Meta Andromeda 2.0 and Contractor Lead Ads: What Changes, and How to Adapt Your Creative

01What actually changed for creative

On 28 October 2026 Meta published the design of a rebuilt Andromeda, the retrieval stage of its ads system, on its engineering blog. Retrieval runs before ranking and before the auction: it narrows the pool of ads eligible for a given person down to a few thousand candidates in milliseconds. Meta says the new version scores every eligible ad at more than a billion ads per second on a single GPU server, roughly 100 times the throughput of the 2024 system, using a mixture-of-experts model with about 300 times fewer inference operations.

There is no setting to switch on, no opt-in and no new report. If you want the full teardown — the architecture, the numbers, and what Meta did not disclose about the 2.4% conversion claim — that is covered in our analysis of what the update means for window and door lead generation. This article is about the part a contractor actually controls: what goes in the creative. If you would rather have the account rebuilt than rebuild it yourself, that is what our Meta Ads agency does.

A note on terminology before we start, because this space is full of words that sound like synonyms and are not. A concept is one reason to buy, expressed in one visual world, for one kind of homeowner. An ad is one file — a video, an image, a carousel — that expresses a concept in one format. Iteration is refining an existing concept, and diversification is making a different one; Meta uses those two words in a specific way and so do we. A hook is the opening two to three seconds of a video or the first line of a static. Everything else is defined at the point it first appears, and the glossary collects the lot.

Three things about the announcement matter here, and only three.

That is the whole strategic premise. Everything below is operational: what to shoot, how to brief it, how to know when it is tired, and how to test it when you generate eleven leads a week rather than eleven thousand.

02Why contractor creative is homogeneous

Contractors produce near-identical advertising for a structural reason, not a lazy one. The work is visual proof: a before, an after, a truck, a crew, a price. Those are the assets that exist at the end of every job, and they are genuinely the assets that sell. The problem is that they sell identically for every company in a metro.

Here is the audit we run first in any contractor account. Every live ad gets sorted into a template, and the templates are remarkably consistent across trades.

Notice what is missing from the list: the failure, the diagnosis, the number, the objection, the neighbourhood, the material. Contractor advertising almost never shows the problem in close-up, because the camera comes out when the job is finished and the mess is gone. That single habit is why so much of this category looks the same — the only subject is the happy ending.

The pattern holds across the trades we run digital marketing for contractors in, from roofing to kitchen remodelling, which is why a single audit template works for all of them.

03Why volume no longer buys you options

The intuition behind producing thirty ads a month is that more tickets means more chances. Two independent findings say that is not what happens.

The first is Meta's own grouping behaviour from December 2025, described above: near-duplicates share their learning. The second is a spend-distribution pattern that shows up whenever anyone measures it. Both are about how the platform allocates budget, and both point the same way. In a 2026 analysis of roughly 3,000 Meta ad sets, the top one or two creatives captured between 70% and 90% of ad-set spend, averaging 87%, while more than 70% of creatives received under 5% of spend and never gathered enough data to be judged — the pattern practitioners call zombie creative, meaning assets that are live but were never actually tested. The same study reported that the concentration fell when accounts deliberately controlled creative counts and set budget floors per creative.

Put those together and the arithmetic changes. Thirty near-identical ads do not give you thirty tests. They give you one test, a lot of unsold inventory, and a report where 70% of the rows are noise. Ten distinct concepts give you ten tests, of which the system will still concentrate spend on two or three — but the two or three it picks tell you something you did not already know.

This is not a theoretical concern. It is the first thing we look for when a contractor says paid social has stopped working: an account with thirty live ads that turns out to contain three ideas. You can see the same dynamic worked through in our interlock paving Meta Ads case study, where the account structure and the creative count were changed together rather than separately.

This is why the practical unit of contractor creative planning is a concept: a single reason to buy, expressed in one visual world, for one type of homeowner. Formatting, aspect ratio, headline variants and copy length are execution details inside a concept. They are worth doing, and they are not new bets.

