How to refresh your content audit for the AI era

Most content audit templates still in use across Australian government were built before AI systems started reading and citing public content. They track URL, owner, last reviewed date, traffic, and a plain-language score. They do not track whether a page is structurally legible to a large language model, whether it is being cited accurately in AI answers, or whether it carries the metadata needed to be retrieved correctly. A content audit in the AI era needs new columns. This guide is for content designers, content strategists, and digital leads in government who already run audits and need to bring them up to date.

Before you start

You do not need new tooling for this. You need your current audit spreadsheet, a sample of 15 to 20 representative pages, and access to a few public AI tools: ChatGPT, Perplexity, and Google's AI overviews.

Choose pages that carry weight. High-traffic service content, eligibility information, and pages that answer a common question. Those are the pages AI systems are most likely to read and surface, so they are where audit gaps cost you most.

Step 1: Add five new columns to your AI-era content audit

Keep every column you already have. The AI-era audit is an extension, not a rebuild. Add five columns, and define each one so your team scores them consistently:

• Structure score: how cleanly the page breaks into headings, short paragraphs, and lists that a machine can parse into discrete answers.

• Retrieval quality: whether an AI system can find and pull the page when someone asks the question it answers.

• AI citation accuracy: whether AI tools represent the page correctly when they cite or paraphrase it.

• Schema completeness: whether the page carries structured metadata, such as page type, review dates, and FAQ markup, that tells a machine what it is.

• Semantic specificity: whether the page names things precisely, such as legislation, agencies, and programs, rather than using vague references.

These five columns are the difference between an audit that describes your content and one that predicts how AI systems will treat it.

Step 2: Score structure and semantic specificity from the page itself

Two of the five columns you can score just by reading the page. Do these first because they are fast.

For structure, check whether each section answers one question under a clear heading. A page that runs three concepts together under a vague heading scores low. A page with a heading like "Who can apply" scores high.

For semantic specificity, check whether the page names things. "The Disability Discrimination Act 1992" is specific. "Relevant legislation" is not. AI systems extract and cite specific, named claims far more reliably than general ones.

Step 3: Test AI representation on a sample, without paid tools

This is the step most audits skip, because people assume it needs expensive software. It does not.

Take your sample of 15 to 20 pages. For each one, write the plain question a user would ask to reach it. Then put that question to ChatGPT, Perplexity, and Google's AI overview.

Record three things: whether your page is retrieved at all, whether the answer matches what your page actually says, and whether the source is attributed to you or to someone else. That gives you real scores for retrieval quality and citation accuracy across a representative sample, for the cost of about an hour.

Step 4: Score schema completeness and apply a simple rubric

For schema completeness, check whether the page carries structured data. Most government CMS platforms support basic schema, yet many pages still ship without it. Mark each page as complete, partial, or missing.

Then score all five columns on the same simple rubric, so the audit stays readable:

• 2: meets the standard, no action needed

• 1: partial, needs work

• 0: fails, priority fix

A three-point scale is enough. A rubric people can apply in seconds is worth more than a precise one nobody finishes.

Step 5: Use the audit to scope the work that follows

An audit is only useful if it points to a decision. Sort your sample by total score, lowest first. The pages at the bottom are your scope.

Group the fixes by column. If most pages fail on structure, that is a content design and templating job. If they fail on semantic specificity, that is a writing and review job. If they fail on schema, that is a CMS and metadata job. The pattern in the scores tells you which discipline to bring in first.

This is also the discovery step for any larger AI representation work. You cannot scope a fix until you know which pages fail, and why.

What good looks like

A refreshed audit gives you three things the old version did not: a sample you have actually tested against live AI systems, five new columns scored on a rubric your team can apply quickly, and a ranked list of pages that tells you what to fix and in what order.

The 2018 audit told you what content you had. The AI-era content audit tells you whether that content will survive the way people now find answers.

If you run audits on a cycle, which pages would you be most worried to see an AI system answer on your behalf right now? Start your sample there.

Get the AI-era content audit add-on

Five columns to add to an existing content audit, and a scoring key for all five. Keep every column you already have and add these alongside them.

Structure score	Retrieval quality	AI citation accuracy	Schema completeness	Semantic specificity
How cleanly the page breaks into headings, short paragraphs, and lists that a machine can parse into discrete answers.	Whether an AI system can find and pull the page when someone asks the question it answers.	Whether AI tools represent the page correctly when they cite or paraphrase it.	Whether the page carries structured metadata, such as page type, review dates, and FAQ markup, that tells a machine what it is.	Whether the page names things precisely, such as legislation, agencies, and programs, rather than using vague references.

Scoring rubric. Score every new column on the same three-point rubric:

• 2: meets the standard, no action needed

• 1: partial, needs work

• 0: fails, priority fix

Photo by Nick Morrison on Unsplash