# How to Run a Content Audit With AI (and Agents)

**Author:** John Morabito (Founder, /winston)
**Published:** June 14, 2026
**Reading time:** 13 minutes
**Canonical:** https://www.winstondigitalmarketing.com/playbooks/content-audit-with-ai/

A content audit with AI is a structured review of every page you have published, where agents do the volume work and a human owns the judgment. The loop is five stages: inventory every URL, classify each page by intent and performance, decide keep, merge, prune, or refresh, prioritize by effort against payoff, then execute the winners with an agentic refresh workflow. The AI joins hundreds of URLs to traffic, ranking, and citation data in one pass. The person decides what is still true and what is worth saving.

## Why audits get skipped, and why AI fixes that

The reason most content audits never happen is not that teams doubt the value. It is that the manual version is punishing: export a URL list, cross-reference it against Search Console and analytics by hand, open each page to judge quality, hunt for duplicates from memory, and somewhere around page eighty the will to continue runs out. The audit becomes a spreadsheet that gets built once and acted on never.

AI removes the part that made it unbearable. An agent can crawl the full site, join every URL to its performance and citation data, flag the duplicates and the decayed pages, and hand you a draft recommendation for each one. What is left is the work that was always the point: deciding what to keep, what to combine, what to kill, and what to fix. The audit stops being an archaeology project and becomes a standing process you can actually run.

## Stage one: inventory every URL

You cannot audit what you have not listed. Start with a complete crawl and reconcile it against three other sources, because each one knows about pages the others miss: your sitemap (what you think you published), Search Console (what Google indexed and sends traffic to), and analytics (what actually gets visited). Orphan pages, forgotten landing pages, and old campaign URLs surface only when you union all four lists.

For each URL, pull the fields the later stages will need: title, primary topic, word count, publish and last-modified dates, indexed status, internal links in and out, clicks and impressions, average position, and whether it earns any AI citations. This is the join that makes the audit tractable. An agent assembles it in minutes; doing it by hand is exactly the step that kills manual audits. If a page is drawing citations from AI engines, note it, because a page that AI names is worth more than its raw traffic suggests.

## Stage two: classify by intent and performance

Every page gets scored on two axes, and the intersection decides its fate. The first axis is search intent: what the page is for. Is it trying to rank for a query, convert a visitor, inform a reader, or support another page in a cluster? The second axis is performance: is the page winning (traffic, rankings, or citations), holding steady, or decaying?

Doing this with AI means the model reads the page, infers its intent, and reconciles that against the query data, rather than a human guessing at hundreds of pages one at a time. Where the inferred intent and the actual ranking queries disagree (a page you wrote to convert that only ranks for an informational term), you have found a mismatch worth flagging. That gap is often the real reason a page underperforms.

## Stage three: decide keep, merge, prune, or refresh

Now the classification turns into an action. Four buckets cover almost every page, and the two axes from stage two point each URL at one of them.

| Action | When it applies | What you do |
|---|---|---|
| Keep | On-intent and performing | Leave it, log it, and maintain it on the refresh cadence |
| Merge | Two or more pages competing for one intent | Combine into the strongest URL, redirect the rest into it |
| Prune | Off-intent, thin, duplicative, or targeting dead queries | Remove or redirect, so it stops diluting the pages that matter |
| Refresh | On-intent but slipping on aged facts or fresher competitors | Update the page and re-submit it |

An agent can draft the recommended action for every URL with its reasoning attached, which is what makes the stage fast. But the delete, merge, and redirect calls are where a human has to sign off, because a dataset does not know that a low-traffic page closes deals from a single referral source, or that merging two pages would break a link a partner relies on. Draft with the agent; approve with a person. The honest finding on most audits is that there are more prune-and-merge candidates than the team expected, and clearing that dead weight is often what lets the surviving pages climb.

## Stage four: prioritize by effort and payoff

A raw action list is not a plan. Sequence it, or the audit dies in the same drawer the last one did. Score each item on two things: the effort to do it and the payoff if it works. That produces an obvious order.

