Meta Ads in the Age of Andromeda: The 2026 Playbook

If you run Meta ads for a local business, the platform you learned three years ago no longer exists. Meta has rebuilt its ads delivery system around large-scale AI models with names like Andromeda and GEM, and the practical result is simple: the machine now does most of the targeting, and your job has shifted to feeding it clean signals, distinct creative, and enough budget to learn.

That is the whole thesis of this guide. Below we explain what Andromeda and GEM actually are, using Meta's own engineering publications rather than secondhand hype, what changed for small local advertisers, what still works, and the exact 2026 playbook we run for restaurants and local service businesses, including the budget math that determines whether your account can learn at all.

What is Andromeda, actually?

Andromeda is Meta's ads retrieval engine, announced on December 2, 2024 on the Meta engineering blog. Its job is the first stage of ad delivery: out of the millions of ads that could be shown to a given person, it selects the few thousand candidates that enter ranking and the auction. It runs on NVIDIA Grace Hopper hardware and Meta's own MTIA accelerators.

Meta reports that Andromeda enabled roughly a 10,000x increase in model complexity at the retrieval stage, and cites a 6 percent gain in recall and an 8 percent improvement in ads quality on tested segments. Those are Meta's numbers, not independent measurements, but the direction is consistent with what buyers see in accounts: the system has become dramatically better at finding the right people for an ad on its own.

The practical translation for a restaurant or plumber in Austin: Andromeda decides whether your ad even makes it into consideration for a given person. Your targeting settings still gate who is eligible to see the ad; within those constraints, the retrieval engine chooses among eligible ads on predicted performance, which is where your creative and conversion signals do the work.

What is GEM, and how does the 2026 stack fit together?

GEM, the Generative Ads Recommendation Model, is the newest layer. Meta announced it on November 10, 2025, describing it as an LLM-scale foundation model for ads recommendation that began shipping on Reels in the second quarter of 2025 and transfers what it learns to the serving models through knowledge distillation. The primary source is Meta's own engineering post; InfoQ published a useful third-party technical summary in December 2025.

The 2026 delivery stack is therefore three layers working together:

1. Andromeda handles retrieval: which few thousand ads, out of millions, get considered for this person.
2. Lattice, Meta's 2023 consolidation of its ranking models, handles ranking: which of those candidates actually win.
3. GEM sits above both as a foundation model, teaching the serving models what it learns at scale.

Meta reports GEM drove a 5 percent lift in ad conversions on Instagram and 3 percent on Facebook Feed. Again, company-reported figures. The point for a small advertiser is not the exact percentages. It is that every layer of the machine is now a learning system, and learning systems reward advertisers who give them consistent, honest data.

What changed for small local advertisers?

Three things changed, and they compound.

First, Advantage+ became the default path, not the power-user option. Starting in February 2025, Meta streamlined campaign setup around Advantage+, renamed Advantage+ Shopping to Advantage+ Sales, extended the automation to leads and app objectives, and completed the rollout in the second quarter of 2025. Social Media Today covered the consolidation, and Jon Loomer's breakdown remains the clearest practitioner walkthrough. Meta reports roughly a 22 percent average return on ad spend improvement for Advantage+ Sales and about 10 percent lower cost per qualified lead in early Advantage+ leads testing, and it launched an "Opportunity Score" to push advertisers toward its recommended setup. Treat those figures as Meta's marketing of its own product, but the structural fact stands: manual campaign types are being folded into automated ones.

Second, the endgame is full automation. As the Wall Street Journal reported in June 2025, Mark Zuckerberg's stated goal is full campaign automation by the end of 2026: an advertiser sets goals and a budget, and AI handles creative, targeting, and allocation (coverage here). As of mid-2026, URL-to-campaign automation is in testing with select advertisers. Meta also reports that more than one million advertisers already use its generative AI ad tools, producing over 15 million ads per month.

Third, manual levers matter less, inputs matter more. When retrieval and ranking are this automated, the things you still control are your creative, your conversion signal quality, your budget adequacy, and your offer. Everything else in this guide is about those four inputs.

What still works in 2026?

The practitioner consensus is remarkably uniform on this, and it matches what we see in the accounts we run. Three principles hold.

Creative is the targeting input. Under Andromeda, the system reads your creative to decide who should see it. A video about happy hour targets happy hour people; you do not need an interest stack to tell Meta that. Shops like Triple Whale and Foxwell Digital have written the best public explanations of this shift.

Broad targeting routinely beats narrow interest stacking. This is practitioner consensus, not a Meta rule, but it is consensus for a reason: narrow audiences starve the learning system, and the retrieval engine already knows more about your likely customer than an interest list does. For a local business, "broad" means a sensible radius around your location with age limits where relevant, and almost nothing else.

Three to five genuinely distinct creative concepts beat many near-duplicates. Practitioners consistently report that the system treats near-duplicate variants as one entity, so ten versions of the same ad with different headlines gives the machine one idea to test, not ten. Distinct concepts means different angles entirely: the dish, the room, the story, the offer, the review.

