Reviewed by Product Specialist at thinQit. Updated 18 November 2026.
An AI generated website change should arrive with five pieces of evidence: the brief it answers, a before and after record of every edited element, a live preview, test results, and a readback from the published page. thinQit, the AI delivery workspace behind Cody and Codex, treats these five as the minimum a reviewer needs to approve a change.
What evidence should support an AI generated website change?
The evidence should let a reviewer answer three questions without investigating: what was asked, what changed, and does it work. Anything that does not help answer one of those questions is decoration. A good evidence pack is short, specific and checkable against the published result.
AI delivery is fast, and that moves the risk. When a change takes two weeks, dozens of small human checks happen by accident. When it takes two minutes, those checks only happen if the process builds them in. The thinQit guide on why AI assisted delivery needs evidence, previews and approval gates makes the same point: the aim is to make the output legible, not to slow it down.
This article goes one level deeper than that guide. It lists the individual pieces of evidence, explains what each one proves, and says where each one stops being useful.
What does the brief prove, and why does it come first?
The brief proves that the change answers the right request. It is the written statement of the outcome, the page or component in scope, and the constraints, recorded before any code or copy is produced. Without it, a reviewer can only judge whether the change looks good, not whether it is correct.
In Codex, the thinQit website and web app builder, this step is part of the flow. You describe the business and who it is for, the assistant asks a few focused questions, and it writes a plan for you to approve. The Codex builder starts the build only when you say the plan looks right, so the approved plan is the reference every later piece of evidence is checked against.
A brief is useful evidence when it names the scope in plain terms. "Change the pricing headline on the homepage" can be checked. "Improve the homepage" cannot. The boundary of this evidence is that it says nothing about the result. A perfect brief can still produce a wrong change, which is why the next four items exist.
What a reviewable brief contains
- The outcome in one sentence, written in the requester's own words.
- The exact page, section or component that may change, and what must stay untouched.
- The inputs supplied, such as a logo, photos, a PDF or existing copy.
- The person who approved the plan, and when.
How should a before and after record look?
A before and after record lists every changed element with its old value and its new value, one row per change. It proves the scope: the reviewer can see that the headline changed, the footer did not, and nothing else moved. It reads like a receipt, not a press release.
The row format matters. Each row names the page, the field, the value before and the value after, using the exact text a visitor will see. A summary such as "updated the site" is not a record, because it cannot be checked. A row such as "homepage, first heading, old text, new text" can be checked in seconds.
For code changes this record is the diff inside the pull request. How Cody turns a described change into a pull request walks through that path, where the AI developer maps the repository, plans the steps and writes the code. The diff shows what changed in the files, while the before and after rows show what changed for visitors. A reviewer who needs to approve a visible change wants both: the diff explains how, the rows explain what.
The limit of this evidence is that it is a statement of intent. It tells you what the change was supposed to do. It does not prove the result behaves.
Why does a live preview count as evidence?
A live preview counts because it is the only item that shows the real experience. A description of a page is not a page, and a screenshot is one frozen moment. A preview on a real URL lets the reviewer click, resize and try to break it. It catches the failures that a change record cannot show.
Typical failures only a preview reveals include a link that points nowhere, a form that does not submit, a layout that collapses on a phone and a tone that reads wrong in context. Codex puts the preview next to the chat and offers computer, tablet and phone views, so the check across screen sizes takes a few clicks rather than a separate test session.
Two rules keep preview evidence honest. First, the preview must show the change under review, not an earlier version. A stable preview link that updates after every change helps here. Second, the reviewer should look at the pages next to the edited one. A change to a shared heading style affects every page that uses it, and the edited page alone will not show that.
The boundary of this evidence is time. A preview shows the result as it is now. It does not show what the live site will do after publication, which is why the last item on the list matters most.
Which test results belong in the evidence pack?
The test results that belong in the pack are the ones tied to the change: tests that cover the edited behavior, and a statement of what was run and what passed. They prove the change did not break something that used to work. Results that are not tied to the change add noise.
For a code change, Cody writes tests first and shows the results, so the pack contains a pass count and the names of the tests that cover the edit. For a content or layout change, the equivalent checks are narrower and more practical: every link in the edited section resolves, every image loads, the heading order is logical, and the layout still renders on a phone.
