Meta's Andromeda Update: What It Means for Hospitality Ads

Andromeda is the part of Meta’s ad system that decides which ads get a chance to show, and in October 2026 Meta’s engineers explained how they’ve rebuilt it. The engineering is impressive, but the measured gain is modest, so for anyone running Meta ads for a restaurant, bar, hotel or leisure venue, the real change is where your effort goes: less fiddling with targeting, and more genuinely different ads.

This guide covers how it works in plain English, what’s actually new, and what to do about it, with the technical detail at the end if you want it or want to forward it to your agency. We’ve stuck closely to what Meta has published, because plenty of agency advice about Andromeda goes further than Meta has, and where something is our own opinion, we’ll say so.

How Meta Picks Which Ad Someone Sees

Every time someone opens Instagram or Facebook, Meta has a split second to pick one ad from millions, and because it can’t look closely at all of them, it works through a few steps:

1. First comes a targeting check, where Meta drops any ad that isn’t aimed at this person, so if they’re outside the area you’ve chosen, your ad stops there.
2. Next, Andromeda gives every remaining ad a quick score and passes only a shortlist on, or as Meta puts it, it “quickly scans tens of millions of ads and selects a few thousand strong candidates”.
3. Then comes ranking, where bigger, slower models look properly at that shortlist and predict how likely each ad is to get the result you’ve asked for, such as a booking or a click.
4. Finally, an auction decides which one ad wins the spot.

A useful way to picture it is that Andromeda is a recruiter skimming a huge pile of CVs, and ranking is the hiring manager who interviews the shortlist.

Meta first wrote about Andromeda in December 2024, and agencies such as Confect say it was fully rolled out by October 2025, although Meta hasn’t given a date itself.

What’s New in the Rebuilt Andromeda

The first sift used to be the rough one, which meant a good ad could be cut before the careful step ever saw it. Meta says the new version reaches quality that “approaches ranking-level”, so every ad that’s allowed now gets a proper look, and in recruiter terms, every CV in the pile gets read.

It also builds a richer picture of each person. Rather than one profile that averages everything out, Meta now keeps several, each tracking a different kind of interest, and matches each ad against the ones that are relevant to it. There’s more on how that works in the technical section below.

Here’s a simplified example of our own, since Meta doesn’t say what each profile focuses on. Picture someone who lives a mile from your venue. This month they’ve been looking at places for a friend’s birthday dinner, they’re also a regular at Sunday roasts, and on Fridays they’re the after-work drinks type. One averaged picture of them would make all of that a lukewarm match for everything, whereas a model built like the one Meta describes can judge your group-dining ad against the birthday-planning side of them, which is hot right now, and your Sunday roast ad against a different side.

Meta also says each ad is combined with “real-time serving context” before it’s scored, meaning the circumstances when it would be shown. It doesn’t say exactly what that includes, but our guess is that the same 30-second video could work on a Saturday night on the sofa and miss on a Tuesday morning commute with the sound off.

The system change is big, but the measured gain is modest. The only result Meta has shared is that “since 2025, Meta Andromeda has delivered a 2.4% increase in conversions on Facebook”, where a conversion is the result you’re paying for, such as a booking or a sign-up. Meta’s advice to make genuinely different ads isn’t new either, as it dates back to 2025, but our read is that now every ad is scored and matched more finely, a stack of ads that all look alike gives you even less than it used to.

Why Your Ads Need to Be Genuinely Different

This is the part that matters most for your account, because with Andromeda doing more of the matching, the ads you give Meta decide which of the people in your area it can find. Meta puts it plainly: “the focus has shifted from niche targeting to creative diversification as the best lever to find the most relevant audiences.” In other words, within the area and audience you’ve set, different ads are how Meta reaches different people, and that does more than narrowing your targeting.

The catch is that Meta groups ads that look alike. Its system “groups together ads that share key visual and thematic attributes”, meaning ads that look similar or are about the same thing, and those ads are “seen as variations of the same creative”, with Meta sharing what it learns between them.

Meta draws a useful line between two things here. Iteration is tweaking an ad you already have, and Meta’s example is two ads “with identical visuals, but different text CTAs” (the button text, like Book Now). Diversification is making genuinely different ads, such as a Reels ad aimed at one audience and a partnership ad with a creator who appeals to another.

