Meta Andromeda 2.0: What the 28 October 2026 Update Changes for Window and Door Lead Generation

01What Meta published on 28 October 2026

On 28 October 2026 Meta's engineering team published a technical account of what it calls the 2026 Meta Andromeda, under the title Meta Andromeda: Powering the Next Generation of Ads Retrieval Through Model-System Co-Design. It is a twelve-author post in the machine learning section of Engineering at Meta, and the useful way to read it is as a description of a stage of ad delivery that advertisers have never been able to see.

Andromeda is the retrieval stage of Meta's ads system, first described publicly on 2 December 2024. Retrieval runs before ranking. It takes the pool of ads that are eligible for a given person — campaign objective, budget, targeting rules, policy — and narrows it to a few thousand candidates in milliseconds. A more expensive ranking stage, now led by models such as GEM, then scores those survivors in detail. Only after that does the auction decide the price.

The 2026 version changes the first of those three steps in three ways: a mixture-of-experts model for retrieval, a single fused GPU kernel per user-ad pair, and the co-location of index matching, scoring and top-K selection on one GPU server. Section three translates each one. What matters first is what the post is not.

One naming note, because it matters when you read vendor material: Meta calls it "the 2026 Meta Andromeda". Trade coverage calls it the second generation, and "Andromeda 2.0" has become the shorthand. All three refer to the same 28 October engineering post.

02Why a retrieval upgrade changes lead generation at all

Here is the mechanic that makes a GPU throughput figure commercially interesting to a window dealer: an ad that is not retrieved cannot win, whatever you bid. Retrieval happens before ranking and before the auction. If your creative is not in the few thousand candidates that survive the first cut, your bid, your budget and your landing page are never consulted.

And the reason Meta rebuilt retrieval is the part of the post that most coverage skipped. Meta says generative AI creative tools have helped advertisers produce "hundreds of millions of new ads", and that the pool keeps growing because of Advantage+ automation and AI creative tooling. The retrieval stage was rebuilt to handle a candidate pool that had outgrown the architecture designed for it.

Two consequences follow, and they pull in opposite directions for a contractor.

The good news: a smaller advertiser with genuinely different creative is no longer structurally disadvantaged by a modest budget. If retrieval quality improves across the pool, being one of the few accounts in your market with four distinct concepts should matter more than being the account with the most interest layers.

The bad news: the cost of keeping up rises, because near-duplicates no longer buy you additional chances. A dealer producing forty ads a month from one photoshoot is, in Meta's own terms, producing roughly one creative. The system is not short of ads. It is short of ideas it can tell apart.

03The three changes, and which numbers matter

Meta's post contains a lot of engineering figures. Almost none of them are advertiser outcomes, and confusing the two is how a dealer ends up paying for a "100x performance" narrative. Here is the full set, with what each one is and is not.

The three architectural changes are worth understanding at a conceptual level, because two of them bear directly on what kind of creative gets matched.

That last clause is the most misinterpreted sentence in the whole post. "Meta moved targeting onto the GPU" is being retold as "targeting no longer matters". What Meta wrote is that index matching enforces advertisers' targeting rules on the GPU. Your audience still bounds eligibility. What changed is where within those bounds the matching happens, and how many candidates get considered.

04What Meta did not say

Andromeda is unusually fertile ground for invented certainty. These are the claims circulating in contractor marketing that have no source in Meta's documentation.

• A creative quota. Meta has published no minimum or maximum number of ads per ad set for Andromeda. Its help documentation caps ad sets at 50 ads and notes one ad can hold up to 10 creative assets, and its ad-volume guidance warns that too many ads can hurt performance. The numbers you see quoted — eight, ten, fifteen, twenty — are agency convention, and Meta's only published figure of that kind was 20 diversified ads for Advantage+ shopping campaigns in November 2024.
• A similarity percentage. No threshold, no similarity score, no "ads above 60% get collapsed". Meta says look-alike ads are grouped as variations of one creative. It has never published a number at which that happens.
• A penalty. Meta's guidance describes shared learnings and delivery at the creative level. It does not describe near-duplicates being blocked or suppressed. The practical cost is opportunity, not punishment: twenty near-copies teach the system what one concept can do.
• An account-level rollout date. The post gives none. Neither did the December 2024 post.
• Instagram results. The 2.4% figure is Facebook-only, with no method published, and no Instagram equivalent is given anywhere in the post.
• Credit for the quarterly numbers. Meta's Q3 2026 results credited Andromeda improvements with +1.4% Facebook conversions. The +8.3% clicks and +15.7% conversions figures reported for Q2 2026 cover broader ranking work, not Andromeda alone. The GEM post of 13 October 2026 claimed +18.9% conversions across surfaces, also without a baseline. Three different numbers, three different systems — do not add them up.

