Optimisation agents and the vanity metric trap

Optimisation agents adjust bids, budgets and creative continuously to hit a target, so the target decides everything. Point an agent at vanity metrics such as headline ROAS or click-through and it will deliver them, often at the expense of real profit. Point it at true profit, set reversible limits, and keep a human over the decisions that change strategy.

Why are optimisation agents dangerous with the wrong target?

Because they are relentless. A human media buyer chasing a 400% ROAS might notice that the campaign is driving heavily discounted orders with high return rates. An agent will not, unless it has been told those things matter. It will happily scale the campaign that looks best on the number it was given.

This is a well-worn tension in our listening: brands optimise for ROAS and CPA while ignoring returns, fees and true profit, and agencies get rewarded for metrics that flatter rather than inform. Agents make the problem faster, not different.

What should an optimisation agent optimise for?

Whatever is closest to real business value that you can measure reliably and quickly enough for the agent to learn from. In order of preference:

1. Contribution profit after product cost, discounts, returns and fees.
2. Gross margin on orders, if returns data arrives too slowly.
3. Qualified conversions with a quality filter, for lead generation.
4. Revenue-based ROAS, only with guardrails on discount depth and audience mix.

Our sister app FMedia is built around true-profit ROAS for Meta and Google, and it is worth reading their thinking if your agents currently optimise on platform-reported revenue.

Which optimisation decisions can agents make alone?

Using our checkpoint map, we let agents act alone on reversible, mechanical moves:

• Pausing creative that shows clear fatigue.
• Shifting budget between proven ad sets within an agreed ceiling.
• Adjusting bids within a defined range.

And we keep humans on anything that changes strategy:

• Launching into new audiences.
• Abandoning an audience that converts slowly but may matter long term.
• Changing the offer or deepening discounts.
• Moving budget between channels beyond the ceiling.

When should a human overrule an optimisation agent?

When the agent is right on its metric and wrong on the business. Common examples:

An agent can tell you what is working. Only a person can tell you whether it should.

How does audience research make optimisation smarter?

Optimisation agents learn from what happens after spend. Research agents learn from what audiences say before it. Combining them means the optimisation loop is testing creative that already addresses real tensions, rather than burning budget discovering them. Our research runs on SOMIN, an AI audience-research platform; its paid media intelligence connects audience insight with campaign performance.

For agencies wanting an outside view of measurement, the Analytic Partners case study is a good reference on pairing audience insight with rigorous analytics.

How should results be reported to clients?

Honestly, and in plain language. Each week, our conductors report what agents changed, why, and what the human overruled. We include what went wrong. Clients evaluating agency performance beyond vanity metrics deserve to see the decisions, not just the dashboard. Every change should be traceable, as we argue in our audit trail piece.

What should an optimisation guardrail look like?

Guardrails work best when they are numeric, specific and written into the charter before launch. Examples we commonly use:

• A daily ceiling on how much budget an agent may move without approval.
• A maximum discount depth any variant may promote.
• A minimum share of spend reserved for audiences the brand is building, regardless of short-term efficiency.
• A frequency cap, so no audience sees the same creative past the point of fatigue.
• An alert, not an action, whenever performance changes faster than a set rate, because sudden jumps are often measurement problems.

Guardrails should be revisited monthly. If an agent hits a guardrail repeatedly, either the guardrail is wrong or the strategy is, and a human needs to decide which.

What about creative fatigue?

Fatigue detection is one of the most useful jobs for optimisation agents, because it is tedious for people to watch and easy to miss. Agents spot declining response early and can rotate creative within the approved pool. The human job is making sure the pool is refreshed with genuinely new ideas, not minor variations of the tired ones.

Finally, agree in advance how long a test must run before an agent may declare a winner. Agents will happily call results early on thin data, especially when the target metric is noisy. A minimum learning period, written into the charter, protects you from scaling yesterday's lucky streak into tomorrow's budget, and it gives the human reviewer a fair basis for the call.

If your agents are currently chasing a number you do not quite trust, brief the agents at ask@agentc.marketing and we will help you re-point them.

Frequently asked questions

What should AI optimisation agents target?

The metric closest to real business value that can be measured reliably, ideally contribution profit after costs, discounts and returns, rather than revenue-based ROAS alone.

Can an optimisation agent change budgets without approval?

At AgentC, only for reversible moves within an agreed ceiling, such as shifting spend between proven ad sets. Strategic changes wait for a human.

Why do vanity metrics mislead agents?

Agents pursue whatever target they are given. If that target ignores returns, fees or discount depth, the agent will scale campaigns that look strong but earn little.