Switch It On in August. Freeze It in November.

August is when a shop has time and no urgency, which is exactly the wrong combination for making decisions about December. The reflex is to leave it. The problem with leaving it is that every automation you might introduce needs to be wrong a few times, in front of real customers, at low stakes, before you can trust it at high ones.

Three months is about right. Long enough to see the failure modes that only appear with volume, short enough that nobody has forgotten why it was configured that way. Counting back from the first serious week of demand, that lands in the month most merchants treat as dead.

Why does August decide December?

Because the season starts earlier than the calendar suggests, and because burn in takes longer than anyone budgets.

The scale is worth stating plainly. Consumers spent 257.8 billion dollars online between 1 November and 31 December 2025, up 6.8 percent year on year, with 25 separate days clearing 4 billion dollars against 18 the year before. Cyber Monday alone took 14.25 billion. The shape of that is not one spike but a plateau with spikes on it, which is harder to staff for than a single bad Monday.

And the ramp begins sooner. October 2025 online spending reached 88.7 billion dollars, up 8.2 percent, with holiday decorative goods already up 130 percent against September. If your plan starts on 1 November, it starts after a month of real seasonal demand has already been served, badly, by whatever you had running in September.

What should you switch on now?

Things that answer a question you already answer manually, in the same way, every time. Nothing that requires judgement, and nothing that touches money without a person.

The order status question is the obvious first candidate, because it is the highest volume enquiry in most shops and it has exactly one correct answer sitting in your systems. Automating it means connecting a lookup rather than writing clever text. Low stock alerts are second, because they change a decision rather than an interaction, and because discovering a bestseller went to zero on 3 December is the single most expensive thing that can quietly happen to you. Feed and listing monitoring is third: a daily check that your product count in the shopping channel matches your product count in the catalogue, which is a two line comparison and catches the failure that takes a week to notice otherwise.

Notice what is missing. Nothing on that list is a new tool with a new login. Peak season is not the moment to introduce vendors, because the failure mode of a new vendor is not a bug, it is a support ticket to somebody else's company on a Sunday.

Which support work is safe to automate?

The repeated question with a single correct answer, and nothing beyond it. The ranking is the same one we set out in choosing which tickets a bot should take first, and peak season sharpens it rather than changing it: volume rises fastest in exactly the categories where the answer is mechanical.

Set the boundary before the volume arrives, in writing. Automated replies handle where is my order, what is your returns window, do you ship to this country, and what size am I looking at. A person handles anything about a specific refund, anything where the customer is upset, and anything involving an exception to policy. The line is not about difficulty, it is about whether being wrong creates an obligation, which it does the moment an automated answer promises something you have to honour.

Staffing follows the same logic. If you bring in seasonal help, the training window is October, and what they need is not product knowledge but permission: what they may decide alone and what they must escalate. A temporary person who cannot approve a 12 euro refund generates more work than they absorb.

What should you freeze, and when?

Structural change, from about 1 November, with a written exception for content. The technique is standard practice outside retail too. GitLab's own tooling implements deploy freeze windows whose stated purpose is to prevent unintended production releases during a period you specify, configured as start and end times with a time zone, so the freeze is a rule the system enforces rather than an agreement people remember. If your setup allows that, configure it. If it does not, a note in the calendar and one person who says no is a workable substitute.

Be honest about the cost, because the freeze has one. Google's release engineering practice argues the opposite case for normal periods: frequent releases mean fewer changes between versions, which makes testing and troubleshooting simpler, and the same chapter emphasises canarying changes and rolling back features that misbehave rather than avoiding change. Both things are true. During a period when you cannot afford a bad hour, you accept a larger, riskier batch later in exchange for stability now.

Content is the category people get wrong in both directions. Freezing your ability to change a price, fix a typo in a product title or swap a sold out item off the home page turns a sensible policy into a straitjacket during the exact weeks when merchandising matters most. Keep those levers live. What you are freezing is anything that changes how the site works: templates, checkout steps, integrations, plugins, theme updates.

Why does AI referred traffic change the preparation?

Because it grew faster than any other channel and it rewards work you cannot do in November. Traffic to retail sites from generative AI tools rose 693.4 percent across the 2025 season, and 1,200 percent year on year in October alone.

The part that matters commercially is what those visitors did. Shoppers arriving from AI sources were 16 percent more likely to convert, generated 8 percent more revenue per session and stayed 44 percent longer. That is a channel behaving like a well qualified referral rather than like cheap traffic.

You cannot buy your way into it late. Whether an assistant can describe your product accurately depends on what your product pages actually say: sizes, materials, delivery windows, return terms, in text rather than in an image. That is an August job, and it is the same job that improves every other channel, which we went through in how AI shoppers read a product page. A catalogue cleaned up in August is still clean in December; a catalogue neglected until December stays neglected, because nobody rewrites 400 product pages during their busiest fortnight.

How much ecommerce automation is too much?