04Video versus static: what the data really says

Every contractor article on this subject says video wins. The truth is more useful: video wins a specific job, static wins a different job, and the two datasets that get quoted at each other are not actually contradicting one another.

The reconciliation is straightforward. Video earns attention and does the cold-audience prospecting — it interrupts a scroller and holds them. Static closes more cheaply once someone is already interested, and it fatigues faster. Contractor accounts usually sit at the prospecting end of that spectrum, which is why the contractor-specific data leans so heavily toward video: these campaigns are mostly cold, local and interruption-based.

05The concept grid

A concept is not a topic. "Roofing" is a topic; "your neighbour's roof is eleven years old and you cannot see the decking" is a concept. The grid below is how we generate them without repeating ourselves, and it works across trades because it iterates on why now rather than on product.

Five concept families work across every trade we handle. They are not product categories; they are the moment the homeowner crosses from noticing to considering.

The fifth family is the one almost nobody runs, and it is the most valuable in a considered purchase. "Here is when you should not buy this" is the only claim in the category that sounds like advice rather than advertising, it is unrepeatable by a competitor who only has offer creative, and it produces the highest-intent form fills we see in contractor accounts.

Which concept to start with depends on the trade. HVAC accounts usually get the fastest read from the repair-versus-replace threshold, because the decision is forced by a failure; roofing accounts get it from the storm-damage diagnosis. If you want trade-specific benchmarks before choosing, the HVAC contractor marketing and roofing marketing pages set out what each looks like in practice.

06The 20-minute job-site capture

The reason contractors buy agency video is not that they cannot hold a phone. It is that nobody has told the person on site what to capture, so what comes back is a wide shot of a finished house. A shot list fixes that in a single visit, and it costs one installer twenty minutes.

Hand this to whoever is on the job. It is deliberately written for someone with gloves on.

Eight items, twenty minutes, and it produces the raw material for three or four concepts — enough to keep a bench of approved-then-held assets, meaning concepts that are shot, edited and approved but deliberately not launched yet: the failure, the diagnosis, the number (overlay), the neighbour, and often the objection. Compare that with a monthly agency shoot that produces one polished concept and eleven variations of it.

Two practical rules keep this from degrading. First, do not let the same person both shoot and edit, because the person who was there will edit out the messy detail that makes the footage work. Second, keep a subject log — one line per capture: trade, failure type, city, season, product. Without it you will end up with nine versions of the same rusted pan and no way to see that you have.

Contractors who run this well treat the capture as part of internet marketing for contractors rather than as a marketing favour — the crew is producing the only asset the category cannot copy, which is footage of real work at a real address.

07The one-page brief

Most contractor creative problems are briefing problems wearing a production costume. A brief that says "before/after video, 30 seconds, upbeat" guarantees the template. A brief that states the claim cannot.

Five fields. If a creative cannot be described in all five, it is not a concept yet.

One more field worth adding to the template as a checkbox rather than a paragraph: what does this concept look like in the first frame? Because that is what Meta's Creative diversity column compares — the images and video thumbnails in an ad set — and because the opening frame is also the thing that decides whether a scroller stops. If two concepts open on the same kind of image, revise one before you spend money on it.

08Hooks, and the first three seconds

Hook work and concept work are different activities, and conflating them is why accounts end up with twenty openings and one idea. The concept is what the ad argues; the hook is how it earns the first three seconds in which to argue it.

The published contractor data puts a lot of weight on that window. Roofing creative analysis reports that vertical 9:16 before/after performs 2–4x the click-through of produced 16:9 horizontal, with the first three seconds carrying around 80% of engagement. Practitioner fatigue data is consistent with that: hooks fatigue in 2–3 weeks, visuals in 4–6, formats in 6–8, offers in 6–10, and the core message only after 8–12.

