Editing a live ad can reset its learning. What Meta and Google say about significant edits, and why new creative should arrive as new, paused ads.
Nubu Team··13 minute read
The new creative is rendered, approved and visibly better than the ad it is meant to replace. The live campaign is one tab away in Ads Manager, and the fastest route from A to B looks harmless: open the ad that is already running, swap in the new video, save. Same ad, same name, no extra rows in reporting, nothing new cluttering the account.
That instinct, the edit in place, is one of the most expensive tidy-ups in paid media. Both major ad platforms describe, in their own documentation, a learning process that sits underneath every ad you run, and both are clear that certain changes send that process back to the start. Changing the creative is on the list. The account looks cleaner for a day, and then the delivery system quietly begins its education again, on your budget.
This post makes the case for the opposite habit: new creative should arrive in your ad account as a new object, paused, and a human should flip the switch. It sticks to what Meta and Google actually publish rather than folklore, looks honestly at the one model where replacing content inside a live ad genuinely is the right call, and ends with a playbook you can apply whether or not you use Nubu.
Start with Meta, because Meta documents this most explicitly. Meta's own guidance describes an initial learning phase at the start of each campaign: the delivery system explores which audiences and placements work best for the ads in an ad set, and while that exploration runs, performance can be less stable. This is not a penalty or a probation. The system is spending your budget on finding out who responds to this specific ad, and its early guesses are necessarily its widest ones.
Exiting the phase is tied to results rather than to the calendar. The clearest number Meta publishes sits on its Learning Limited page: an ad set is flagged as learning limited when it is unlikely to receive around 50 optimisation events in the week after its last significant edit. That phrase, "last significant edit", matters and we will come back to it. Meta's guidance on cost stability points at the same yardstick from another angle: costs tend to settle once an ad set has generated roughly that many optimisation events since the last significant edit, and once an ad set has exited the learning phase, Meta says, delivery should stabilise.
Google's equivalent is the Google Ads learning period, and it attaches to a different layer of the account. Google's Help Centre explains that Smart Bidding needs a learning period in a handful of situations: when a bid strategy is newly created or reactivated, when a setting for the bid strategy is changed, and when campaigns, ad groups or keywords are added to or removed from the strategy. Calibration can take up to three weeks or one to two conversion cycles, though it is often faster when there is more conversion data to draw on. How long it lasts depends on how many conversions the campaign collects, how long a click typically takes to become a conversion, and which bid strategy is in play. The same page notes that the algorithms keep learning even after the visible Learning status clears, and that conversion history from previous campaigns can shorten the initial learning period.
The two systems differ in mechanics, but they agree on the shape of the thing. Performance on both platforms rests on accumulated evidence. The evidence is attached to specific objects and specific settings. Disturb those objects and you pay for the education twice.
What resets the learning phase: the significant edit lists#
Meta is unusually specific here, which is helpful, because it means we do not have to speculate. Ads Manager tracks the date of your last significant edit precisely because, in Meta's own words, that edit may restart the learning phase. And Meta's Help Centre defines the term: a significant edit is pausing your ad set, or making a change to the optimisation event, the audience or the creative. Changes to budget or bid strategy may also count, depending on how large they are. A separate Help Centre page on editing ads adds that editing certain details of a campaign, ad set or ad may restart the learning phase or send the ad back through the ad review system. Meta's stated recommendation follows directly: avoid making changes while a campaign is still in the learning phase.
Read that list again with the tempting shortcut in mind. "Just swapping the video" is not a small edit in Meta's taxonomy. The creative is one of the three named categories of significant edit, sitting alongside the audience and the optimisation event, the two levers everyone already treats as structural. There is a logic to that. Everything the delivery system has learnt about your ad is, in effect, an answer to the question "who responds to this creative?". Replace the creative and every one of those answers now describes an ad that no longer exists. From the system's point of view you have not refreshed an ad. You have discarded the evidence and kept the container.
Google's documented triggers are structural rather than creative: a new or reactivated bid strategy, a changed strategy setting, campaigns, ad groups or keywords added or removed. Google does not publish a per-ad significant edits list the way Meta does, which is itself a reason to phrase claims about Google carefully. What its Help Centre does establish is that learning attaches to the bidding configuration, that changes to that configuration cost calibration time measured in days to weeks, and that a freshly created object starts with no history of its own. However you cut it, the platform's confidence is earned per object and per configuration, and it does not transfer to whatever you swap in.