- **Prune first.** Removing thin and duplicative pages is low-effort and clears the signal, so the pages you keep get read more cleanly by both Google and AI engines. Do it before you invest in anything else.
- **Merge next.** Consolidating split authority is medium-effort and often the highest single-move payoff, because two mediocre pages become one strong one instead of competing with each other.
- **Refresh the cite-critical pages.** Pages that earn or are close to earning citations, on fast-moving topics, get the refresh work first, because that is where freshness pays. The full cadence for this lives in our content-refresh system for AI search (https://www.winstondigitalmarketing.com/playbooks/content-refresh-system-ai-search/).
- **Refresh the rest on signal.** Everything else waits until a decay signal fires rather than getting reworked on a blind schedule.

## Stage five: run the refresh with agents

The audit produces the plan; the agentic workflow executes it. For the refresh queue in particular, the mechanical stages are agent work: draft the updated sections, flag the stale facts, restructure a section into a liftable answer-first chunk, prepare the re-submission. The human reviews each draft for truth and voice before it ships, which is the same staged build with a person on the judgment that we run for net-new writing. That craft (keeping the volume without the machine sound) is the whole subject of how to build an AI content pipeline that doesn't sound like AI (https://www.winstondigitalmarketing.com/playbooks/ai-content-pipeline-human-quality/).

Point the same agents at the maintenance loop and the audit stops being a one-time event. The inventory refreshes itself, the classification re-runs on a schedule, and pages that trip a decay signal queue themselves for the next cycle. That is the difference between an audit you do once and a content operation that stays clean. When the refresh work is heavier than the team can carry, it is the work we run inside our content creation service (https://www.winstondigitalmarketing.com/services/content-creation/).

## The honest version

The audit is not the deliverable. The prioritized action list is, and it is worthless if nobody works it. AI makes the boring 80 percent (the inventory, the joins, the classification, the first-draft recommendations) cheap enough that the audit actually gets finished, but the deletes, the merges, and the "is this still true" calls stay human on purpose. If a tool promises a fully automatic audit that prunes and republishes with no one reviewing, it will eventually delete something that was quietly earning its keep. Ready to have us run it on your library? Get in touch (https://www.winstondigitalmarketing.com/contact/).

## Frequently asked questions

**What is a content audit with AI?**
A content audit with AI is a structured review of every page you have published, where AI and agents do the heavy lifting: pulling a full URL inventory, joining it to performance and citation data, classifying each page by search intent and quality, and proposing a keep, merge, prune, or refresh decision for a human to approve. The AI handles the volume work that used to make audits get skipped, cross-referencing hundreds of URLs against traffic, rankings, and duplication. The human owns the judgment calls: whether a page still reflects the truth, whether two pages should become one, and whether a decline is worth reversing. The output is a prioritized action list, not a spreadsheet nobody acts on.

**How do you classify pages in a content audit?**
Classify on two axes and let the intersection decide the action. First, search intent: what the page is trying to do (rank, convert, inform, or support another page). Second, performance: whether it earns traffic, rankings, or AI citations, holds steady, or has decayed. A page that is on-intent and performing gets kept and maintained. Two pages competing for the same intent get merged. A page that is off-intent, thin, and drawing nothing gets pruned. A page that is on-intent but slipping gets refreshed. Doing the classification with AI means joining the inventory to Search Console, analytics, and citation data in one pass instead of eyeballing pages one at a time.

**Should you delete or refresh underperforming content?**
It depends on why the page is underperforming. Refresh a page that is on-intent, structurally sound, and only slipping because the facts aged or a competitor published something fresher. Prune a page that is off-intent, thin, duplicative, or targeting a query nobody searches, because keeping dead weight dilutes the signal on the pages that matter. Merge when two or more pages split the authority for one intent, then redirect the losers into the winner. The honest default is that most libraries have more prune-and-merge candidates than teams expect, and clearing them is what makes the remaining refresh work pay off.

**Can AI agents run a content audit end to end?**
Agents can run the mechanical stages end to end, but not the judgment calls. An agent can crawl the site, build the inventory, join it to performance and citation data, flag duplicates and decay, and draft the keep, merge, prune, or refresh recommendation for every URL. What it cannot do reliably is decide whether a borderline page is still true, whether a merge would break a real user path, or whether a declining page is worth saving for strategic reasons a dataset does not capture. The workable model is agent-run detection and drafting with a human approving the plan before anything gets deleted, merged, or republished.

Service: https://www.winstondigitalmarketing.com/services/content-creation/
Audit: https://www.winstondigitalmarketing.com/contact/#audit