That last table is the most useful operational fact in this article. It means a good concept is not one ad — it is a platform for four to six weeks of experimentation, in this order: swap the hook, then the visual treatment, then the format, and only then consider a new concept.

If you are deciding how much of the budget belongs in paid social at all before optimising hooks, the comparison in the best form of advertising for contractors sets paid social against search, local services and referral.

Two cautions. Hooks that overstate get leads that over-promise — a "you might not need a new roof" hook has to be followed by a real diagnosis, not a soft close. And hook variation is not diversification: Meta groups ads with similar visuals and themes, so four hooks on the same footage is one concept with four openings. That is a legitimate refresh strategy, not a reason to skip new concepts.

09Format, placement and the carousel trap

Formats are where contractor accounts quietly lose money, because the format that photographs the work best is the one the reporting cannot evaluate. Meta documents the format-level behaviour of its own ad products in its Business Help Center, which is worth reading directly rather than through an agency summary.

On 27 October 2026, a day before the Andromeda post, Meta was reported to have added individual performance data for 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.

The mechanical decisions underneath all of this are unglamorous and worth getting right once: one ad set should carry your formats rather than one ad per format; a single ad can hold up to 10 creative assets, which is where placement variations belong; and ad sets cap at 50 ads, with Meta's own ad-volume guidance warning that too many ads can hurt performance. The goal is more concepts, not more objects.

For contractors whose paid social sits inside a wider residential builder or homebuilder programme, the format mix usually has to reconcile with a longer sales cycle than a repair trade; our paid social for residential homebuilders work covers that case separately.

10Fatigue signals and the refresh calendar

Meta gives you two official signals. The Creative fatigue status appears in the delivery column when cost per result reaches twice or more your past ads, and Meta's recommendation is a new image or video that is materially different from the original. The Creative diversity column labels an ad set High, Medium or Low by comparing its images and video thumbnails, and Meta marks it as estimated and in development.

Both are lagging. By the time a fatigue flag fires you have already spent two to three weeks on a declining asset, which is why practitioners watch upstream signals instead.

Here are the thresholds we actually use, stated as numbers so they can be argued with. Frequency (average impressions per person in a seven-day window) is a watch signal above 2.5 and an act signal above 3.5 sustained for a week. Click-through is a watch signal at 10–14% below that asset's own baseline and an act signal at 15–20% held for five days. Cost per result is a watch signal at 15–24% above the asset's baseline and an act signal at 25% or more held for three days. These are practitioner conventions, not Meta's published thresholds — Meta's own only trigger at twice the cost per result.

Contractor accounts fatigue faster than the general benchmarks imply, for a structural reason: the addressable audience is small. A contractor serving a thirty-mile radius may have only tens of thousands of targetable homeowners, so the same person sees the same ad far sooner than in a national e-commerce account. Regional agency data for HVAC puts it plainly — after three to four weeks on the same creative, frequency climbs past 3.0 and cost per lead rises 25–40%. That is a compressed calendar, and it is why the production rhythm matters more than the individual ad.

Three signals are worth watching weekly, in this order, because that is the order they move in: spend concentration (is 90% of the budget on one asset?), frequency (is the same person seeing it more than three times a week?), and click-through against the asset's own baseline (not against the account average). Cost per lead is the fourth and last, by which point you are reacting.

11Testing design at contractor volumes

Contractors cannot run the testing programmes the creative literature describes, because lead volume is too low. A dealer generating ninety leads a month cannot detect a 10% difference between two concepts — the sample is nowhere near. Pretending otherwise produces confident conclusions from coin flips, and it is the most common analytical error we find in contractor accounts.

Run the arithmetic once and the constraint becomes obvious. At the $42.95 median Meta cost per lead for home improvement in the WordStream and LocaliQ 2026 benchmarks, a contractor spending $4,500 a month produces roughly 105 leads a month — about 24 a week. Split across three ad sets and no single set is anywhere near the roughly 50 optimisation events per week that Meta's guidance uses as the learning-phase threshold. That is arithmetic on a published benchmark, not a forecast, and the benchmark is US data; but the structural conclusion holds in any market: at contractor budgets, statistical confidence arrives over months, not days.

