TerradsTerrads
← All posts

Meta publishes one creative-volume number, and it isn't any of the ones you've been told

At least 20 diversified ads in Advantage+ shopping campaigns, and 20-30% of budget on testing. That is Meta's recommendation. The 8-15 per ad set and the Entity ID threshold are nobody's.

Chris Jankulovski8 min read

Meta's published recommendation is at least 20 diversified ads in Advantage+ shopping campaigns, with 20–30% of overall ad budget on testing new ads, and new ads launched weekly by the advertisers doing it best. That is one number, scoped to one campaign type, and the word diversified is doing most of the work. The figures circulating instead — eight to fifteen per ad set, fifty a month, 50 to 100 conversions per Entity ID — appear in no Meta document I could find.

Ask how many ads you should be running and you will get an answer within seconds. The trouble is that you will get a different answer from every source, each stated as a fact about how Meta's system works, and almost none of them carrying a link.

Here is a sample from the first page of results, all of them other people's claims and none of them mine: that accounts running three creatives per ad set under-deliver against accounts running eight to fifteen; that brands testing twenty-plus new ads a month see 65% higher ROAS; that you want three to five genuinely new concepts per month for every $50,000 of monthly spend; that budget should split 70% proven, 20% testing, 10% experimental; and that Andromeda needs 50 to 100 conversions per Entity ID before it exits the learning phase.

Five guides, five incompatible numbers, no citations. So I went to find what Meta has published, and the answer turns out to be both less and more than the internet suggests.

The number Meta actually publishes

It is on a Best Practices page from November 2024, Three steps to optimize your performance with creative diversification, and it is unambiguous:

As a best practice, we recommend maintaining at least 20 diversified ads in Advantage+ shopping campaigns and dedicating 20-30% of your overall ad budget towards testing new ads.

The same page addresses cadence, and is careful about it in a way the republished versions are not:

The best creative diversification leaders routinely launch new ads on a weekly basis but if you're just getting started your schedule may look different based on your resources.

That is the recommendation, in full. Three things about it are worth more than the number itself.

It is scoped. Twenty ads in Advantage+ shopping campaigns, not twenty per ad set, not twenty everywhere. A guide that reports it as a general threshold has dropped the only clause that tells you when it applies.

The adjective is load-bearing. Twenty diversified ads. Meta's own definition of the term, from a companion page, is "a wide range of ad creatives with different themes, messages, and visuals to cater to diverse audience segments." Twenty crops of one photograph satisfy the count and none of the definition.

The cadence is explicitly conditional. Weekly is what the strongest practitioners do, and Meta says outright that yours may differ based on your resources. It is an observation about leaders, not a threshold you are failing.

So a budget split is published — 20–30% on testing — and it is not 70/20/10. A volume figure is published, and it is not 8–15 per ad set. The industry did not invent numbers out of nothing so much as replace Meta's with its own and keep the authority.

The claim that is genuinely fabricated

One of the five is different in kind. "Andromeda needs 50 to 100 conversions per Entity ID to exit the learning phase" describes a specific mechanism inside a specific named system, and that system has exactly one technical document behind it: Meta Andromeda: Supercharging Advantage+ automation with the next-gen personalized ads retrieval engine, published December 2024.

That post is about the retrieval stage:

Retrieval is the first step in our multi-stage ads recommendation system. This stage is tasked with selecting ads from tens of millions of ad candidates into a few thousand relevant ad candidates.

Retrieval narrows an enormous pool to a shortlist before ranking and the auction get involved. I downloaded the page and searched the source rather than reading a summary of it. The phrase "learning phase" does not appear. Neither does "per ad set". The word "entity" appears exactly once, inside the page's own structured-data markup, as the property name mainEntityOfPage.

The learning phase is a real, separately documented part of Meta's delivery system. Andromeda is a retrieval engine. Welding the two together produces a confident sentence about a mechanism that has never been described anywhere, and the precision of "50 to 100" is what makes it persuasive.

What Meta's documents do say about volume

It is worth being accurate about this too, because "Meta never mentions volume" would be as wrong as the claims this post is correcting. Meta talks about volume constantly — it just never separates it from diversity.

The advertiser-facing page describes generative tools making it easier to "upload more volume and diversity of creative," and the engineering post uses the word repeatedly while explaining why retrieval had to be rebuilt: the number of ad creatives in Meta's systems is expected to grow significantly, and Andromeda exists to handle that growth. More than a million advertisers made more than 15 million ads in a month with those tools.

None of that is advice about your account. It is Meta explaining why it needed a bigger filter. The pairing is always volume and diversity; what travelled was the volume half, because a count is actionable and "different themes, messages and visuals" is a judgement call.

The figures Meta does attribute

For completeness, since the argument here is about sourcing rather than scepticism, here is what the three pages claim, all attributed:

  • 8% improvement in ads quality from Andromeda, and a 6% recall improvement to the retrieval system, on selected segments.
  • 22% increase in ROAS for advertisers who turned on Advantage+ creative's AI-driven targeting features, and an estimated 7% increase in conversions for businesses using image generation.
  • 9% lower cost per action on average with the streamlined setup for Advantage+ sales campaigns.
  • 11% higher click-through rate and 7.6% higher conversion rate on average for campaigns using image generation within Advantage+ creative.
  • 19% lower CPA on average for businesses using partnership ads.
  • A 7% increase in link clicks in one named case study using background image generation.

These are averages across Meta's advertiser base, published by the company that sells the features, and they measure adoption rather than creative strategy. Weigh them accordingly. But they are attributed, dated and findable, which is more than the numbers this post opened with.

What to do with this

  1. Use Meta's figure where Meta's figure applies. At least 20 diversified ads in an Advantage+ shopping campaign, 20–30% of budget on testing. Outside that campaign type, Meta has published nothing and neither has anyone else.
  2. Spend the effort on the adjective. Describe each creative in one sentence without naming its format. If two sentences match, you have one creative twice, and your count is measuring production rather than diversification.
  3. Treat cadence as a resourcing question. Meta says so itself. Weekly is what the leaders manage, not a line you drop below.
  4. Be suspicious of any number attributed to Andromeda. It has one technical document. If a claim about it is not in that document, ask where it came from.

The broader pattern is the one that also runs through ad specs, safe zones and creative fatigue: a number repeated in twenty articles is evidence that twenty people copied each other. Meta publishes a great deal about its own systems, and much of what circulates as "what Meta says" was never in any of it — including, this time, a real recommendation that got overwritten by louder ones.

Sources

All three pages were fetched and read in full on 19 September 2026. The engineering post was searched as raw source text rather than in summary, which is how the absence of "learning phase", "per ad set" and any entity concept was established. Meta revises and removes pages without notice — if a link has moved, re-check the claim rather than trusting this page.

Terrads generates ads that differ in theme, message and visual rather than in crop, with each placement composed for its own frame.