A short test checklist for a visible change
- Every link in the edited area opens the intended destination.
- Every form in the edited area submits and shows a confirmation.
- The layout holds at phone, tablet and computer widths.
- The page still carries one main heading, with sections nested logically beneath it.
- The tests covering the edited code pass, with the count shown.
The boundary here is coverage. Passing tests prove only what the tests exercise. A reviewer should read which behaviors were covered, because a green result on the wrong tests is false comfort.
How do you prove the change is actually live?
You prove it by reading the published page back and checking that the new values are present and the old ones are gone. This is the only evidence that describes the result visitors see. A merged pull request, a successful deploy message and a write receipt are progress, not proof.
The reason is simple. Between a merge and a live page sit builds, caches and deploy queues, and any of them can leave the old version in place. A readback fetches the live URL and compares what it finds with the after values in the change record, row by row. When the rows match, the change is verified. When the deploy has not landed yet, the honest status is "merged, awaiting deploy", and the check runs again later.
This is the discipline behind how Sophia, the thinQit SEO and GEO teammate, works: each content change is tied to its workflow, then read back from the live site after publishing. The AI teammates follow the same rule, because a claim without a readback is only a claim.
A screenshot of the live result, taken after the readback, is a useful companion. It adds the visual view that text matching cannot give, such as a broken layout around the edited text. The boundary of readback evidence is that it covers the pages you check. A change to a shared component deserves a readback on more than one page.
How much evidence does a small change need?
A small, reversible change needs the lighter version of the pack: the brief, the before and after row, and the live readback. A large or customer facing change needs all five items and a human approval before publishing. The size and reversibility of the change decide the weight, not how confident the AI sounds.
A practical split looks like this:
- Single sentence or heading edit: brief, before and after row, live readback.
- New section or page: all of the above, plus a preview and link and layout checks.
- Form, checkout or data change: all five items, with tests that cover the behavior.
- Shared component or style change: all five items, with the readback on several pages.
The approval step sits on top of the evidence. The thinQit guide to approval gates in AI delivery explains why a person should own the moment of going live: someone reads the evidence, opens the preview and decides. The evidence makes that decision fast, and the gate makes it accountable.
What does weak evidence look like?
Weak evidence is evidence that cannot be checked or that describes effort instead of results. It is common because it is easy to produce and it sounds reassuring. A reviewer who learns to spot it avoids most approval mistakes.
- The vague summary. "Improved the page" names no element and offers nothing to compare.
- The confident silence. A record that never mentions uncertainty, even though the build made a judgment call. Honest flags where the system was unsure are part of good evidence.
- The stale preview. A preview link that shows an earlier version of the work.
- The merge as proof. A closed pull request offered as evidence that the live site changed.
- The screenshot alone. One frozen view with no change record behind it.
Each of these fails the same test: could someone who did not run the build understand what happened from the record alone? If the answer is no, the pack is incomplete, and approval should wait.
Curious what it costs?
See the pricing and pick what fits you.
Frequently asked questions
Who reads the evidence, the requester or a separate reviewer?
Both can, but the person approving the change must read it. In thinQit, teams set approval expectations for each workflow, so a production change is checked against the agreed brief before it is accepted. The requester knows the intent, and a separate reviewer brings fresh eyes.
Is a pull request enough evidence for a website change?
No. A pull request shows the code diff, which explains how the change was made. It does not show what a visitor sees, whether the preview behaves, or whether the published site carries the new content. Add the before and after rows and a live readback.
How long should the evidence take to read?
Under a minute for a small change. If a reviewer cannot answer what changed and whether it works in that time, the evidence is too vague or too long. Good evidence is specific and short, with one row per change.
What if the readback shows the old content after a merge?
Treat the task as unfinished. The usual cause is a deploy that has not landed yet. Wait, run the readback again, and call the change live only once the new values are served. If the old ones persist, the change needs a fix, not an approval.
Do content changes need the same evidence as code changes?
They need the same logic with different checks. A content change needs the brief, the before and after text, a preview and a live readback. Code changes add test results. The question is the same for both: what was asked, what changed, and does it work.
Sophia is thinQit's AI SEO & GEO specialist. She runs continuous technical audits, maps search and answer-engine intent, and tunes content so it ranks on Google and gets cited by ChatGPT, Perplexity, Gemini and AI Overviews.
