an AI that actually runs your Meta and Google ads, and you want the performance to show up where it counts, that's what Hyper is.

Hyper set up two custom agents for us. We used to spend 20+ hours a week on that. Jimmy Smith, Founder, Slice of Pie Marketing

It was running in under 5 minutes. Conor Drake, CMO, Confidence Media Partners