Meta Ads

Creative is the targeting: Meta’s Andromeda and e-commerce video ads

Since Meta rebuilt the first stage of its ad delivery around a system called Andromeda, the ad itself does most of the work interest targeting used to do. That changes what you should produce: not more versions of one ad, but more ads that are truly different. Here is what Meta says, what media buyers observe, and a worked example that turns one product into six concepts.

By the Dokiri Studio team · November 6, 2026 · 8 min read

In this article

What Andromeda is, in Meta’s words

Andromeda is the retrieval engine of Meta’s ad system: the first step, which narrows tens of millions of eligible ads down to a few thousand candidates for each person before ranking models choose what they see. Meta described it on its engineering blog in December 2024. Its business news post of March 2025 adds the part that matters to advertisers: as businesses upload more creative to support a diversification strategy, Andromeda helps pick the right creative for each person.

Meta’s own list of performance practices, which it calls Performance 5, puts creative diversification next to account simplification: make a variety of ads and let the system match each message to the people most likely to respond.

Why broad targeting and varied creative replaced interest stacking

With Advantage+ audience, delivery can look beyond the audience you describe. Meta’s developer documentation lists the inputs that stay fixed: location, minimum age, language and custom audience exclusions. The rest guides delivery but can be expanded.

So the strongest signal left in your hands is the ad. Bluesense Digital, an e-commerce agency whose presenter publishes long Meta Ads breakdowns on YouTube, puts it this way: the system reads what the ad says and shows, from the spoken words and on-screen text to the people and the copy, and finds the viewers it fits (video). In his reading, a line like “for nurses on 12-hour shifts” steers delivery toward nurses better than an interest box ever did.

  • Stacked interest ad sets split your data. An ad set usually leaves the learning phase after about 50 results in the week after its last significant edit, per Meta’s Help Center. Five small ad sets get there later than one.
  • The person on screen targets too. In his experience, a woman of 60 and a man of 20 holding the same product don’t reach the same buyers.

Why ten near-identical ads may count as one

Media buyers widely report that Meta groups ads that look alike and serves them to the same people. Bluesense describes near-duplicates being filed together, so a stack of similar ads buys roughly the reach of a few (video). Meta hasn’t documented how it measures similarity, so treat this as field observation, not a published rule.

What Meta does say points the same way. Its page on managing ad volume warns that when too many ads run at once, each one delivers less, fewer leave the learning phase and results get worse. It recommends fewer ads per ad set, with diverse creative inside them.

What makes two ads different

  • Different: a new persona, a new problem or desire, a new offer, a new format, a new person and place on screen.
  • Not different: a new background colour, a new button, the same script with a new first word, the same shots recut.
  • In between: a new script in exactly the same format. Since Andromeda, he finds it can be read as the same ad, so add a visible change too.

A concept is persona × angle × offer × format

Bluesense’s working definition of a concept has three core parts and an optional fourth (video). Change any one and, by his rule of thumb, you have a new concept that can reach new people.

  • Persona: a problem, a desire and a buying trigger, as narrow as you can make it. “Women 25–45” is not a persona; “a nurse who drinks warm water all night because she can’t leave the ward” is.
  • Angle: the argument. Pain, a contrarian truth, proof, a comparison, a story.
  • Offer: what they get and why now. A bundle, a gift with purchase, a guarantee, a deadline, not only a discount.
  • Format: the shape of the ad. Problem → solution, lifestyle, before / after, unboxing, promo.

His order for testing has moved between videos; in October 2026, for e-commerce, it was persona, then offer, angle and format. In every version the offer is the biggest lever and the hardest to change, and most teams test format first only because it is the easiest.

Worked example: one product, six concepts

Take an insulated one-litre steel bottle whose product page says it keeps drinks cold for 24 hours and has a leak-proof lid. Each of these six concepts differs from the others on at least two axes, with a different person and place on screen.

