Most teams stall halfway through content audits. Someone exports a crawl, pastes in Search Console clicks and gives the first 80 URLs careful verdicts, but never scores the long tail before the quarter ends.
With an agent-run audit, you finish with a scored row for every URL, a keep, refresh, consolidate or remove verdict with the evidence attached, and a short list of pages you should build next. You run content gap analysis inside the same procedure, and once you've scripted the steps, you can schedule the audit instead of waiting for the next quarterly review.
What a content audit is and how it differs from auditing AI agents
A marketing content audit is a systematic review of every page you've published. It starts with an inventory, evaluates each URL on quality, performance and relevance, and ends with a decision about what to keep, update, merge or remove.
ROT content is the usual target. The acronym stands for redundant, outdated and trivial. Redundant pages duplicate something else on the website, outdated pages are no longer accurate, and trivial pages never had a reason to exist.
Auditing with AI agents is a different job from auditing AI agents. Agent-governance and compliance audits review what an AI system did, which tools it called, and whether it stayed inside its permissions. Here, the agents are the labor. They read pages, pull data and draft verdicts, and you keep a log of their actions so the audit itself stays reviewable.
Content gap analysis finds the topics buyers and competitors cover that your website doesn't. We run it as the last stage of the audit, because you can't tell what's missing until you know what exists and which pages already work.
What you need before you start
Gather five inputs before writing a single prompt. Agents produce weak verdicts when they guess at context you could have handed them.
- Crawler export: A Screaming Frog crawl with status codes, canonicals, word counts, titles, crawl depth and inlinks. The free version stops at 500 URLs.
- Search Console and GA4 access: Read access to clicks, impressions, CTR and position per URL, plus sessions and engagement per landing page.
- CMS read access: Publish and modified dates, authors, content types and the full page body.
- Context documents: A company overview, ideal customer profile, personas, competitor list and topic clusters. These tell the agent what "on intent" means for your business.
- Model or agent access: An API key or subscription for the model you'll run, and whichever agent tool you choose in the step below.
How to audit content with AI agents step by step
Build the inventory, define the rubric and prompts, then configure the agent to score each URL with read-only data access. Use the verdicts to find content gaps and route findings through human approval in these eight steps:
- Build the inventory from crawler, Search Console and GA4 data.
- Define a scoring rubric with weights and thresholds.
- Write a prompt chain that scores one URL at a time.
- Choose where the agent runs.
- Connect it to your data with read-only access.
- Score every URL and assign verdicts.
- Run a content gap analysis on what the inventory doesn't cover.
- Send every finding through a human approval gate.
Build the content inventory with a crawler, GSC and GA4
Merge everything into one sheet with one row per URL. Join on the normalized URL (lowercase, trailing slash handled consistently, tracking parameters stripped), then add clicks, impressions, average position, sessions and engaged sessions for the same date range.
Screaming Frog can pull Search Console and GA4 data during the crawl through its API Access settings, which saves the manual join. Two API limits shape the result.
- The Search Console API caps detailed exports at 50,000 rows per day per search type and returns the top rows by clicks, so low-traffic URLs can vanish from large pulls. The bulk export to BigQuery avoids that cap.
- On the GA4 side, check each response's thresholding and sampling metadata, because high-cardinality page dimensions can collapse rare URLs into an "(other)" row.
The common gotcha is an inventory that only contains linked pages. A crawl from the homepage misses orphans, which are URLs with no internal path to them. Screaming Frog finds them by comparing the crawl against XML sitemaps, GA4 and Search Console, then running Crawl Analysis to populate the Orphan Pages report.
Export the redirect chains report too. A redirected URL belongs in the inventory as a pointer to its destination, never as a page to score.