The honest summary is that the engineering is verifiable and the marketing claim is not. That is normal for platform announcements, and it does not make the operational advice wrong. It makes the arithmetic untestable in your account, which is exactly why your own measurement has to be built before you change anything.

05Creative is now the targeting control

Meta has been steering advertisers this way for eighteen months, and the direction is documented rather than inferred.

In April 2025 Meta published guidance stating that the focus had shifted from niche targeting to creative diversification, which it called the best lever for finding the most relevant audiences. In December 2025, in a post titled Demystifying Creative Diversification, Meta spelled out the mechanism in the terms that matter here: its ads system analyses creative and "groups together ads that share key visual and thematic attributes", and "when multiple ads look or feel alike, these are seen as variations of the same creative, meaning learnings and delivery optimizations are shared at the creative level, not just the individual ad level."

The same post draws a line most accounts never notice. Creative iteration is the manual A/B habit — two ads with identical visuals and a different text call to action. Creative diversification is distinct assets built for different personas or use cases. Meta's own figures, which are vendor claims, put advertisers leaning into diversification at 11% higher click-through rate and 7.6% higher conversion rate. The post also states plainly that creative similarity, meaning images that are visually too alike, can drive fatigue.

Ads Manager exposes this as a Creative diversity column that compares the images and video thumbnails in an ad set and labels it High, Medium or Low. Meta marks it as estimated and in development. Note what it compares: thumbnails. For a video ad, that makes the first frame — not the script — the thing being measured for distinctness.

The look-alike trap in this specific category

Window and door advertising is unusually prone to accidental duplication, because the category has a small set of visual clichés that every dealer reaches for. If your account contains these, you do not have six creatives. You have one.

• The stock couple at the window. Same pose, same light, same smiling profile — often literally the same stock licence as a competitor's.
• The clapperboard before/after. Facade, split screen, no address, no product name.
• The showroom pan. Slow dolly across sample frames with a logo sting.
• The price banner. "0% financing", "free quote", "save $1,000" over the same photograph.
• The truck and crew. Wide shot, arms crossed, no differentiation between companies.
• The neighbour testimonial. Talking head on a porch, unlabelled city, generic praise.

None of these is a bad ad. The problem is that a retrieval model comparing thumbnails across a market sees a wall of near-identical imagery, and a dealer whose entire account is built from those six templates has given the system nothing to route. The competitor who films a failed seal being cut out of a fifteen-year-old unit has given it a distinctive object to match.

06What "different" means for a window and door dealer

The unit of work has changed. It is no longer "how many ads can we produce this month", it is "how many genuinely different reasons to replace windows can we put in front of people". For this trade there are more than most dealers use, because the purchase has several independent motivations.

Nine is not a monthly target. It is a library, and the windows and doors marketing material we publish treats it as one that grows with each installation season. For a dealer spending around $4,000 a month in a single market, three concepts running at once with a fourth in production is a defensible rhythm; replacing a concept when Ads Manager flags creative fatigue — which Meta defines as cost per result at twice or more your past ads — is a better trigger than a calendar.

The production implication is uncomfortable and worth saying plainly: the median window ad does not require a video shoot, it requires a subject. Most of the nine concepts above can be shot by an installer with a phone in twenty minutes during a normal working day. What they cannot be is invented by a designer working from a template, because a template is how you get six versions of the same idea. The window company marketing strategies we use put creative concept ownership with the operations team for exactly this reason.

07Your ad copy is a retrieval input

Meta's post lists what the specialised user towers draw on. One entry in that list is new enough to be worth acting on: LLM-derived semantic features. Meta does not say which language models produce them, and it does not say how the ad side is encoded beyond a joint encoding that includes real-time serving context. So the following is inference, labelled as such, not something Meta documented.

Two cautions. First, this is not a licence to stuff keywords — Meta's system is matching ads to people, not to query strings, and a keyword-stuffed primary text reads as spam to the human who matters. The rule is one specific claim per creative, stated in the words a homeowner would use. Second, we have no evidence that Meta reads your landing page semantically for retrieval purposes, so treat the ad as the surface you actually control at this stage and put the page work where it belongs — in the AI search side of this trade, which is a different pipeline with different rules.

08The signal problem in a three-week sale

Better retrieval does not fix a weak conversion signal; it amplifies whatever the signal asks for. And window and door has a structural signal problem that most accounts never resolve, because the sale is slow and the events are few.

The purchase is a considered one. In This Old House's 2026 windows survey of 1,000 US homeowners, 68% researched for one to four weeks and 84% collected at least two quotes before deciding. The average project came in at $7,183, roughly $517 per window. That is not a click-to-buy funnel, and Meta's optimisation machinery is built around events that happen in volume.