What does the 2026 playbook look like?

Four moves, in order: a simple structure the system can learn on, a budget that clears the learning threshold, a sustainable cadence of distinct creative, and clean conversion signals. Everything else is a variation on those four. Here is how we install each one for a local restaurant or service business.

Account structure

Keep it almost embarrassingly simple:

• One campaign per real business objective. For most local businesses that is one: reservations, orders, or booked calls. A second campaign only if you genuinely run a second objective, such as catering leads alongside dine-in.
• One or two ad sets per campaign. Broad local audience, radius set to your realistic draw (for an Austin restaurant, that is usually 5 to 10 miles, not the metro).
• Three to five distinct creative concepts per ad set, per the consensus above.
• Advantage+ where offered. Fighting the default path in 2026 means fighting the delivery system itself.

What you should not build: separate ad sets per interest, per zip code, or per demographic slice. Every split divides your conversion events, and conversion events are the scarce resource, which brings us to the math.

The 50-event budget math

This is the single most important number in Meta advertising, and it comes from Meta's own documentation: an ad set needs roughly 50 optimization events per seven days to exit the learning phase and stabilize.

From that, practitioners derive a minimum budget formula:

Minimum daily budget = (target cost per result × 50) ÷ 7

Worked example for a restaurant optimizing for reservation or lead events that cost $5 to $10 each:

• At $5 per event: ($5 × 50) ÷ 7 ≈ $35 per day
• At $10 per event: ($10 × 50) ÷ 7 ≈ $70 per day

So a realistic floor for a restaurant is roughly $35 to $70 per day per ad set. If your budget only supports one ad set at that level, run one ad set. A $30 per day budget split across three ad sets is not a small version of a good account; it is three ad sets that never learn.

Two rules protect the learning you paid for: never swing a budget more than 20 percent in one move, and never edit targeting mid-flight. Both reset learning, and a reset means buying those 50 events again.

For context on whether those event costs are plausible: per WordStream's 2025 Facebook ads benchmarks, the most recent dataset the firm has published, restaurants are among the cheapest verticals on the platform, with lead campaign costs per click around $0.74 and cost per lead around $3.16, against $27.66 across all industries. We go deeper on that math in our answer on whether Facebook ads are worth it for local restaurants.

Creative cadence

The machine consumes creative faster than most owners produce it. Our working cadence for a local account:

• Launch with 3 to 5 distinct concepts. For a restaurant: a signature dish close-up, the room at full energy, the owner or chef on camera, a customer review card, and the current offer.
• Review every two weeks. Retire the clear loser, keep the winners untouched, and introduce one new concept. Do not pause winners to "freshen things up."
• Shoot vertical video first. Meta shipped GEM on Reels before anywhere else, and phone-shot vertical video is the one format a small business can produce at a sustainable cadence.
• Batch production monthly. One hour of phone footage on a busy Friday yields a month of concepts. Polished production is optional; distinct angles are not.

Conversions API and signal quality

The delivery system is only as smart as the signals you feed it. The browser pixel alone misses a meaningful share of conversions, so connect the Conversions API (CAPI) so events flow server-side as well:

• Many POS, ordering, and reservation platforms offer built-in Meta tracking integrations. Confirm exactly what yours supports, and turn on server-side event sharing wherever it is offered.
• Send real conversion events (reservation completed, order placed, lead submitted), not just page views.
• Deduplicate pixel and CAPI events with a shared event ID; the platform integrations handle this for you.
• Check Events Manager monthly for event match quality and fix anything flagged.

This is unglamorous plumbing, and it is also the highest-leverage hour a local advertiser can spend, because every downstream decision Andromeda and GEM make about your account is built on these events. If your pixel is still broken or thin while this gets fixed, our POS data playbook shows how to build custom audiences straight from your point-of-sale system in the meantime.

What we'd do first

If we took over your Meta account tomorrow, the first week would look like this:

1. Verify the signal. Pixel plus CAPI connected, real conversion events firing, dedupe confirmed.
2. Collapse the structure. One campaign, one broad local ad set, Advantage+ path.
3. Run the budget math. Estimate the cost per event, apply the 50-event formula, and set a daily budget the account can actually learn on. If the math does not clear roughly $35 per day, we would tell you to fix that constraint before spending at all.
4. Ship 3 to 5 distinct concepts and put the two-week review cadence on the calendar.
5. Leave it alone. No mid-flight edits, no budget whiplash, and a scheduled read of results at the two-week mark.

Meta ads is often only half of a local acquisition system; the other half is search intent, which we cover in our Google Ads guide for restaurants, and both belong inside a coherent marketing program rather than running as disconnected experiments.

If you would rather have an operator who has run this at Meta scale look at your account directly, request our free business audit and we will show you exactly where your setup stands against this playbook.