Define the scoring rubric for each page
Score each URL against five or six criteria with fixed weights, so two runs on the same page land in the same verdict band. These are the starting weights we'd use, and you should tune them after the first calibration pass:
| Criterion | What the agent checks | Weight | Fails when |
|---|---|---|---|
| Intent match | Does the page answer the query it ranks for and serve a named persona | 25 | Top queries ask a different question than the page answers |
| Depth | Does the main content add original detail beyond common knowledge | 20 | The page summarizes others without new information |
| Freshness | Are facts, prices, screenshots and dates still accurate | 15 | A stated fact is wrong or a dated claim has expired |
| Cannibalization | Does another URL target the same intent | 15 | A second page splits impressions for the same queries |
| Performance direction | Are clicks and impressions rising, flat or falling | 15 | Clicks fell for three consecutive months |
| AI visibility | Is the page cited in AI answers for its topic | 10 | Competitors are cited on the page's core prompts and it isn't |
Anchor the depth and freshness criteria in Google's guidance on helpful, people-first content. Its self-assessment questions ask whether a page adds original information and whether dates get changed without substantial edits, and both translate directly into rubric checks.
A total of 70 or higher passes. Between 50 and 69 means refresh, and below 50 means a human reviews the page for consolidation or removal. Any single hard fail (a wrong fact, a direct cannibal) flags the page regardless of the total.
Write the agent prompts
Run a chain of three short prompts per URL instead of one long one. Short prompts are easier to debug, and each step's output becomes the next step's evidence.
PROMPT 1 · EXTRACT
You are auditing one page. Inputs: page body, URL, title, publish and
modified dates, top 20 GSC queries with clicks and impressions, GA4
sessions for the last 6 months, company context, persona list.
Return JSON: {url, primary_intent, persona_served, claims_with_dates[],
outbound_sources[], word_count}. If a field cannot be determined from
the inputs, set it to "unknown". Do not infer.
PROMPT 2 · SCORE
Using the extraction and the rubric below, score each criterion 0-100.
For every score, cite the input field or source URL that supports it.
Return JSON: {url, scores{intent, depth, freshness, cannibalization,
performance, ai_visibility}, evidence{criterion: [source]},
unknowns[], hard_fail: true|false, hard_fail_reason}.
PROMPT 3 · VERDICT
Apply the decision matrix to the scores. Return JSON:
{url, verdict: keep|refresh|consolidate|remove, target_url (if
consolidate), reasons[], confidence: high|medium|low}.
Any verdict of consolidate or remove must set confidence no higher
than medium and list the evidence a reviewer should check.The explicit "unknown" instruction matters more than any other line. Models fill gaps with plausible guesses unless you give them a sanctioned way out.
Choose where the agent runs
Pick the runtime by audit size and team skill, since every option below can execute the same prompt chain. Here's how the options compare as of November 2026:
| Option | Best fit | Skill needed | Cost model |
|---|---|---|---|
| No-code tools (n8n, Zapier, Make) | Scheduled pipelines that read a sheet, call a model and write results back | Low | Paid plans by executions or tasks. n8n's community edition is free to self-host |
| Claude or ChatGPT with file uploads | A pilot on a few hundred URLs from a CSV | Low | Consumer or team subscription |
| MCP clients (Claude, Cursor, Claude Code) | Live reads from GA4, the crawler and the CMS during scoring | Medium | Client subscription plus server setup |
| LangChain and LangGraph | Custom Python or JavaScript pipelines with state and retries | High | Open source, with paid LangSmith tracing and scheduling |
| CrewAI | Splitting extraction, scoring and verdicts across role-based agents | Medium to high | Open source, with a limited free hosted tier |
| Microsoft AutoGen | Existing AutoGen projects only, since the repository is in maintenance mode and points new work to Microsoft Agent Framework | High | Open source |
| Output.ai | TypeScript workflows built with Claude Code, with prompts, evals and traces kept in one repo | High | Open source (Apache-2.0) |
For a first audit, a chat upload is enough to calibrate the rubric. Move to a scheduled pipeline once the prompts stop changing.
Connect the agent to your data with least-privilege access
Give the agent read access and nothing more. Search Console has a read-only scope (webmasters.readonly) and GA4 has analytics.readonly, so request those scopes and add the service account to GA4 as a Viewer.
