Running a small website with AI, day to day

A website is a project for about a week and an obligation for about a decade.

The building part gets all the attention, and AI tools have made it faster than ever. Then the site is live and nothing more is said about it, which is the part that actually goes wrong. Prices change and the page still shows last year's. A plugin update waits four months. A mailbox stops collecting enquiries and nobody notices until a customer phones to complain that nobody answered their email.

After launch, a website needs copy, images, updates, backups, email, speed checks and facts kept current, and an AI assistant can draft or run most of these with a person approving. That last clause is the whole article. An assistant is very good at the looking up, the drafting and the proposing. It does not decide what goes live on your business's website. The useful question is not whether to use one but where your own judgement goes in the loop.

What follows is the running part, job by job, with the point where you stop and read.

What a site needs every month

Strip away the advice industry and the list is short. Once a month somebody needs to:

• Apply the updates that are waiting, on the platform, the plugins and the theme.
• Confirm a backup exists and is recent, rather than assuming it.
• Check that the contact form, the mailboxes and the forwarders still deliver.
• Look at whether anything got slower, and why.
• Correct the business facts that drifted: hours, prices, the phone number, the service you stopped offering in March.
• Read enough of the traffic to notice a page that died or one that took off.
• Publish or refresh something, so the site is not frozen at its launch date.

Seven items, about an hour on a quiet month. It does not happen because it is seven logins and nobody's actual job. The knowledgebase guide to monthly WordPress maintenance tasks walks the platform half properly, including what most people skip: delete the plugins you stopped using instead of deactivating them, because a deactivated plugin with a known vulnerability is still on disk.

An assistant changes the economics of that hour. It holds all seven in one conversation, fetches the state of each, drafts what needs drafting and proposes what needs changing, which turns seven logins into one session where you mostly read and say yes or no. The exact monthly sequence with the prompts in order is a piece of its own . If WordPress is most of your site, AI and WordPress covers the plugin and theme half in depth.

The line between drafting and approving

Everything here sits on one distinction, worth stating plainly before any of the jobs.

An assistant may draft. A person approves.

Drafting is where being wrong costs nothing, because nothing has shipped. Three versions of a page's copy, a slow-page report, the list of waiting updates, a first pass at a returns policy: get any of it wrong and the only loss is the minute it took to read.

Approving is where being wrong costs something real. Publishing the copy. Running the update. Changing the DNS record. Not harder, just the ones with consequences, and the ones a person signs.

For hosting work that line is enforced rather than suggested. Every change an assistant makes through the Flashcloud AI Connector is a two-step confirmed action: the assistant proposes exactly what will happen and nothing runs until the customer approves it. The tool call does not perform the change, it returns a proposal with the exact record, path or value, plus a confirmation token. You read it, you say yes, and then it runs. Say no and the token dies. The full tool list, and the things that have no tool at all, are in AI agents and your hosting account.

For copy, images and legal text there is no such gate, which is why you have to build one yourself. Nothing stops an assistant writing a claim about your business that is not true, and nothing stops you pasting it onto the site. The approval step there is a habit, not a feature.

One safety net before you start approving things: Flashcloud keeps daily backups with 30-day retention, so anything an assistant changes can be put back. Our backup options covers what the daily backup includes, which is files, databases, email and DNS, and how a restore runs from the portal. A bad approval becomes a bad afternoon instead of a bad quarter.

Copy, images and legal pages

This is what most owners reach for an assistant first, and where the output needs the most editing. Ask for website copy and you get fluent, generic prose full of adjectives no reader believes and no customer can check. "Industry-leading service." "Passionate about quality." It reads like a website and says nothing.

The fix is to hand the assistant the house rules a copywriter would work to. These are ours, and they transfer to any small business:

• Facts, not adjectives. "We answer the phone until 7pm" beats "exceptional customer service". A fact can be checked, which is why it persuades.
• A banned-word list. Write down the fog words your industry overuses, the world-class and best-in-class family, the ones about unlocking potential. Every competitor uses them, so they carry no information. Paste the list into the prompt and the assistant stops reaching for them.
• Every claim carries a source. A number needs an origin, a score needs the link to the review, a comparison needs the date you checked it. No source, cut the claim.
• Say the mechanics. Not "fast turnaround" but "most jobs back within two working days". The mechanic is what people remember.
• Never invent capability. An assistant that does not know your delivery radius will produce a plausible one, and that plausible number is the most dangerous thing in the draft.