It matters because, as Meta explains, when ads are too similar its system “may not recognise it as different” and could show them to the same people. Meta says that overlap limits your reach and “ultimately leads to creative limited or creative fatigue status”, which are the statuses you’ll see in the Delivery column in Ads Manager (fatigue being Meta’s word for people seeing the same image or video too often).

None of this means dropping your brand style. Our advice is to keep the cues that make you recognisable, like your logo, colours and end frame, and vary the main visual and what each ad is about.

You’ll also come across some rules that aren’t in anything Meta has published, at least that we’ve found:

What This Looks Like in a Real Ad Account

Hector Serrano, Head of Content at the dictation app Wispr Flow, described how his team handles this on Lenny’s Podcast in November 2026. Wispr runs hundreds of creator ads a month from a deliberately mixed group, from mums and lawyers to a 45-year-old ecommerce founder, and Hector brings each new creator in to fill a specific gap rather than to add more of the same. He warns that once creators can see which ads get the most spend, they tend to copy them, leaving you with lots of ads that look and sound alike, and he says some of his best performers were ads he’d never have backed (“what the heck is this?”), while some of Wispr’s best ads are also its plainest, like a long screen recording with a line of text on top.

For a hospitality business, the equivalent might be a regular’s phone video of their Sunday lunch arriving, a guest showing off their hotel room or a chef talking through tonight’s specials, running alongside your professional shoot. You can’t know in advance which one Meta will match to the most people, so it pays to run genuinely different ideas and let the results decide.

What to Do Next

Here’s what we’d do this week. These are our suggestions based on what Meta has said rather than Meta’s own instructions, so treat them as a starting point and test them in your own account.

1. Check your location settings first, because Andromeda only scores your ad for people your targeting allows. For a local restaurant, bar or café, that usually means a tight radius around each site, often 1 to 3 miles in a city, while hotels and destination venues will want to reach further, to wherever their guests travel from. If you use Advantage+ Audience (Meta’s automatic targeting), interests and age ranges are treated as suggestions, but location, minimum age and exclusions stay as strict limits, and there’s more on setup in our Meta ads guide for hospitality.
2. Count the genuinely different ideas you have running, rather than the number of ads. The same video with three different headlines is really one idea, whereas a close-up of your signature dish, a behind-the-bar cocktail clip and a creator’s first visit are three, because they look different and they’re about different things. If most of an ad set is one idea, that’s the first thing to fix.
3. Replace near copies with one or two new ideas, each clearly about one occasion, like bottomless brunch, a spa day or a private dining room. Since Meta matches each ad to the side of a person it’s most relevant to, our read is that an ad that’s plainly about one thing gives it a clearer match than one that tries to cover everything.
4. Make each ad work in the moment with vertical video, captions and a first second that works with the sound off, because where and when your ad appears probably matters as well as who sees it.
5. Give new ads about a week before you judge them, which is our working rule rather than Meta’s. Meta scores every eligible ad but won’t spend on all of them, so low spend on its own isn’t a verdict, and if a quiet ad looks a lot like one of your other live ads, a genuinely different idea is a better next test than a new caption.
6. Keep your ad sets few and well funded. If you run several sites, a few ad sets that each cover several locations usually learn faster than one ad set per site, because Meta needs around 50 results a week in an ad set to finish its learning phase, and ad sets with overlapping areas, which is common in cities, end up competing for the same people. On a small budget, such as £500 a month across the whole account, we’d run a single ad set with three to five genuinely different ads rather than ten near copies.
7. Plan a steady supply of new ideas. Confect’s study of 3,014 ecommerce advertisers puts the time an ad keeps working at 2 to 4 weeks, and while that’s ecommerce rather than hospitality, local audiences tire quickly too, so a few fresh ideas each month is a sensible rhythm, and more if you’re spending more.
8. Make sure you can see what’s working, because none of this helps if you can’t tell which ads bring people in. Our guide to tracking bookings by booking platform covers venues that take bookings, and for walk-in venues like cafés, an offer that’s only in the ad and redeemed in store is a simple way to see which ads work.