Google's official GA4 MCP server is read-only. Screaming Frog ships a built-in MCP server that runs inside your licensed installation. For the CMS, the official WordPress MCP adapter exposes only the abilities your website registers, and Webflow's MCP server respects each user's existing role. In general, MCP access should be read-only by default, with write access a per-person opt-in where it is enabled for the workspace.
Apply these controls before connecting the agent, because the risks are documented. The OWASP MCP Top 10 lists token mismanagement and scope creep among its top entries, along with tool poisoning, where injected misleading context manipulates the model's behavior. The controls are:
- Use short-lived tokens and keep keys in a vault.
- Review any community MCP server's code before installing it, and never give an audit agent write access to the CMS.
- Before signing with a hosted platform, check its SOC 2 report, data retention terms and sub-processor list.
Score every URL and assign keep, refresh, consolidate or remove
Map rubric results to verdicts with a fixed matrix so the agent applies the same rule on page 12 and page 412:
| Rubric result | Traffic signal | Verdict |
|---|---|---|
| Total 70+ and no hard fail | Any | Keep |
| Total 50–69, or freshness hard fail | Has clicks or inlinks | Refresh |
| Cannibalization hard fail | Two URLs split the same queries | Consolidate into the stronger URL |
| Total below 50 | No clicks, no links, no replacement need | Remove |
For cannibalized pairs, pick the winner by existing rankings, backlinks and completeness, merge the best material into it, and redirect the loser. Google treats a permanent 301 or 308 redirect as a strong signal that the target should be canonical. When a removed page has no replacement, return a 404 or 410. Google's core-updates guidance treats deletion as a last resort, so a weak page with real substance goes on the refresh list first.
A finished row should look like this:
{
"url": "/blog/expense-policy-template",
"verdict": "consolidate",
"target_url": "/guides/expense-policy",
"reasons": ["Both URLs rank for 'expense policy template'",
"Target has 4x the referring domains"],
"evidence": {"cannibalization": ["gsc_export:query_overlap"]},
"unknowns": ["ai_visibility"],
"confidence": "medium"
}Run a content gap analysis on what the audit leaves out
The scored inventory shows what you have. Content gap analysis compares it against three outside references:
- Pull keywords your closest competitors rank for that you don't, group them by topic, and check which topics have no URL in your inventory.
- Review the SERPs for those topics to see which subtopics, formats and questions every ranking page covers.
- Run your buyer prompts through ChatGPT, Claude, Perplexity and Google AI Overviews and record where competitors get cited and you don't.
Classify each gap as a new page or a rewrite. Create a new page when the intent is distinct and no existing URL serves it.
Rewrite when a URL already ranks or gets cited for that intent, because a rewrite keeps the URL and its history while a second page would split relevance. When the subtopic has little search demand of its own, expand the existing page with a section.
Route findings through a human approval gate
A named reviewer must sign off before anyone acts on a verdict. Reviewers can accept low-risk keep verdicts in batches.
An editor must confirm the stated problem before approving a refresh verdict. Reviewers must check consolidate and remove verdicts page by page, because redirects and deletions are hard to reverse.
Escalate any finding where the agent marked confidence as low, listed unknowns on a hard-fail criterion, or recommended removing a page with backlinks or conversions. A person applies the actual changes (redirects, merges, deletions) in your CMS through the review gates you already use.
Keep agent findings traceable and accurate
Every finding needs a source trail a reviewer can follow in under a minute. Require a cited input field or URL for each score, log every tool call and model response with a timestamp, and store the prompt version that produced each verdict.
Then check the agent the way you'd check a new analyst:
- Re-score a random 10% sample by hand each run, plus every remove verdict.
- Confirm that each cited source URL exists and says what the agent claims.
- Track the disagreement rate between agent and reviewer over time, and when it climbs, inspect the rubric before you blame the model.
Humans stay in the loop because the error rates are not zero. A single-practitioner benchmark of six models on a 100-page content audit found the best-value model, GPT-5.4, invented 3.4% of its reported issues, while two smaller models invented 16.9% and 19.4%. That's one test on a small sample, so treat it as a warning rather than a rate to plan around.