None of this is an argument against ever changing anything. Both platforms expect campaigns to evolve. It is an argument for knowing which changes are structural in the platform's eyes, and for making structural changes deliberately rather than as a side effect of shipping a file.
Advertising tools handle "the creative has changed" in one of two ways, and each is defensible in the right context. It is worth understanding both, because the wrong lesson from this post would be that updating a live ad is always a mistake.
The feed-driven model: replacing content in a live ad#
The first model updates the content of a live ad in place. It is how feed-driven and catalogue-driven creative works: the ad object stays live, and what it displays is regenerated when the underlying data changes. Smartly, probably the best-known platform built around this model, describes it plainly in its own documentation: you sync a data sheet or product feed, and the platform automatically creates, updates and archives ads at scale based on changes in the source, with a choice between applying feed changes automatically and reviewing them before publishing. Smartly's documentation also says its automation aims to minimise archiving ads and creating new ones when feed data changes, updating the existing ads where possible, and its marketing describes catalogue products appearing as video with prices and promotions updating automatically.
For the job that model was built for, it is the right call, and it deserves to be defended on its merits. A price is data, not a new creative idea. If a product's price drops, stock runs out, or a promotion ends at midnight, you want the live ad corrected immediately, not a new ad waiting for review while the old one advertises a price you can no longer honour. Tearing down and recreating thousands of catalogue ads every time a feed row changed would churn objects endlessly, and updating in place is the sane engineering answer for that workload.
But notice what the model implies, structurally. The content of a live object changes underneath whatever the delivery system has learnt about it. For a price field inside a dynamic template, that is precisely the point. For a finished piece of brand creative, it means something stranger: the ad your performance history describes and the ad now being served are different films wearing the same identity. Any conclusions you draw from that ad's numbers now straddle two different pieces of work, and Meta's significant edits list says the change to creative may have restarted the learning underneath as well.
The second model treats new creative as a new object. The finished ad is uploaded alongside the incumbent rather than into it, arrives paused, and is activated deliberately by a person. It costs you more rows in the account, and in exchange it buys four things.
The incumbent keeps its learning. Nothing about the running ad changed, so there is no significant edit, no learning phase reset, no unstable week for your best performer. Whatever evidence the platform has accumulated stays attached to the object that earned it.
The challenger gets a clean run. A new ad was always going to need its own learning. Shipping it as a new object means it enters that phase on its own terms, at a moment you choose, with the incumbent still delivering while it warms up.
The comparison stays honest. This is ordinary A/B discipline: two objects, two names, two sets of results, instead of one reporting row whose meaning silently changed halfway through the month. When someone asks which version worked, the account can actually answer.
The switch is a decision. Activation happens when a human decides the numbers, the timing and the budget justify it, not whenever the production pipeline happens to finish.
This is the right default for finished brand creative: new cuts, new concepts, seasonal refreshes, market and language variants. Which model a tool defaults to tells you what it thinks your creative is: rows in a feed, or finished work. Both answers are legitimate. They are just answers to different questions, and we compare the platforms behind them at more length in our piece on Smartly alternatives.
Nubu sits firmly in the second camp, and the reason we upload ads paused is everything above. Nubu is a creative advertising automation platform: it renders finished video and static variants at volume and delivers them to your ad accounts. Delivery is where automation gets dangerous, so the boundary is drawn hard.
On Meta, every rendered creative is pushed to your ad account as a paused ad. It is never activated by Nubu, and it never overwrites an existing ad: if the generated name already exists in the destination, the new ad arrives with a _V2 or _V3 suffix rather than replacing what is there. The pause is not a single checkbox somewhere. It is enforced independently at every stage between your approval and the ad account, so no single bug, misclick or prompt can turn "upload" into "launch".
On Google, the boundary is wider still. Nubu never creates campaigns, never touches budgets or bidding, never activates anything, and never modifies live objects. Demand Gen creatives land as paused ads inside the ad group you chose. Performance Max outputs are assembled into new paused asset groups rather than being edited into existing ones. And a retry after a network failure resumes where it stopped instead of duplicating ads in your account.