ConceptPersonaAngleOfferFormat and opening
1. Night shiftNurse on 12-hour shiftsWarm water by hour four, no time to refillNone, full priceProblem → solution, opens on the pain point
2. School runParent packing two school bags at 7 a.m.Nothing leaks onto the homeworkTwo-pack, one per child3 reasons to buy, opens on a question
3. Desk daysOffice worker trying to drink more waterOne fill in the morning, one after lunchNoneLifestyle, opens on a POV
4. The trailWeekend hikerStill ice at the summitNoneBefore / after, opens on the result
5. The giftFriend of a runner, buying a presentA gift they’ll use every dayGift wrapping includedGift idea, opens on a story
6. This weekShopper already comparing bottlesProof plus a reason to act now15% off until SundayOffer, 15 seconds, opens on the offer

Concepts 1 to 5 speak to people who aren’t shopping for a bottle yet, concept 6 to people who are. In Bluesense’s reading, most brands make only the sixth kind, which is shown mostly to warm audiences, so new reach stalls.

Angle briefs for concepts 1 and 5
For nurses on 12-hour night shifts who can’t leave the ward. Open on warm water at 3 a.m., end on the bottle still cold after the shift. No discount. Tone: tired, dry humour.

For someone buying a present for a friend who runs. Speak to the buyer. Story: the friend who always forgets water on long runs. Gift wrapping included.

What this asks of your production

If diversity is the lever, the unit of production is the concept, not the cut. Bluesense says his team moved from 10 hooks × 3 bodies on one concept, 30 ads, to 2–3 hooks × 1–2 bodies across about five concepts (video). Thirty fully unique concepts would cost far more for gains no one can prove, so he stays in the middle.

  1. 1

    Start from real buyers

    Write personas from reviews, support tickets and return reasons. A language model can widen the list; it shouldn’t invent it.

  2. 2

    Give each persona an angle and an offer

    One row per concept, as in the table above.

  3. 3

    Make two or three ads per concept

    Vary the opening and the person on screen; keep the argument.

  4. 4

    Launch one ad set per concept

    So results read by concept. See account structure.

  5. 5

    Read by concept, then refill

    Retire concepts that didn’t earn spend, iterate on those that did, add new ones every month. How many depends on your spend.

Keep a human on concepting: in his view, language models drift to the average, and everyone prompting the same model gets the same ideas. Hand AI the scripts and production; choose the personas and offers yourself.

With Dokiri Studio

Dokiri makes the production step cheaper. It doesn’t choose your personas or run your campaigns. Two ways to use it here:

  • One concept at a time: paste a row of your table into the Angle, offer or audience field and make 1 ad in the format you picked.
  • One persona, many shapes: run a batch of 5 or 10 concepts with one angle. A planner varies format, person, place and opening across eight hook kinds, uses no format or hook kind more than once in 5, and rewrites or drops near-duplicates.

Each batch concept has a Headline and a Primary text for Meta’s fields, and the ZIP’s ads.csv names every file by concept and format. Videos are 9:16 only: no statics, no feed-ratio versions.

Frequently asked questions

Does Meta confirm that similar ads are grouped together?

Not in those words. Meta says too many ads at once dilutes learning and recommends diverse creative; the grouping of near-duplicates is what media buyers report from their accounts. Planning as if it’s true costs nothing: make your variations genuinely different.

Should I stop using interests?

You can still add them, but with Advantage+ audience they guide delivery rather than limit it. Keep the inputs that stay fixed: location, minimum age, language, and exclusions such as past buyers.

How different do two ads need to be?

Change at least one of persona, angle or offer, and make it visible: a different person, place and opening. A new colour or a new first word makes a variation of the same ad.

Does Meta treat AI-made ads differently in delivery?

We know of no Meta statement saying so, and its own Advantage+ creative tools use generative AI. The real risks are sameness, if you generate fifty near-identical versions, and trust, if an ad shows the product doing something it can’t or a generated person posing as a customer.

How many concepts should I run at once?

It depends on spend. Bluesense’s rough tiers are one to three live concepts under $100k a month, more above that. Our guide to how many new ads you need works it through.

Try it

Tools in this article

Turn your product into ads today

Drop a product photo or a store link: the scripts are written for free, and you pay only for the videos you make.

Make my first ad

Keep reading

All articles