Where Creator Content Fits

One of Meta’s own examples of a genuinely different ad is a partnership ad “featuring a creator who connects with new moms”, alongside a Reels ad for Gen Z. In other words, a different person speaking to a different audience is the kind of variety Meta is asking for, and it’s also the hardest thing for a hospitality business to make on its own.

Different creators tend to film your venue in different ways, with a different face, angle and moment of the visit, as long as you brief them to bring their own take rather than copy whatever’s already working. There are two ways to get that with Joli:

• Partnership ads from creator visits, where creators from Joli’s network of 10,000 vetted creators visit your venue for a gifted experience and post about it, and you ask whether you can run their post as a partnership ad on Instagram or a Spark Ad on TikTok. It’s their choice, and the ad runs from your own Ads Manager.
• Joli Ad Studio, where vetted UK creators film videos at your venue to your brief, ready to run as ads with unlimited usage rights, so you can brief different creators for different occasions and test new openings on the ideas that work.

If you’d like to see how it could work for your venue or group, book a demo.

The Technical Detail

This section is for anyone who wants the specifics, or wants to check what an agency is telling them. Everything here comes from Meta’s own posts, mainly its October 2026 engineering post and the original December 2024 post.

The Pipeline in Meta’s Terms

Meta describes Andromeda as its “machine learning system for personalized ads retrieval, the first stage of the ads recommendation pipeline”. In the 2026 version, three steps run together on the same GPU server. Index matching enforces advertisers’ targeting rules across “tens of feature categories” and updates in real time as ads are created, changed or expire. The retrieval model then scores every eligible ad, and top-K selection picks the shortlist that’s passed on to ranking. Running all three on one machine cut network traffic by about 100 times, according to Meta.

How the New Model Reads People and Ads

The 2026 retrieval model uses “an ensemble of specialized user towers as experts”, which is a mixture-of-experts design. A tower is a model that turns signals about a person into a profile that can be compared with ads, and an “ad-aware” router activates “only the top-K most relevant experts” for each ad to produce its relevance score. Meta describes the interests these experts capture as “diverse, time-evolving, and contextualized”.

Meta lists four kinds of signals feeding the user side: “sequential ad engagement events, user-cohort representations, LLM-derived semantic features, and cross-surface co-engagement signals”. Roughly, that’s the order in which someone has engaged with ads, groups of similar people, the meaning of content as read by large language models, and what people engage with together across Meta’s apps and surfaces. On the ad side, ad content is combined with “real-time serving context” before scoring. Meta doesn’t say what each expert specialises in, how many experts there are, or exactly what the serving context includes.

The Numbers, and What Each One Measures

Meta says the rebuilt Andromeda processes over one billion ads per second on a single GPU server in production (Meta Engineering, 28 October 2026). Most of the other figures in the post are efficiency gains or lab benchmarks rather than results for advertisers:

For comparison, Meta’s December 2024 post reported a 6% recall improvement for the retrieval system and an 8% ads quality improvement on selected segments for the original Andromeda. Those are 2024 figures and aren’t results of the rebuild.

What Meta’s Engineering Posts Don’t Cover

Neither engineering post gives advice on creative, budgets, account structure or testing, and neither mentions grouping similar ads. That guidance comes from Meta’s advertiser posts of April and December 2025, quoted earlier. Meta has also written about the ranking side of its ad system (GEM, October 2026) and about how it shares out computing power across ad requests (Intelligent Delivery, November 2026), and neither asks advertisers to change anything.

Sources

• Meta Engineering, Meta Andromeda: Powering the Next Generation of Ads Retrieval Through Model-System Co-Design, 28 October 2026
• Meta Engineering, Meta Andromeda: Supercharging Advantage+ automation with the next-gen personalized ads retrieval engine, 2 December 2024
• Meta for Business, Demystifying Creative Diversification, 16 December 2025
• Meta for Business, The Creative Advantage: Unlocking the Power of Diversification with Meta Andromeda, 22 April 2025
• Meta for Business, Creative differentiation
• Meta Engineering, Meta’s Generative Ads Model (GEM), 13 October 2026, and Intelligent Delivery, 11 November 2026
• Lenny’s Podcast, A UGC masterclass from the 22-year-old who built Wispr Flow’s growth machine, with Hector Serrano, 1 November 2026