Meta's own guidance treats roughly 50 optimisation events per ad set per week as the point at which an ad set leaves the learning phase. Run the arithmetic on a real dealer: at a $42.95 benchmark cost per lead, a $4,000 monthly budget produces around 93 leads, or about 21 a week. Split across four ad sets — one per neighbourhood, one for retargeting — and no ad set is anywhere near the threshold. The account is permanently in learning, and every structural change restarts the clock.

09Why Meta leads are cheap, and why that is the trap

Meta leads are cheaper than search leads in this category, and the published benchmarks make that unambiguous.

The gap is roughly two to one in Meta's favour on cost per lead, and the explanation is intent. A homeowner searching "window replacement near me" has already decided to get quotes. A homeowner scrolling at 9pm and tapping a before/after is being interrupted mid-thought. That is not a defect of the channel; it is what the channel is. Meta creates demand in the one-to-four-week window where a homeowner is noticing a draft and not yet searching for it. Search captures the moment after the decision.

The trap follows directly. Because interrupted audiences convert more cheaply and less meaningfully, close rates are materially lower. A 138-campaign benchmark of Meta lead-form advertising cited in trade analysis puts close rates in the 8–20% band, against rates above 25% commonly reported for search and LSA traffic in the same trade. Two independent agency accounts of window campaigns report $60–$150 per lead — one published a Sacramento campaign of 54 leads at roughly $78 average, with the best ad nearer $63, on about $4,211 of spend over two months. Both are self-reported and unaudited; they are useful for shape, not for planning.

10The arithmetic that decides your budget

Cost per lead is not a target; it is a ceiling set by job value, margin and close rate. The formula below is a practitioner convention rather than a study, and the divisor of three is a rule of thumb for leaving room for sales cost and overhead — but the arithmetic is what actually governs whether a campaign can work.

Now put that beside the benchmark. At a $42.95 median Meta cost per lead, the whole-home campaign has an order of magnitude of headroom, and the two-window campaign is already within $30 of its ceiling before account management, creative production or a single wasted lead. Those are not the same business, and running them from one campaign with one optimisation event guarantees that the cheap, small-job lead wins the auction for your budget — because it is the one the system can convert most easily.

The structural fix is boring and it is the highest-return change in this article: separate the product lines into separate campaigns with separate budgets, separate creative and separate optimisation events, and put a project-size question on the form so a two-window enquiry cannot masquerade as a whole-home one. That single change does more for a window dealer's Meta performance than any creative variation of the same photograph.

One Canadian caveat before you build a plan on these numbers. The WordStream and LocaliQ benchmarks are US data. Canadian markets for this trade price differently — smaller population per metro, different competitive density, different seasonality around a shorter installation window. Do not convert a US dollar cost per lead into a Canadian target; use the ratio between your own campaigns and hold the absolute figures as a sanity check only. The Canada-focused channel comparison sets out the same caution for the paid search side.

11Account structure after the update

Nothing in Andromeda requires a rebuild. What it does is change the cost of the structure you already have, because a fragmented account splits both the creative pool and the optimisation events.

One new capability is worth knowing about because it lets you split a campaign without splitting the event pool. On 19 October 2026 Meta opened per-creative audiences to Marketing API partners: an age range and gender can now be attached to an individual creative inside an ad set, provided the ad set itself stays open to all genders. For a window dealer that means a creative aimed at the over-55 replacement buyer and a creative aimed at the first-time buyer can live in one ad set and share its learning, instead of in two ad sets that each starve.

Two limits keep this honest. Ad sets are capped at 50 ads, and Meta's own ad-volume guidance warns that too many ads can degrade performance — one ad can hold up to 10 creative assets, which is the right place to put placement variations. The goal is more distinct concepts, not more ad objects.

12Measurement: the 27 October creative breakdown

The useful reporting change came a day before the engineering post. On 27 October 2026 Meta was reported to have added individual performance data for ad creative to Ads Manager's Breakdown: how each image or video performed when delivered as single media, and how each text variation performed. Ads opted into multi-post displays — "Show as carousel" or "Show as collection" — are excluded from that breakdown.

That last exclusion is not a footnote. It means the format most dealers use for installation galleries is precisely the format the new reporting cannot decompose. Practical sequencing: run your portfolio concepts as single-media ads while you are testing which concept earns its budget, and reserve carousel formats for scaling a concept you have already proven.

And a reporting rule that matters more than the tooling: judge concepts, not ads. Meta's own framing is that look-alike ads share learnings at the creative level. A dashboard that ranks eight near-duplicate ads by cost per lead is measuring noise, and it will reliably tell you to scale the variant that got lucky with three conversions in a small sample. Roll near-duplicates into one concept line, then compare concepts.

13The 30-day rebuild

Resist the temptation to do this all in week one. The single largest source of bad decisions in contractor Meta accounts is changing structure and creative in the same week, then attributing the result to whichever change the person who made it prefers. Sequence it, hold each variable for a full week of delivery, and read the metric you actually care about.