Give it those five and the draft arrives usable. Then read it as the person who has to honour it. That is the approval step, and it is the whole job. Prompting for copy that stays true to the business is a longer piece .

Images add one question, whether you may use the thing at all. Generated images are fine for atmosphere and bad for anything a customer takes as evidence: no photo of premises you do not have or a team that does not exist. Then check the file size, because an unresized generated image is routinely two or three megabytes and costs more page speed than any caching plugin wins back. Licensing, honesty and file size get their own article .

Legal pages are the sharpest case. A privacy policy drafted by an assistant is a good structure and a bad finished document, because it describes a generic business rather than yours. The checks that matter: the data you actually collect is listed, the processors you use are named, the jurisdiction is right, the contact route works. The knowledgebase covers adding a privacy policy page mechanically; what to check in an AI-drafted one is a separate piece .

Keeping the business facts consistent

The maintenance job nobody schedules and everybody fails. Your opening hours exist in six places: the site footer, the contact page, the Google Business Profile, the Facebook page, the directory listing somebody made in 2019, and the PDF menu. You change them in one.

That mattered moderately when only humans read websites. It matters more now, because an assistant answering "what time do they close" cannot resolve a conflict between your footer and your profile. It picks one, and not the one you would.

An assistant is unusually good here, because consistency checking is pattern work rather than judgement. Ask it to list every place your phone number, address, hours and prices appear, then to compare them, and the mismatches arrive in a minute. The approval step is trivial: you know the correct value.

The visibility side, how assistants crawl, read and cite your site at all, is a subject of its own: how AI assistants find, read and cite your website covers the mechanics, and checking what AI assistants say about your business is the right first measurement. Keeping the facts consistent across site and web is its own routine , as is what AI helps and hurts in search .

Hosting tasks through the connector

Up to here the assistant has been drafting. This is where it starts doing, and where the confirmed-action model earns its keep.

The Flashcloud AI Connector is an MCP server at https://mcp.flashcloud.com/mcp that Claude, ChatGPT, Gemini, Perplexity, Cursor and any MCP-compatible client connect to over OAuth 2.1 with a Flashcloud login. Paste that address into your assistant's custom connector screen and sign in on our page, so the assistant holds a revocable token rather than your password. Read access is scoped to the categories you approve; actions are a separate permission, off until you turn it on.

Four jobs a small site needs monthly.

Updates. Ask for the state of the WordPress install and you get core, plugin and theme versions read from the account rather than guessed. Then a staging copy, then the updates on staging, which is the order people skip by hand. Each step is a separate proposal. WordPress updates and staging through your AI assistant shows the phrasing.

DNS. Adding the TXT record a third-party service just handed you is the best first confirmed action, because the proposal shows the record type, name, value and TTL and you can check it character by character before it is written. Nameservers are out of scope: that change moves your whole domain and lives in the portal.

Mailboxes. "Create a mailbox for sales@mydomain.com" is the entire instruction. The assistant reads which domains are on the account, checks the address is valid and proposes the creation. Creating mailboxes through your AI assistant makes a point worth repeating: assistants read a lot of untrusted content, including web pages, tickets and DNS records, so the only approval that counts is the one you type into your own conversation.

Cache purge. The task you currently open two tabs for. "Purge the Cloudflare cache for example.com" is one sentence and one approval, and the assistant can check whether caching is on for that domain first. A purge clears the edge only: on WordPress, LiteSpeed Cache and the object cache are separate layers closer to the app, and your own browser is a third. Purging the Cloudflare cache through your AI assistant explains the ordering when a page stays stale.

Some things have no tool, by design. Payments, card details, passwords, two-factor settings, contact details, cancellations and domain transfers can only happen in the Flashcloud portal, by the customer, never through an AI assistant. No permission reaches them by accident and no prompt talks an assistant into them. If one tells you it paid an invoice, it did not.

Speed and logs

Speed is the job owners most want to delegate and understand least, which makes it a good test of the split.