What lands in the account is therefore inert until someone with account access reviews it and turns it on, and that final step happening inside the ad platform is a feature, not a gap. It keeps three things in human hands. Activation timing: the ad goes live when the flight plan says so, not when a render queue drains at 2am. Budget context: the person enabling the ad is looking at the account's spend and pacing, which the production pipeline cannot see. And compliance: a paused ad can be checked by the people accountable for it before it serves; on Google, ads are policy checked as they are created, even though they arrive paused, and any policy finding comes back attached to that creative in Nubu, so problems show up while nothing is spending.
The same rule binds Nubu's AI assistant. It can prepare a Meta or Google upload for you, but it proposes it as an approval card that a human accepts before anything is sent, and everything it triggers lands in the ad account paused, exactly like a manual upload. It has no ability to activate an ad or spend a pound of budget. We wrote about why that boundary exists in AI proposes, humans approve.
Should you edit or duplicate an ad? A practical playbook#
Everything above compresses into five habits. None of them requires Nubu; all of them are easier with it.
1. When the change is the creative, duplicate rather than edit. Meta's significant edits list is categorical: creative changes belong on it. So reserve edit-in-place for the data-like parts of advertising, the feed-driven prices and corrected links, and treat anything that touches the creative, the audience or the optimisation event as a new chapter. If the new version was worth making, it is worth its own object. Duplicating the ad, or uploading the new version as its own ad, costs you a row. Editing the live one can cost you the evidence.
2. Let the incumbent run while the challenger learns. Activate the new ad alongside the old one instead of instead of it. The proven ad keeps delivering with its history intact while the new one goes through learning, and you judge the challenger only once it has had a fair run. Meta's own yardstick for stability is around 50 optimisation events in the week after the last significant edit, which gives you a concrete definition of "fair". Retire the incumbent when the challenger has earned it, not the moment the new file exists.
3. Batch changes, then leave the system alone. Every significant edit restarts the clock, and Meta's guidance is to avoid changes while learning is in progress. Three tweaks on three consecutive days is three restarts. If several changes are genuinely needed, make them together, then give the ad set a quiet week to re-earn stability.
4. Name versions so reporting stays legible. The honest cost of new-object discipline is more objects, and that is only a problem when the objects are indistinguishable. Version your names, and version your tracking. Nubu does both automatically: uploads are named from campaign, flow, market, language and version, collisions get the _V2/_V3 suffix rather than silently replacing anything, and each variant carries its own utm_content baked into the landing URL. Platform reporting and your analytics then agree about which version did what, with no archaeology.
5. Keep spend decisions where the spend is visible. Whatever generates and delivers your creative, let activation happen inside the ad platform, performed by a person who can see budgets, pacing and everything else in the account. A production pipeline that can flip ads live has quietly become a media buyer with none of the context. A pipeline that ships paused objects is a supplier, which is what it should be.
Three caveats, so this stays evidence-led rather than evangelical.
First, Nubu's paused-object delivery covers Meta and Google today. For every other channel, finished creative ships as a ZIP with a manifest, and your team uploads it through whatever process that platform requires. The philosophy travels; the automation, for now, does not.
Second, platform behaviour evolves. This post reflects what Meta's and Google's documentation said at the time of writing, in August 2026, and both companies revise how learning works and what counts as a significant change. Before you build process around any of it, read the current pages yourself: Meta's learning phase overview and significant edits page, and Google's learning period page.
Third, the paused, new-object model does not exempt anyone from learning. Every new ad has to earn its evidence, and there is no workflow that skips that. The choice on offer is narrower and more valuable: whether your proven ad keeps its history while the new one learns, and whether the moment of change is picked by a person or by a pipeline.
Ship the new version without breaking the old one#
If edit-in-place has become your team's habit, it is probably because producing and uploading versions by hand is tedious enough that overwriting feels like the only scalable option. That is a workflow problem, and it has a workflow answer: render every variant, deliver each one as a new, named, paused ad, and keep the switch in human hands. That is what Nubu's delivery features are built to do. If you want to see it against your own campaigns, create a workspace, connect a Meta or Google test account, and push a creative through. It will be waiting in your ad account, paused, exactly where you left the decision.