14What to stop doing

• Ad-volume theatre. Forty ads from one photoshoot is one creative with forty names. If your monthly report leads with ad count, it is reporting effort rather than distinctness.
• AI variation spam. Meta's own guidance warns that creative similarity drives fatigue, and a generative tool asked for twenty variations of one ad will produce twenty near-duplicates. Use generative tools to execute a concept, not to multiply one.
• Interest stacking. Interest layers micro-manage a decision the system now makes at retrieval. Keep the operational limits you genuinely need — service radius, excluded postcodes — and drop the rest.
• One campaign for two buyers. A $1,800 repair and a $16,000 whole-home replacement cannot share a lead ceiling, an optimisation event or a creative. Sharing them means the small job wins your budget by default.
• Optimising to a form fill. It is the cheapest event and the least predictive. Every improvement in retrieval makes this mistake more expensive, not less.
• Boosting posts and calling it paid social. A boosted post has no optimisation event, no concept rotation and no place in this structure.
• Buying "Andromeda-optimised" packages. There is nothing to optimise against. Ask what the vendor will produce — concepts, copy claims, CRM integration — because that is what they can actually deliver.

15Where Meta sits beside Google, LSA and AI search

Andromeda changes how well Meta does its job. It does not change what Meta's job is, and a window dealer who expects paid social to behave like search will keep being disappointed by a channel that is doing exactly what it was asked.

The purchase timeline sets the division of labour. In This Old House's 2026 data, 68% of homeowners spent one to four weeks researching and 84% collected at least two quotes. During that window a homeowner is noticing a draft, looking at a neighbour's new windows, reading about rebates and eventually searching. Meta can be present in the noticing phase. Search and LSA capture the searching phase. The research layer — cost ranges, rebate rules, U-factor thresholds — is increasingly answered inside an AI assistant before a search ever happens, which is why the category marketing strategy treats those pages as infrastructure rather than content.

For dealers who want the paid social layer handled as one component of a full programme, that is the remit of our window and door internet marketing practice. The practical allocation for most dealers we work with is unchanged by the update: search and LSA for capture, Meta for demand creation and retargeting, and the page layer for the research that happens before either. What the update does argue for is a bigger share of the Meta budget going to production rather than to media, because the marginal return has moved from audience construction to concept distinctness. A dealer who shifts 10% of the Meta budget from media into four concepts a month will usually find the media performs better at the same spend.

16Sources and confidence

Every figure above traces to one of the following. Where a source is a company with a commercial interest in the conclusion, it is labelled.

Three caveats travel with this. The engineering numbers describe throughput on Meta's servers, not outcomes in your account, and Meta has published no account-level rollout date. The cost benchmarks are US data collected on different datasets and windows, so the Meta-versus-Google gap is directional rather than precise, and it should not be converted into Canadian dollars. And the creative-diversity thesis is Meta's own advertiser guidance attached to a technical post — it is well evidenced as guidance, but the causal chain from "distinct concept" to "cheaper held appointment" is one you have to test in your own account, because nobody has published that measurement.

17Questions we get asked

18Glossary

Andromeda

Meta's ads retrieval system, first described 2 December 2024 and substantially rebuilt in a post published 28 October 2026.

Retrieval

The first stage of an ad request: narrows the eligible pool to a few thousand candidates before ranking. An ad that is not retrieved never reaches the auction.

GEM

Meta's Generative Ads Recommendation Model, the ranking stage that scores retrieved candidates in detail.

Mixture-of-experts

An architecture with many specialised sub-networks where a routing component activates only a few per input. Meta's 2026 retrieval model uses specialised user towers as experts.

Eligible pool

Every ad permitted for a given person at a given moment by campaign objective, budget, targeting and policy.

Creative diversification

Meta's term for distinct assets built for different personas or use cases. Distinct from iteration, which is A/B refinement of one idea.

Creative iteration

Manual refinement of one concept — same visuals, different headline or call to action. Meta treats these as variations, not new creative.

Creative diversity label

An Ads Manager column comparing images and video thumbnails in an ad set and labelling it High, Medium or Low. Meta marks it estimated and in development.

Creative fatigue status

A delivery warning Meta triggers when cost per result reaches twice or more your past ads. The recommended response is a materially different image or video.

Optimisation event

The action Meta's system is asked to find people likely to take. Cheaper events are more plentiful and less predictive of revenue.

Learning phase

The period an ad set spends gathering enough optimisation events to stabilise. Meta's guidance treats roughly 50 events per ad set per week as the threshold.

Per-creative audiences

A capability opened to Marketing API partners on 19 October 2026: an age range and gender attached to individual creatives inside one ad set.

Cost per held appointment

Media spend divided by appointments attended, rather than by leads captured. The metric that predicts revenue in a considered purchase.