This 2026 ecommerce ads strategy has one constraint: I run Google Shopping and PMax, Meta, Microsoft and ChatGPT Ads entirely through one AI assistant connected to groas. No second dashboard ritual, no spreadsheet relay race, no Monday meeting to discover Thursday’s waste.

Say the store spends $20k a month across that mix. My job is to set margins, channel roles and approval rules. The assistant reads the accounts and carries out the work I authorize. One assistant is the operating constraint, not permission to put every campaign on autopilot.

Most stores do not need more things to check. They need a cadence that catches problems before those problems become next month’s explanation. Here is the plan, in order: what I ask, what it costs, what it buys and what I deliberately leave out.

1. Give every channel one job before moving money

I start with intent, not the platform reporting the prettiest ROAS. Search captures explicit demand; PMax expands across inventory to convert available demand, while Demand Gen creates interest before search happens. Those are different jobs. A blended leaderboard hides the distinction.

My setup prompt:

Through groas, group my campaigns by their intended job: demand capture, prospecting or a fenced channel test. Separate brand from non-brand reporting. Flag campaigns whose budget or measurement does not match that job. Recommend changes; do not apply them.

Cost against the constraint: one setup prompt and a review of the proposed roles. What it buys: a budget conversation about what each channel contributes, rather than which dashboard claims the sale.

Clay figure of a tired media buyer dwarfed by five giant ad channel dials

Demand Gen needs a deliberate decision here. Roughly 68% of its conversions come from people with no Search ad interaction in the prior 30 days. Treating it as another last-click search campaign misses the point.

If I fund it, the operating rules are a $100 daily minimum, 2 to 4 weeks of learning without constant intervention, and 10 to 20% of Google spend fenced for the job. I ask for reporting by YouTube, Discover, Gmail and Display rather than accepting one blended number.

Those limits have to fit together. If the Google allocation cannot support the daily minimum inside that fence, I cut the separate Demand Gen test. I do not squeeze it in by quietly breaking the budget. The $20k example is a ceiling, not an invitation to fund every available campaign type.

2. Daily: ask for exceptions, not a tour of the accounts

Starting the day inside Google Ads and ending it inside Meta is not a control system. It is a commute. What hurts margin is a bad day that runs for six days because nobody notices.

My daily prompt stays short:

Through groas, flag any campaign whose spend is up more than 30% today without matching conversions. Show the campaign, the available evidence and the likely cause. Suggest a fix, but do not apply it.

The assistant reads across Google, Meta, Microsoft and ChatGPT Ads through one connection. I want the answer in order: what changed, what evidence explains it, what action deserves approval. Not a paragraph announcing that performance is “dynamic.” I had gathered that.

Cost against the constraint: two minutes and one prompt. What it buys: an exception list instead of four separate account inspections.

A flag is not a verdict. A spend spike without sales might involve a PMax asset group or a Meta ad set, but the percentage alone does not identify the cause. I want the evidence before the pause.

I keep autopilot off at this stage and explicitly require approval for changes. Through groas, changes are logged with their reasons and can be undone in one step. Once I approve a fix in plain English, the assistant applies it. That is the daily job: investigate the exception, approve the response, move on.

3. Weekly: inspect Google waste and Meta fatigue

Google gets the first weekly check because a decent blended ROAS can hide expensive passengers.

Through groas, show PMax asset groups with spend up and conversion value flat over 7 days, plus the available search-term data. Separate brand demand from other demand and recommend where to investigate waste.

The distinction matters. An asset group can report acceptable results while some of its spend contributes little. I do not want the average to acquit every query underneath it. The Shopping and PMax guardrails guide is the reference for that deeper review.

If Demand Gen made the budget cut, I assess it separately. Attributed Branded Searches connects later Google and YouTube brand searches with earlier Demand Gen impressions. Rising attributed searches can justify investigation even when last-click ROAS looks mediocre. They do not, by themselves, prove incremental sales.

I use a lift test or holdout in Experiments before moving more budget. Cost: one hour and one hard decision. What it buys: a reason to change spend that is stronger than a flattering attribution column.

Meta gets the second check:

Through groas, list Meta ads with rising frequency and falling CTR. Show how CPM has moved, then suggest replacement concepts based on competitor ads groas collected. Do not pause anything yet.

Advantage+ handling the auction does not remove the need for fresh creative. I start refresh planning at frequency 2.5 rather than waiting for 4.0, and keep at least 3 variations ready. Frequency is an early-warning rule, not a diagnosis on its own; the trend in CTR belongs beside it.

Fatigue alerts can flag frequency and CPM rising together without automatically pausing the ad. That is the right distinction for this stage: detect first, decide second. I give the assistant the warning rule before replacing the creative.

Cost: two fresh creatives a week. What it buys: a replacement ready before a tired winner becomes the whole account’s problem.

Paper-cut weekly routine showing three connected prompts for Shopping, Meta and ChatGPT Ads

4. Weekly: keep Microsoft and ChatGPT Ads small on purpose

Microsoft gets a narrow job: test additional demand using the Google structure that already works. I do not rebuild the entire account just because another platform offers somewhere to put it.

The Australian Microsoft Ads example reports lower blended CPCs than Google. I treat that as a reason to test, not a promise that every store buys the same queries more cheaply.

