How an AI Media Buyer Actually Works, Step by Step

The analysis was never the hard part. Any decent tool can read an account and list wasted spend. The hard part is everything between the analysis and the account. Who scores the change. Who approves it. And what happens at 2am when nobody is watching.

Scaley is built by the team behind ZenoX Media, a Google Ads agency. We build it, so weigh that. The playbooks came first. The software is the agency's own checklist with an approval gate on the end.

How does an AI media buyer work?

An AI media buyer reads account data, scores candidate changes, and prepares each proposal with its reason attached. With Scaley, every write waits for a human yes; the team behind ZenoX Media built it around playbooks from work across 200+ ecommerce stores.

1. A read-only audit that reads your history and touches nothing.
2. Scoring: every SKU tagged by margin and velocity, every candidate change judged by profit.
3. The approval queue: one proposal at a time, reason attached, no write without your yes.
4. The log: every approved change on the record, every rejected one having never touched the account.

The gate in step 3 is what makes it a media buyer instead of a recommendation engine. We laid the whole category out on the AI media buyer page, and defined the term itself in AI Media Buyer for Google Ads.

Step 1: the read-only first audit

You connect Google Ads. If you run Shopping, you connect Merchant Center too. Scaley then reads 90 days of history, the audit window, without touching a single setting.

We chose read-only on purpose. It is not a demo trick to make the product look safe. The audit's only job is to earn your trust before it asks for anything.

• Search terms: which queries are eating budget without buying.
• Placements: Performance Max and Display placements with impressions and no conversions.
• Merchant Center diagnostics: disapprovals and feed problems that quietly kill traffic.
• Your catalogue: what each product costs, what it earns, and how fast it sells.
• Budget pacing: which campaigns will run out of month before they run out of budget.

The output is a list of leaks, with the evidence next to each one. One of our published case studies covers a DTC apparel brand spending €180K/mo. It had €4K a month leaking into low-intent search terms and weak display placements. The audit found that before a single change was proposed.

For scale: on our own agency's €20K/mo accounts, we typically find €2-4K leaking. Not every account leaks like that. The apparel brand did, because someone was only adding negatives once a month.

You can disconnect at any time. The audit changed nothing, so there is nothing to clean up on the way out. What you keep is the baseline: a leak list with numbers you can re-check in a month, whether you stay or go.

Step 2: how a change gets scored before it is proposed

After the audit, Scaley watches the account 24/7. Every idea for a change runs through the same scoring pass first, and the pass runs on two axes: margin and velocity.

The Labelizer tags every SKU by what it earns and how fast it sells. It prepares those tags for your Merchant Center custom labels during scoring. The sync itself waits: a tag reaches Merchant Center only after your individual yes on that sync. From those two axes, three tiers fall out:

• Hero: high margin, selling fast. These get the bids and the budget.
• Workhorse: decent margin, steady sales. These pay the bills and get left alone.
• Bleed: low margin, not selling, or both. These get capped or paused.

Velocity is 30-day sales speed, the same window used in our published case studies. After you approve the sync, the tags plug into campaign structure, Target ROAS or budget decisions. Your bids stop chasing revenue and start chasing profit.

In the published jewelry case study, a brand had all 2,400 SKUs in one campaign with a single Target ROAS across the board. The Labelizer tagged the lot. Scaley rebuilt the feed into tiers: Target ROAS of 4.0 on heroes, 3.0 on workhorses and 2.2 on the tail. Blended ROAS rose 38% in seven days. The scaling rules and exclusions engine stayed off. Tiering alone did that.

That is why scoring comes before proposing. A tool that only sees revenue pushes your bestseller harder. A tool that sees margin knows the bestseller is your thinnest product, and that the boring mid-list item is where the profit lives.

Margin and velocity score the products. The changes themselves get scored too, and each kind of proposal carries its own evidence:

• A search-term exclusion shows what the query spent and what it converted. Money out, nothing back, the pattern repeating: both numbers sit on the card.
• A placement exclusion shows its impressions against its conversions. Pile-up with nothing behind it is the signature.
• A budget move never rides ROAS alone. The pass reads ROAS, margin tier, velocity, the season, the day of the week and how the monthly budget is pacing. The same ROAS on a hero SKU and on a bleed SKU are two different facts.

Scoring does not change your account by itself. Every candidate on that list reaches you as one proposal with its evidence attached, and nothing writes until you say yes to that exact one.