An assistant is genuinely useful at diagnosis. Point it at the page and it reads the waterfall, names the heaviest requests, spots the render-blocking script and tells you the hero image is a 2.4 MB PNG that should be a 180 KB WebP. That is reading, where it is strongest. Our guide to why a website is slow covers the same ground from the server side, and the connector can pull site health and the WordPress debug log into the conversation, so the assistant works from your logs rather than general advice.

The approving half needs care, because speed fixes break things invisibly. Minifying and combining CSS and JavaScript can take a layout apart. Lazy-loading a hero image makes the headline number worse. Deferring a script breaks the booking widget that depended on load order. None of it shows on the page you are looking at. It shows on the checkout, on mobile, next week.

So let the assistant produce the list, ranked by what each fix wins back, then apply them one at a time with a look at the live site after each. Boring, and it is the difference between a faster site and a site that is faster and broken.

For traffic, the portal's service page carries visitor traffic over time, top pages, bandwidth and storage, and cPanel adds Awstats and the raw access logs under Metrics. Viewing your website statistics explains what confuses everybody first time: server-side stats count every request including bots, while JavaScript analytics count only humans with JavaScript on, which is why the two numbers never agree and why both are useful. Reading those logs with an assistant, and asking why a page got slow, are articles of their own .

Email and customer chat

Email is the quiet failure. A contact form that stops delivering looks exactly like a slow week in sales, and a small site can run for a month that way without knowing.

The monthly check takes a minute: send a message through your own contact form and confirm it arrives in the mailbox you expect, and that any forwarder still points where you think. An assistant can list the mailboxes and forwarders on the account so you check against reality instead of memory, and read mail headers with you when a message goes missing. Deliverability is mostly DNS, and why emails go to spam covers the SPF, DKIM and DMARC records that decide it. Running the whole email side through an assistant gets its own piece .

Customer chat is the one job here where the honest answer is mostly no, at least for a small site. A widget trained on your pages answers the five questions already in your FAQ, a small win, and confidently invents an answer to the sixth, a real loss when the sixth is about a refund or a booking or a price. If you run one, scope it hard: answer from the site's own content, say "I will pass this to a person" at the first sign of a transaction, route to a human fast. We take the same view of our own queue, which is why Flashcloud support is answered by people, not by an AI chatbot first. What to expect from AI chat on a small site is written up separately .

The monthly routine

The whole thing as a sequence, approval points marked. One session, one assistant, in this order.

1. Ask for the state of everything. "Read my hosting account and tell me what is pending: core, plugin and theme updates, SSL status, disk and bandwidth use, and anything in the last 30 days of traffic that moved a lot." Reading only, nothing to approve. The agenda arrives in one message.
2. Confirm the backup before you touch anything. Check that a recent one exists and looks complete rather than trusting that the schedule ran. If you have never restored one, the guide to testing your website backups is the right afternoon to spend, because an untested backup is a hope. Approve: nothing. Verify: yes.
3. Staging, then updates. A staging copy, the updates on staging, a look at the staged site, then the same updates on production. Approve: each action separately, after reading the proposal.
4. Fix the facts. Give the assistant the current hours, prices and contact details and ask it to find every page that disagrees. Approve: each edit, because you know the right answer.
5. One speed fix. The top item from the ranked list, applied, then a look at the live site. One a month compounds. Approve: the change, then check the page.
6. Test the contact route. Form to mailbox, end to end, by actually sending one. Approve: nothing. A test, not a change.
7. Publish something. A page refresh, a short post, an updated service description. Draft against the house rules above, then read it as the person who has to honour it. Approve: the final text, after reading every claim.
8. Purge the cache and look at the site. One sentence, one approval, then open the page in a browser that has never seen it. Approve: the purge.

Under an hour for a small site, and most of that hour is reading. Not an accident of this list: it is what the work becomes once the looking up and drafting are handled, because your time goes to deciding instead of navigating.

Two notes on where the assistant is not the answer. Anything structural, a redesign, an accessibility review, a store that handles money in an unusual way, is a person's job, and routine operations running smoothly is what buys the attention for it. And if a proposal arrives that you cannot check, do not approve it. Ask the assistant what a visitor would see differently, or open a ticket and ask us. Hosting questions reach a person here, which is the part that has not changed.

None of this needs a plan change. The AI Connector comes with every Flashcloud plan and support is still answered by people, so the hour above starts with the two-minute connection and a first task that approves nothing: ask it what the account looks like today.