The limit is not technical, it is how many things one person can hold in their head while tired. Every automated step is a step somebody has to be able to explain, disable and reverse under pressure, and past about six of those a small team loses track of what is running.

A useful test: for each automation, can you name what it does, what it costs if it misfires, and how to stop it, without opening a document? Anything failing that test is either not worth keeping or not yet understood well enough to survive a busy week. Two well understood automations beat six configured hopefully.

There is a second limit worth respecting, which is coupling. An order status reply that reads your live catalogue depends on the catalogue being available; a stock alert that fires into a messaging tool depends on that tool. Each dependency is fine alone and the combination is where an ordinary outage becomes an interesting one. Draw the dependency chain on paper once. If any single failure takes out three automations at once, that is the one to give a manual fallback.

What breaks first, in practice

Feeds, then support volume, then stock, roughly in that order, and each has a tell you can watch for.

The product feed breaks quietly. A channel silently disapproves a portion of the catalogue after a price or availability format changes, and the first symptom is not an error but a gentle decline in a channel that was working. The tell is a count mismatch, which is why the daily comparison earns its place. Support volume breaks loudly and predictably, rising fastest in delivery related questions as shipping deadlines approach, which is precisely the category most amenable to a lookup. Stock breaks expensively, because the cost is not the lost sale but the cancellation email.

Conversion rate is the number that tells you something is wrong before you know what. A drop with steady traffic points at the site: a broken variant picker, a payment method failing silently, a shipping estimate that turned absurd for one country. A drop with rising traffic usually points at the traffic instead, which during peak often means a channel started sending browsers rather than buyers. Knowing which of those you are looking at takes one glance at sessions alongside orders, and it decides whether you spend the afternoon in your templates or in your ad account.

Keep a rollback path for every change you do make during the season, including the content ones. A price is easy to revert; a bulk edit that touched 300 products is not, unless you exported the previous values first. That export takes a minute and is the difference between an error you undo and an error you rediscover in January.

The one page runbook

Write it for a stranger at six in the morning, because on the worst day of the season the person holding it will be operating with roughly a stranger's clarity.

It needs four things. Contacts: your payment provider, your carrier, your hosting or platform support, with account numbers, because nobody can find those under pressure. Off switches: for every automation you switched on, the exact steps to disable it and revert to manual, tested once in October so you know the steps are right. Thresholds: the numbers that mean stop, such as orders per hour falling below a level you have seen, or a checkout error rate above one you have measured. And the decision rights: who may pause the shop, who may issue a refund above the usual limit, who may take the bot offline.

Test the off switches. This is the step everybody skips and it is the whole point of writing them down. An off switch that turns out to require a vendor login nobody has is not an off switch, it is a phone call you will make while orders fail.

What to watch while it runs

Three numbers, checked at a fixed time each day rather than stared at continuously. Orders per hour against the same hour last week, which catches a broken checkout faster than any error log. Support queue age, meaning how old the oldest unanswered message is, which is the honest measure of whether your support plan is holding. And out of stock rate on your top sellers, because that is the one problem you can still act on with a day's notice.

Mobile deserves a specific check rather than a general assumption. In the 2025 season 56.4 percent of online transactions happened on a smartphone, rising to 66.5 percent of sales on Christmas Day. Buy something from your own shop on your own phone, on mobile data rather than office wifi, in the week before peak. It takes ten minutes and it finds the problems your desktop testing never will. If the storefront itself is what keeps failing that test, that is a rebuild to schedule for January rather than a patch for November, and it is worth doing on a foundation that handles mobile checkout properly rather than another layer of fixes.

Unfreezing without a January outage

The backlog you accumulated is now the risk. Six weeks of deferred changes want to ship on the same afternoon, and shipping them together recreates exactly the situation the freeze was protecting you from, with the added disadvantage that nobody remembers what any of them do.

Sequence it deliberately. Ship one change, watch it for a day, ship the next. Start with the smallest and least connected. Anything that touches checkout goes last, in the quietest week you can find. If something breaks, you know which change caused it, which is the entire benefit of not shipping them as a batch.

Then do the part that pays for next year, while it is still fresh: write down what actually went wrong. Not a formal review, half a page. Which automation misfired and in what circumstance. Which question your bot could not answer. Which stock alert fired too late to act on. That half page, read in August, is worth more than any seasonal checklist somebody publishes, including this one, because it is about your shop rather than about shops in general. The stock thresholds in particular are worth revisiting against a full season of data, which is the point at which forecasting stops being theoretical, as we set out in what a forecast actually needs before it is worth trusting.

The uncomfortable summary

Most peak season failures are not capacity failures. They are changes made too late, by people with no time to watch what the change did. The retailers who have a calm December are rarely the ones with the best tooling; they are the ones who stopped changing things in time and knew how to switch each piece off.

That is an unglamorous conclusion for an article about automation, and it is the one the data keeps pointing at. Automate the repeated question in August. Measure it in September. Stop touching anything structural in November. Then spend the season doing the work no software does for you, which is deciding what to promise customers and keeping the promise.