Step 3: the approval queue, one proposal at a time

Scoring produces candidates. The queue is where a candidate becomes a proposal.

One proposal at a time. Nothing arrives in a batch. You read one change, you decide on one change, you move to the next.

Scaley attaches the reason to the proposal, with the numbers behind it: what the query spent, what it converted, what the change does. Not a one-line "optimization available" badge.

Scaley writes nothing until you say yes. Not yes to a category of change. Yes to this specific one.

On screen, a proposal is one card: the change in plain words, the numbers behind it, and the reason pulled from the playbook. You approve it or reject it. That is the whole interface between the analysis and your account.

Some things never reach the queue at all. From our own automation doctrine, written after watching real accounts break, three proposals are banned outright:

• Big budget moves off a short trend, which is how a month of margin disappears over a bank holiday.
• Structure changes. Nothing renames campaigns or rebuilds ad groups while you sleep.
• Brand exclusions. Automated negative mining will eventually flag a term that looks wasteful and is your own brand plus a modifier. Cheap to approve, expensive to discover later.

The queue does cost time. In the published home decor case study, a group running four brands on about €420K/mo combined was about to hire a third junior buyer. Their lead reviews the queue in 45 minutes a day instead. The group saves 15+ hours a week, and the third hire never happened.

Scaley finds overnight. You decide in the morning.

Step 4: what logged and reversible means in practice

"Logged and reversible" shows up in every automation pitch, ours included.

Logged: every proposal keeps its reason and its numbers, and every yes you give is on the record. If something changed on your account, the log says what changed and why. Our own test for automation is blunt. If a log can't say what changed and why, that's not automation. That's a black box with extra steps.

Reversible, three ways:

1. A rejected proposal never touched the account. There is nothing to undo, because nothing happened.
2. The changes that get proposed are small and reversible: a negative keyword, a budget move, a label sync. Campaign structure is not on that list, which is why structure changes never arrive in the queue overnight.
3. An approved change you regret is on the record with the day it landed. You start the reversal, not Scaley. The log entry says what changed and why, so you know what to undo and when it went in. The record of what happened stays.

These are boring changes, proposed one at a time, with a receipt attached. That is how the mechanism can run at 2am without a person watching: the part that runs unattended is reading and scoring, never writing.

Where the playbooks come from

The scoring pass does not run on general knowledge. When Scaley prepares a change, it reads the playbook library ZenoX Media built while running ads for 200+ ecom stores. That library holds 262,792 Slack messages across 636 channels. It also has 8,183 voice clips, 109 hours total, of buyers talking through real decisions. The written part is 95 SOPs and 100+ playbooks, 193 files in total.

That library belongs to ZenoX Media, the Google Ads agency behind Scaley. Scaley reads it when it prepares a change. No model was trained on it.

That library is where the reason attached to each proposal comes from. It is the difference between "add negative keyword, predicted savings" and "we have seen this query pattern before." The second reason tells you what that pattern did last time.

The agency behind that library is rated 4.8/5 from 38 reviews on Trustpilot. Agency reviews, not Scaley product reviews. The product is young, and pretending the agency's reviews are its own would be fake.

Where the chat tier fits

Three tiers, one mechanism. The gate does not move. What moves is how much of the mechanism you get.

Tier 1 is chat only. You ask, it answers with the playbook behind the answer. No account connection, no writing.

Tier 2 adds the read-only connection per account. Now it reads your Google Ads and Merchant Center data and suggests moves in conversation. It still writes nothing. Reading and suggesting is the whole job.

Scaley Suite is the tier that writes: $199/mo plus 4% of spend. It has a 7-day free trial, and billing does not start until day 8. Every write still waits for your individual yes, same as everything above. The base fee and rate apply together at every spend level.

Each tier is honest about what it cannot do. Tier 1 cannot see your account, so its answers stay general until you connect one. Tier 2 can read your account and never write it. The Suite can write, and only when you say yes to the exact change. There is no tier where the software decides for you. The ladder stops one step short of that on purpose.

If you want to try the mechanism before the full Suite, the GPT for Google Ads lander is the low-risk door. Chat first, connect read-only when it has earned it, hand over the write gate when you are ready. All three prices sit on the pricing page.

Scaley is built for stores running their own Google Ads, and for dropshippers testing products fast. If that is you, start the 7-day Suite trial and connect the account. The first audit is read-only and changes nothing. The worst possible morning is the one where you learn exactly where the money leaks.