A Builder's Honest Review of Cogny
We build Cogny, so we can't say 'the best AI marketing agent on the market' — that's for someone else to earn. What we can do is show you the ledger: 167 of our own measured marketing changes, 70 worked, 78 partial, 19 didn't work, misses included. This is the honest part no tutorial can copy.
Cogny is an AI marketing agent. It watches your numbers, proposes next moves, and — once you approve them — ships changes into your ads, site, analytics or CRM. Then it opens a ticket and measures the result in your own data a few days later, closing it as worked, partial, didn't work or inconclusive.
We build it. So we can't seriously write "the best AI marketing agent on the market" — that line has to be earned by someone we don't own, or it's just marketing noise. What we can do is the thing almost no vendor will do: show you the misses.
This review is the honest part. It's the credibility anchor that the rest of the build-vs-buy-marketing-harness series leans on — if you're deciding whether to build your own marketing stack or buy one that measures itself, start with the record.
The verdict ledger: our own misses, in the open
We market Cogny with Cogny. Every change one of our own agents makes to our site, ads, or analytics becomes a measured ticket on the verdict ledger. No cherry-picking, no "as measured in the best week ever" — each change is read in our connected data and closed with a verdict.
As of late 2026, that ledger holds 167 measured changes:
| Verdict | Count |
|---|---|
| Worked | 70 |
| Partial | 78 |
| Didn't work | 19 |
| Inconclusive | 0 |
That's a 42% worked rate, a 78 partial rate that we don't dress up, and 19 changes that flat-out didn't work. No 100% success rate, no "every change shipped wins." Some changes we made moved nothing and bought no result, and you can read them all on the ledger by kind of change: ads, site and SEO, tracking and measurement, email and community.
Why publish this? Because the whole product promise is a bet, and a bet you can't verify is just a claim. If we tell you every change you make gets a verdict, we should be willing to stand under the same microscope. The ledger is how.
What really works
Honesty doesn't mean false modesty. Here's what the 70 "worked" tickets and the day-to-day product experience actually support.
The loop is the product, and the loop is real. The difference between Cogny and "an AI that writes you a report" is that a report is the last step for most tools, and it's just the proposal for us. You ask about a metric, the agent proposes one concrete change, you approve it, it ships, and it gets measured. The step people habitually skip — did it work? — is automated and free on every plan, Free included.
Measurement is boring and that's the point. We compare equal before-and-after periods in your own connection — 3 days later for ads, 7 days later for content and SEO. It's not fancy, but it's repeatable, and it's the thing that separates a marketing change from a marketing guess.
Running from your own AI is what customers actually use. Claude, ChatGPT, Cursor, Codex — your own agent drives Cogny over a standard connector, and everything it writes is logged with the tool, the time and the result. Most of the "worked" verdicts came through that path, not the app.
What's partial
The 78 partial verdicts are where the honest nuance lives. "Partial" usually means one of three things:
- It worked on the metric we aimed at, but not the one you'd actually point to first. A change lifted one conversion and nudged cost per conversion sideways. That's real progress, not a headline.
- It worked, then faded. What wins for three days can drift back after the novelty or the budget re-balancing. Partial is how we say "yes, but watch it."
- It moved the campaign but not the account. Spend that wins in one campaign frequently just moved from another. The ledger counts that as partial, because the business result you care about is the whole account, and that's a harder bar.
Partial isn't failure, and it isn't a win you can brag about. It's the honest middle ground that a binary "worked / didn't" would hide.
What didn't work
Nineteen changes didn't work, and we're not going to dress them up. The reasons are the normal ones, and naming them helps you avoid them:
- A change that was an okay idea aimed at the wrong lever. Good copy on a page nobody's traffic lands on still gets a "didn't work" for the page.
- A well-measured change that shifted spend with no net gain. It shipped, it deployed, and the account total simply didn't move. That's a real miss.
- Changes that needed a bigger commit than one ticket. Sometimes one change isn't enough and the honest verdict is that the single move — on its own — didn't do it.
The key word is ledger. These aren't scrubbed after the fact. They're permanent rows on a public page, the same way your tickets are permanent in your own account.
What this means if you're reviewing Cogny
Searching "cogny review" or "ai marketing agent review" usually surfaces either marketing pages or vendors reviewing themselves. A review written by the people who built the thing has a built-in blind spot, so read accordingly. What makes ours worth your time is that the claims being made are falsifiable:
- You can run the same loop in a free workspace and get your own verdict in a few days.
- You can read our ledger and see our misses, not a highlight reel.
- You can hold any statement against a specific, sourced number on a public page.
That's a review framework no testimonial page can fake, because it doesn't ask you to trust us — it asks you to check.
The case for the measured agent, not a stack you build
If you're weighing buying Cogny against building your own concatenation of a data pipeline, a reporting tool and prompts, this is the honest fork in the road. Building your own gives you control and zero vendor dependency; the price is that you own the measurement, the scheduling and the maintenance. Nothing measures itself unless you make it, and very few solo founders or small teams actually build that part — it's the first thing that gets cut.
A measured agent inverts it: the measurement is the product. Whether that's worth $9 a month or $499 is a real decision, and it's the one the build-vs-buy harness is written to help you make. But it should be decided on whether you want a tool that tells you what it did — misses included — not on a review that promises otherwise.
Read our record, then check it yourself
The most honest thing we can say about Cogny is also the simplest: every change gets a verdict, misses included. The verdict ledger is the whole record — 167 changes, 70 worked, 78 partial, 19 didn't work. We don't claim a 100% success rate, and you shouldn't expect one from any tool that actually tests things.
The only independent check that matters is the one you can do in a free workspace, no card required. Connect one source, ship one change, and read your own verdict three or seven days later. If it didn't work, you'll see that too — that's the deal.
FAQ
Is Cogny actually any good?
That's ours to earn, not to claim. The closest thing to an answer we can give is our public ledger: 167 of our own measured marketing changes, 70 worked, 78 partial and 19 didn't work. The product promise isn't "everything wins" — it's that every change gets measured and the result is closed with a verdict. A free workspace will produce your own verdict, not a marketing claim.
Does every change Cogny makes work?
No. Nineteen of our own 167 measured changes didn't work, and they're on the public verdict ledger rather than scrubbed. We never claim a 100% success rate, and any tool that does is lying or not measuring. What's guaranteed is the verdict, not the outcome.
What does "partial" mean in a verdict?
A partial verdict is an honest middle ground: the change moved the metric it aimed at, worked then faded, or helped one campaign at the expense of the account total. It's real progress — not a headline and not a miss.
Can I verify this without paying?
Yes. Free is $0 a month, no card, with 15 MCP calls a month and 50 one-time chat credits, and every plan gets measured tickets. Connect one source, ship one change, and read the verdict a few days later — misses included.