Three adjustments make testing honest at this scale.

One structural point that overrides all three: because Meta concentrates spend on one or two creatives, the honest question is never "which ad won" but "did the rotation contain a concept capable of winning". That is a different report, and it is the one to bring to a quarterly review.

If the volume problem is the binding constraint rather than the creative, the fix is usually channel mix rather than more concepts — which is the subject of our contractor lead generation services work.

12What AI should and should not do

Generative tools are the reason Meta rebuilt retrieval, and they are also the fastest way to make your account more homogeneous. Both things are true.

The same distinction governs how AI is changing the rest of a contractor's marketing, which we cover separately in AI marketing for contractors: the useful applications compress real work, and the useless ones manufacture volume.

The first list is execution. The second list is either fabrication or the exact behaviour Meta's December 2025 guidance warns about — creative so similar its system cannot tell it apart, which produces shared learning and a single budget of attention rather than twenty bets.

There is also a credibility cost that never appears in a dashboard. Contractor advertising runs on the claim that a real person did real work at a real address. An AI-generated before/after is detectable by the audience it is aimed at, and it is indefensible if a homeowner asks. The line we hold in accounts we manage: AI may compress, reformat, caption and translate real footage; it may not manufacture the evidence.

13The creative operations calendar

Creative operations is a calendar problem before it is a creative one. Every trade has two or three windows a year when demand is structurally high, and a concept produced the week after that window is worth a fraction of the same concept produced before it.

Two operating rules follow. First, shoot in the off-season — the material for a January bath campaign is captured during October demolition work, because you cannot produce it in the week you need it. Second, hold a bench: two approved concepts per ad set, ready to launch, not yet launched. When a fatigue flag fires, a bench of two converts a two-week problem into a two-hour one.

Two service-line considerations sit on top of this calendar. Remodelling work has a different seasonality and a longer consideration window than repair trades, which is why we run it through a separate remodeling contractor marketing programme. And the calendar itself should agree with the annual plan rather than contradict it — the wider version is set out in contractor marketing strategies for 2026.

The rhythm that survives contact with a busy contractor is monthly, not weekly: one capture day per month across the crews, one editing session, three concepts entering rotation, one retiring. Contractors who try to run a weekly creative sprint stop doing it within a quarter, and the account goes back to one tired concept by spring.

14Five failure modes

• Reporting ad count. "We shipped twenty-two creatives this month" is a report about effort. Twenty-two ads inside three concepts is three tests. Report concept count, and report the group each ad belongs to.
• Shooting the ending. The camera comes out when the work is finished, so the only subject is a completed job — which is the same subject every competitor has. The failure is the differentiator and it is only there at the start.
• Buying video, not concepts. A production budget spent on one polished explainer yields one candidate. The same money spread across four rough captures yields four candidates and a bench.
• Testing in the learning phase. Below roughly 50 optimisation events per ad set per week, a comparison is measuring noise. Contractor volumes require patience or a leading metric, usually both.
• Ending every ad in "free quote". Static "free quote" creative is reported as the most expensive category per lead in contractor accounts, and the reason is structural: it asks for the biggest commitment at the coldest moment. A photo-based diagnosis is a smaller ask that qualifies harder.

There is a sixth, subtler one worth naming because it is fashionable right now: attributing every change to Andromeda. Meta published an architecture post with no account-level rollout date and no way for an advertiser to observe the switch. When a contractor's costs move, the cause is usually creative age, seasonality, auction density or a structural change somebody made — and the diagnosis matters, because the remedies are different.

