How We Test the Best AI for Meta Ads: An Open Method You Can Repeat

Vendor “we tested” pages often compare two periods of Ads Manager ROAS. Use a live control, count delivered orders, and add the fee and the tax.

• Why before-and-after charts fail
• The Advantage+ sales baseline and the budget each cell needs
• A scorecard, plus tax by market

Delivered orders, not a dashboard gap

Pay for the tool that adds delivered, profitable orders on top of a sound Advantage+ sales campaign. Use Meta’s A/B test or Conversion Lift, then the store backend. One tool at a time, attribution locked, full cost included.

Vendor pages often lack a control, and season, price, stock and creative move between periods. Meta advises against switching campaigns off and on. From 12 January 2026 the Ads Insights API stopped returning 7-day and 28-day view windows, per Meta’s developer note as covered by PPC Land. From 3 March, trade coverage moved likes, saves, comments and shares to a one-day engage-through window and said billing was unchanged. That billing point was not confirmed on a Meta page. A window crossing either date mixes two rulebooks. Haus’s 640 incrementality tests found higher incremental ROAS from manual campaigns for 58% of brands that tested both against Advantage+, including some where Advantage+ looked better in Ads Manager.

The baseline, then the design

Test against a new Advantage+ sales campaign. It is included with the spend. An old manual campaign with tired ads is an easy win and a weak conclusion. Judge each job on its own: agents that edit budgets, bids and ads; rule engines such as Birch, formerly Revealbot; image, video and copy generators; creative analytics; and multivariate testers.

Use the unified sales campaign that replaced Advantage+ shopping. The ASC and AAC APIs are deprecated in Marketing API v25.0. Leave Advantage+ audience, placements and campaign budget on. Industry reports say new sales campaigns start that way. Load at least 20 varied ads. Write-ups say Meta recommended that in November 2024 and also warns that too many ads at once means fewer exit learning. Neither point was confirmed from Meta. Every cell needs the same pixel and Conversions API.

Write the rule first. Example: cost per delivered new-customer order falls 10% or more versus baseline, after the fee. Creative, analytics and rules use an A/B test with one change and no audience overlap. Agents that shift budget across campaigns use Conversion Lift or a geo holdout. Meta treats A/B as a tactical comparison and Lift as a read on conversions the ads caused. Lock 7-day click, choose engage-through before launch, and keep 1-day view identical. Run full weeks. Open outside Diwali, Ramadan and Black Friday–Cyber Monday unless the festive auction is the question.

Aim for 100 purchases per cell, at least 14 days and usually 28. A 20–30% test share is cited in industry reports as Meta’s figure and was not confirmed from Meta.

Monthly spend	Test budget	Purchases per cell per week at ₹1,500 CPA	What fits
₹10 lakh	₹2–3 lakh	About 17–25	One 28-day A/B test, about 70–100 purchases per cell
₹25 lakh	₹5–7.5 lakh	About 40–60	An A/B test in 18–28 days
₹1 crore	₹20–30 lakh	About 165–250	A/B plus Conversion Lift or a geo holdout

For two weeks, require approval on every automatic change. Cap daily spend near 120% of plan, keep a one-click rollback, and export the change log weekly.

Where a 15% lift becomes a loss

Record Ads Manager purchases with 1-day view on and off, UTM-tagged store orders, and delivered orders after cancellations, returns and RTO, split new versus returning. One agency analysis lists existing customers, view-through credit, weak signals and thin creative as four wastes inside Advantage+. Count COD only when delivered, using that cell’s RTO rate.

A ₹40 lakh brand, average order ₹2,000 and about 40% COD, spends ₹6,00,000 per cell for 28 days. In-house ads: 400 purchases. The tool, same structure: 460, a 15% gap. After RTO and cancellations: 320 and 340, about 6%, with new customers at 220 and 230. The ₹40,000 fee makes delivered cost ₹1,875 versus ₹1,882, and each extra order costs ₹2,000. At a ₹1,900 break-even, stop.

Weight the incremental delivered-order cost at 50%, fees plus unrecovered tax, FX and cards at 20%, logged hours at 15%, and approval, caps, rollback and the log at 15%. Ship at 10% or better after fees and tax, at a preset confidence level such as 90%. Kill within 5% either way after 28 days, or if worse. Extend once, by 14 days, for a 5–10% gain that is not yet confident. Record losses. A single lift number does not travel across accounts.

Market	Invoice	What to do
India	Meta Ireland shows no GST	Self-assess 18% IGST, pay via GSTR-3B Table 3.1(d), claim the credit
UK and EU	No VAT with a valid number on the account. Otherwise 20% in the UK, or the local rate	Reverse charge
Australia	Sources conflict on 10% GST once an ABN is entered	Check Meta’s billing page
United States	Third parties cite 0–10% state sales tax, not confirmed with Meta	Budget above the set spend
Gulf and South-East Asia	Not verified here	Read billing settings

A Meta Ireland invoice shows no GST, so Indian advertisers self-assess and can miss it. Facebook India billing shows GST and TDS instead. Add dollar FX and card fees. March 2026 coverage sets the website default at 7-day click, 1-day engage-through and 1-day view, and a video engaged view at 5 seconds, down from 10. The older 12% lower cost across 15 tests used platform attribution. Week 0 locks baseline, UTMs, hypothesis and caps. Weeks 1–4 change nothing except a broken ad. Week 5 decides after RTO, fees and tax. At ₹10–20 lakh, use 28-day A/B tests. Lift often lacks volume.