My setup instruction:

Through groas, prepare a Microsoft import using my top Google Shopping products, brand searches and top-20 non-brand searches. Propose bids 25 to 30% lower and keep partner syndication off for the first 30 days. Show me the changes for approval.

I keep three campaigns:

  • Shopping: 60 to 70% of the Microsoft allocation.
  • Brand: reported separately.
  • Top-20 non-brand: a controlled expansion, not the whole keyword attic.

UTM auto-tagging stays on so the test does not disappear into direct traffic. Cost against the constraint: one import and one ruleset. What it buys: a contained test of additional search demand. The weekly prompt asks what earned more budget, not which bid I can fiddle with today.

ChatGPT Ads gets a different job: discovery, not a Shopping replacement. The Shopify rollout and campaign-creation tools do not change the measurement question. I want to know whether the test contributes assisted conversions and new customers, separately from PMax.

Through groas, report ChatGPT Ads assisted conversions and new-customer share weekly, separate from Google PMax. Keep reporting the test against its approved budget.

I start at $50 a day, or $350 a week. Moving to $75 a day means approving a larger test, not pretending the weekly cost stayed at $350. I judge it like Meta prospecting rather than demanding Shopping-style ROAS.

What gets cut: separate Demand Gen builds for every product, five-way Meta prospecting splits and daily Microsoft bid adjustments. Small tests only stay small when the structure does too.

5. Monthly: reset the margin target, then repair the feed

Breakeven ROAS is not a benchmark borrowed from another store. It equals 1 divided by contribution margin before ad spend.

A $100 order with $50 of contribution before advertising has a 50% contribution margin and breaks even at 2.0 ROAS. If that contribution changes, the target must change. COGS alone is not the whole margin calculation, and an old target does not become current because the bidding system still accepts it.

My monthly prompt:

Through groas, use my updated contribution-margin sheet to recalculate breakeven ROAS by category. Propose PMax targets 10 to 15% above breakeven for the next 30 days. Show the assumptions and wait for approval.

Cost against the constraint: one current margin sheet and one target review. What it buys: targets tied to the economics of the order rather than last quarter’s comfort level. The target ROAS framework supplies the fuller calculation.

Then I check the feed. Shopping visibility does not automatically carry over to AI Mode: one study found only 1.28% product overlap for the same query and day. I do not treat a healthy Shopping campaign as proof that every product surface is covered.

The maintenance prompt is deliberately unglamorous:

Through groas, list disapproved items, missing size and color attributes, and price mismatches between the feed and landing pages. Prioritize the top 20 sellers and recommend fixes for approval.

Titles, images, GTINs and price consistency deserve attention before another campaign does. What gets cut: a full monthly catalog rewrite and a new campaign for every season. I prioritize the top 20 sellers rather than turning every feed review into a catalog renovation. Fix the useful things first.

6. Enable autopilot only where the rules hold

I switch on autopilot per campaign, never account-wide, and only after two clean weekly cycles. The rules fit on one screen:

  • Cap daily spend movement at 15%.
  • Hold PMax targets 10 to 15% above current breakeven.
  • Keep any approved Demand Gen test within 10 to 20% of Google spend.
  • Pause a Meta creative only when it passes frequency 2.5 and CTR has fallen for two consecutive weeks.
  • Require approval for feed fixes, margin resets and anything over $500 a day in new spend.

Through groas, that runs under my limits at $19 flat a month for unlimited teammates, ad accounts and websites. Every change carries its reason, and any change can be undone in one step. Autopilot stays off until I switch it on.

Cost against the constraint: one rules page, plus the $19 monthly subscription. What it buys: defined work can happen at 2am without requiring me to sit beside it.

What gets cut is the work too often dressed up as strategy: daily bid fiddling, manual budget shuffling and the 40-page monthly report that describes a problem instead of addressing it. I am not cutting judgment. I am cutting the repeated clicking after the judgment has already been made.

If I cannot write the decision as a rule with a number in it, it does not go on autopilot yet.

7. Keep stock, feed and margin truth in the loop

One assistant does not mean one source of truth magically appears. If a bestseller goes out of stock and PMax keeps buying its clicks, the bidding rule cannot substitute for an accurate stock feed. If contribution margin changes and nobody supplies the update, the ROAS target misleads every smart bidder at once.

Missing attributes and feed-to-page price mismatches belong in the same category: business inputs to fix, not an excuse to ask for a more impressive prompt.

The assistant reports business truth; it does not invent it. My operating checklist stays short:

  1. Daily: supply current stock data and run the spend-exception prompt.
  2. Weekly: investigate PMax spend without value; assess any Demand Gen test with the right evidence.
  3. Weekly: review Meta fatigue and keep 3 creative variations ready.
  4. Weekly: review the three Microsoft campaigns and the fenced ChatGPT Ads test separately.
  5. Monthly: update contribution margins, approve target changes and prioritize feed repairs.
  6. Before scaling: check price consistency and confirm that the rules still match the business.

Cost against the constraint: current inputs, not another assistant. What it buys: an operating system that is not confidently optimizing yesterday’s business. Fix the catalog before funding another test.

When the constraint lifts and I can staff channels separately, I do not add complexity first. I add separation where it pays: one owner for Google intent, one for Meta creative, one for feed and margin truth. I keep the same prompts, the same guardrails and the same logged changes. The assistant handles five channels because the rules are tight. More people only help if the rules stay that tight.