The split is human-led strategy, AI-led execution: AI can replace the hours of reading and data merging, but it can't replace the judgment that decides whether a page matters to your business.
A worked example on a 500-page website
The numbers below are estimates for a 500-URL audit, built on current list prices for Claude models on Anthropic's pricing page. Assume the three-prompt chain uses about 6,000 input tokens and 1,500 output tokens per URL, including page text, Search Console rows and the rubric. That's 3 million input tokens and 750,000 output tokens for the full audit.
| Approach | Model or method | Estimated cost | Estimated time |
|---|---|---|---|
| Agent, mid-tier model | Claude Sonnet 5.5 at $2 input and $10 output per million tokens | About $13.50, or about $6.75 through the Batch API's 50% discount | Under a day of runtime, plus 15–25 hours of human review |
| Agent, low-cost model | Claude Haiku 5.5 at $0.10 input and $0.50 output per million tokens | Under $1 | Same runtime, more review time for weaker verdicts |
| Manual audit | Analyst reading each page | Staff time | Up to 2,000 hours at four hours per page |
The manual line uses the four hours a careful page analysis can take, multiplied across 500 pages. Few teams spend that much, which is exactly why they sample instead of covering every URL.
The review hours are our estimate for a 10% spot check, every consolidate and remove verdict, and the gap list. Token spend is the smallest line in the budget.
What you spend depends on how long reviewers need, so the cheapest model isn't automatically the cheapest audit.
Measure AI search visibility in the audit
AI visibility belongs in the rubric because a page can rank well and still never appear in AI answers. Two terms need separating first.
A mention is your brand named in the answer text. A citation is your URL used as a source for the answer. The two often tell very different stories, since a brand can be named with no link and a page can be cited in an answer that recommends a competitor.
Track both across ChatGPT, Claude, Perplexity and Google AI Overviews using a fixed prompt set. Answers change between runs, so run each prompt several times and report a range, such as cited in 4 to 6 of 10 runs, instead of a single reading.
CheckThat, GrowthX's AI visibility measurement product, tracks daily brand mentions, sentiment and citation sources across those engines with editorially reviewed prompts. Its measurement model groups results into Presence, Reputation, Perception and Influence, which maps cleanly onto the audit's question of whether a page gets cited and how the brand gets described.
Turn the audit into a continuous process
A quarterly audit describes a website that has already changed. The continuous version runs the same steps on a cadence. Schedule a weekly recrawl, refresh Search Console and GA4 data once each source finalizes (Search Console data typically lands two to three days late), and rescore only the URLs whose inputs changed.
Direction matters more than level when choosing alerts. A website-wide traffic average can hold steady while dozens of pages quietly decline, so alert on per-page trends. Flag any URL whose clicks fell three months running, any page newly flagged as a cannibal, and any core prompt where a competitor started getting cited. Inside the GrowthX platform, Portfolio sorts every page into direction bands (Freefall, Declining, Stable, Growing, Surging) alongside a 0–100 Health score from 77 technical checks, which is the same signal this alert design produces by hand.
Fix common problems with agent audits
- Context window limits: Score one URL per call and pass the cannibalization candidates in as a short list of overlapping URLs.
- Over-flagging: Require a specific wrong or expired claim before freshness can fail, and treat a missing date alone as an unknown.
- Cannibalization false positives: Two pages can share keyword phrasing and serve different intents, such as a template and a strategy guide. Check query overlap in Search Console before accepting the flag, and let a reviewer confirm every pair.
- Stale exports: Stamp every export with its date range and refuse to score inputs older than the current cycle.
- Inconsistent scores across runs: Setting temperature to 0 helps but doesn't guarantee identical output, and one study measured answer instability of up to 15% at zero temperature. Constrain verdicts to a fixed set of labels, run borderline pages twice, and send disagreements to a human.
Start with a 50-page calibration run. Compare the agent's verdicts against your own, adjust the rubric weights where you disagree, and only then point it at the full inventory. If maintaining this workflow across separate tools is the bottleneck, book a demo to see how the GrowthX platform runs it in one workspace.