15The 30-day creative rebuild

If you are running this rebuild inside a larger account and city-level structure, the Toronto market version of the same work is documented in Meta Ads management in Toronto, and the equivalent for kitchens and baths in our One Stop Remodeling Meta Ads case study.

What you should expect at the end of it is not a dramatic cost-per-lead improvement. It is something more useful: a report where the rows mean different things, a production rhythm the crews can sustain, and a bench that turns a fatigue flag from an emergency into a scheduled swap. The improvement follows in the second quarter, when the concepts that survived have had enough volume to be judged.

16Sources and confidence

Every figure above traces to one of the following. The creative-performance numbers in this space come almost entirely from agencies and vendors with a commercial interest, so they are labelled as such rather than presented as research findings.

How this page is built for machine reading. The structure is deliberate rather than decorative: a summary block above the first section, one idea per paragraph, an explicit question-and-answer module, defined terms at the point of first use, and structured data for the article, the FAQ, the glossary and the procedure. Heading order and semantic elements follow the W3C HTML and accessibility guidance; the vocabulary used in the markup is the public schema.org standard; and every diagram on the page is inline SVG rather than a late-loaded image, so the page carries no external requests and no layout shift, consistent with the Core Web Vitals guidance.

Four caveats matter more than any single figure here. The creative-performance numbers are dominated by agencies publishing their own results, which is not a reason to dismiss them but is a reason not to treat a 30–60% saving as a budget you can book. The two largest datasets on format conflict on cost per acquisition, and the conflict is best explained by job rather than by winner. Almost all of this data is e-commerce-weighted or US-weighted, and a local contractor audience of tens of thousands behaves differently from a national one. And nothing published anywhere tests the specific chain this article proposes — that a distinct concept produces a cheaper held appointment — so that remains a hypothesis to run in your own account rather than a finding to cite.

17Questions we get asked

18Glossary

Andromeda

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

Retrieval

The stage that narrows eligible ads to a few thousand candidates before ranking and the auction. An ad that is not retrieved never has its bid read.

Concept

One reason to buy, expressed in one visual world, for one type of customer. The unit of creative planning; distinct from an ad or a format.

Iteration

Refining an existing concept — new headline, new opening frame, new aspect ratio. Meta treats these as variations rather than new creative.

Diversification

Meta's term for distinct assets built for different personas or use cases, as opposed to iteration.

Creative diversity label

An Ads Manager column comparing an ad set's images and video thumbnails and labelling it High, Medium or Low. Estimated and in development, per Meta.

Creative fatigue status

A delivery warning Meta triggers when cost per result reaches twice or more past ads, resolved by a materially different asset.

Spend concentration

The share of ad-set budget landing on the top one or two creatives. The earliest practical signal that the rotation has narrowed.

Hook

The first two to three seconds of a video or the opening line of a static. Fatigues fastest — practitioner data puts it at two to three weeks.

Zombie creative

An ad that receives so little spend it never gathers evaluable data. Reported at over 70% of creatives in one 2026 analysis of ad sets.

Bench

Approved concepts held ready but not launched, so a fatigue flag can be answered in hours rather than weeks.

Learning phase

The period an ad set spends gathering enough optimisation events to stabilise, around 50 per ad set per week per Meta's guidance.

Hook rate

Three-second video views as a share of impressions. A leading indicator of creative health that is readable at low spend.

CTR (click-through rate)

Clicks divided by impressions. Moves before cost per result does, which is why it is used as an early signal.

CPL (cost per lead)

Media spend divided by enquiries captured. Cheap and plentiful, and the least predictive metric in a considered purchase.

CPA (cost per acquisition)

Media spend divided by completed actions. A stricter measure than cost per lead when the action is defined as something worth having.

Frequency

Average number of times one person saw an ad in a window, usually seven days. The first fatigue signal that moves.

Learning phase

The period an ad set spends gathering enough optimisation events to deliver stably.

Optimisation event

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

HVAC

Heating, ventilation and air conditioning — used here as the trade label, not as a technical description.