Human-in-the-loop website changes: why every change should be approved

Three levels of automation

The model you choose can differ by type of change. What matters is that you choose it deliberately, and that you know exactly what any tool will publish without a person reading it first.

Why every content change should be approved

Your company owns what the website says

In Moffatt v. Air Canada (14 February 2024), a Canadian tribunal held the airline responsible for incorrect information given by the chatbot on its own website, and rejected the argument that the chatbot was responsible for its own statements¹. It was a small claims decision in another country, but the principle is intuitive: visitors treat everything on your site as your word, whoever or whatever produced it.

Advertising rules apply to your own site

In the UK, the CAP Code has covered marketers' own marketing communications on their own websites since 1 March 2011². It says marketing communications must not materially mislead, and that marketers must hold documentary evidence for objective claims before publishing them³. A tool that rewrites "we support 40 integrations" as "we integrate with everything" has made a claim you now have to defend.

Search engines care about purpose, not authorship

Google says appropriate use of AI or automation is not against its guidelines, but using it mainly to manipulate search rankings is a spam violation⁴. Its guidance asks creators to think about who made content, how and why, and to make AI involvement clear where readers might wonder⁵. Its spam policies name generating many pages without adding value for users as scaled content abuse⁶. A review step is where someone checks that a change helps readers, not just a metric.

Language models get facts wrong

Drafting tools can produce fluent sentences that are untrue: a feature you do not have, a client you never worked with, a certification you do not hold. These errors are easy to catch when a person who knows the business reads the change, and easy to miss when nobody does.

What a good approval step shows the reviewer

Approval only works if the reviewer can decide in a couple of minutes. A good proposal puts everything needed on one screen:

• The exact change: current and proposed text side by side, on the actual page.
• The evidence: the signal that prompted it, such as "1,900 impressions and 11 clicks in 28 days" (illustrative), with a link to the data.
• The reasoning: one or two sentences on why this change should help.
• The checks: length, uniqueness, links, accessibility basics and any claims flagged for substantiation.
• The risk level: low (a title on a blog post), medium (services page copy), high (pricing, legal, comparisons).
• The way back: confirmation that the previous version is stored and how rollback works.
• The measurement plan: what will be measured, for how long, and what counts as a result.

Approve, edit or reject: how to decide

Five questions cover most decisions:

1. Is every statement true, and could we show evidence for it?
2. Does it sound like us, and would we say it to a client in a meeting?
3. Does it make the page more useful to the buyer reading it?
4. Is the evidence behind it strong enough to justify the change?
5. Is now the right time, given other changes on this page?

Yes to all five: approve. Right idea, wrong words: edit, and keep the edit, because the edit records your voice for the next proposal. Any firm no: reject, and write down why. A recorded rejection ("we do not name competitors on services pages") stops the same proposal coming back.

Match review depth to risk

Keeping approval fast

The usual objection to human review is that it slows everything down. It does not have to.

• Batch low-risk changes into one weekly review rather than interrupting someone every day.
• Keep proposals small. One change is quick to judge; ten changes in one proposal are not.
• Set a review slot in the calendar, so the queue never grows stale.
• Record why edits and rejections were made, so proposals get closer to approvable over time.
• Automate everything around the decision: finding, drafting, checking, publishing, rollback and measurement.

This is how approval fits into a continuous website optimisation routine without becoming the bottleneck.

Keep an audit trail

For each change, record:

• The page, the previous version and the new version.
• Who approved, edited or rejected it, when, and the reason given.
• When it was published, and whether it was rolled back.
• The measurement window and the outcome, including inconclusive results.

The trail answers the questions that come later: why did this page change, who agreed to that claim, and did it make any difference? It also separates three things that are easy to blur: a decision, a deployment and a result.

Rollback is part of the approval

Approving a change should never be a one-way door. Keep the previous version of every change, restore it in one step, and roll back automatically if the page breaks after publishing. Knowing the way back makes reviewers more willing to approve sensible changes quickly.

Where Artiweave stands

Artiweave is approval-first. It proposes specific changes from crawl, Search Console and performance evidence, shows the reasoning and checks for each, and publishes nothing until a marketer approves it. Approvals are an explicit step with a recorded reason, published changes roll back automatically if something breaks, and outcomes are reported honestly. Read more about the approach in our guide to AI website optimisation, or start with how to improve a B2B website without a redesign.

Frequently asked questions

What does human in the loop mean for website changes?

It means software can find opportunities and draft changes, but a named person approves, edits or rejects each change before it is published, and the decision is recorded with its reason.

Is it ever fine to let AI publish website changes automatically?

For low-stakes, easily reversed technical work such as compressing images, automatic publishing with monitoring can be reasonable. For copy and claims under your brand, a person should approve each change, because your company is responsible for what the website says.

Does human review slow optimisation down?

Not much, if the review is designed well. Small proposals, batched low-risk changes and a fixed weekly slot keep it to minutes, while software handles finding, drafting, checking, publishing and measuring.

Do UK advertising rules apply to my company website?

Yes. Since 1 March 2011 the CAP Code has covered marketers' own marketing communications on their own websites, including the requirement to hold evidence for objective claims before publishing them.

Sources

1. Moffatt v. Air Canada, 2024 BCCRT 149, Civil Resolution Tribunal of British Columbia (14 February 2024), on CanLII
2. Advertising Standards Authority: Landmark agreement extends our digital remit (1 September 2010; in force 1 March 2011)
3. CAP Code (UK Code of Non-broadcast Advertising), section 3: Misleading advertising, rules 3.1 and 3.7 (accessed November 2026)
4. Google Search Central Blog: Google Search's guidance about AI-generated content (8 February 2023)
5. Google Search Central: Creating helpful, reliable, people-first content, Who, How and Why (accessed November 2026)
6. Google Search Central: Spam policies for Google web search, scaled content abuse (accessed November 2026)

Nothing goes live until you approve it.

Artiweave proposes evidence-based changes to your B2B website with the reasoning and checks for each, publishes only what you approve, and keeps a full record of every decision. It is in private beta.

Join the waitlist