AI Detection for SEO Agencies: How to Audit a Client’s Existing Content for AI Risk

You take on a new client and inherit three years of blog posts, landing pages, and product copy you didn’t write. Some of it might be AI-drafted, some thin, some perfectly fine, and you have no idea which is which. Before you promise anyone a ranking turnaround, you need a way to look at that pile and figure out what’s a liability and what’s worth keeping.

That’s what an AI content audit is for. Not to catch anyone, not to slap a verdict on every page, but to triage an existing library so you spend your hours on the pages that actually need them. Here’s how to run one without scaring the client or burning a budget chasing scores that don’t matter.

Key takeaways

• You’re auditing for risk and quality, not authorship. The goal is to find pages that read like autopilot and decide which ones are worth fixing.
• Detector scores are probabilities on a sample, not proof. Sample strategically across page types and dates instead of running every URL.
• The combination that should top your fix list is high AI score plus thin content plus flat traffic, not a high score alone.
• Document your method before you touch anything. A repeatable audit you can hand a client beats a one-off score dump.
• Substance first, rhythm second. Rewriting for a lower score without adding value just makes a thin page harder to spot.

What you’re actually auditing for

It helps to be honest about the question. A client rarely cares whether a page was typed by a person or a model. They care whether the content helps their rankings or quietly drags them down. So an AI content audit isn’t really about detection at all. It’s a content-quality audit that happens to use a detector as one of its instruments.

A detector measures how predictable the writing is and how little the sentence lengths vary. We’ve covered the statistics these detectors actually run on, but the short version: they’re pattern-matching for the smooth, even hum that machine drafts tend to produce. That hum often overlaps with *thin and generic*, which is the thing that genuinely hurts a client. So a high score is useful as a flag. It’s just not the finding itself.

Keep one fact pinned to the wall before you start. Detectors output probabilities, and they get things wrong. OpenAI retired its own AI classifier in July 2023 because the accuracy was too low to be useful. A 2023 study led by Stanford’s Weixin Liang found detectors disproportionately flagged essays by non-native English speakers as AI, even though humans wrote all of them. If a client’s writer is ESL, or the brand leans on a rigid template, expect false positives. You’re gathering signals, not handing down verdicts.

Scope the audit before you run a single check

The instinct is to dump the whole sitemap into a tool and see what turns red. Don’t. On a site with thousands of URLs that’s slow, costly, and noisy. You want a sample that represents the library.

Pull your sample along three axes:

• Page type. Blog posts, category pages, product descriptions, and landing pages tend to come from different people with different processes. Sample each.
• Publish date. Content from 2021 was likely human-written by necessity. Content from 2024 onward is far more likely to be AI-assisted. The split tells you where any AI use crept in.
• Performance. Grab your top traffic pages, your top conversion pages, and a band of zero-traffic stragglers. A flat, high-scoring straggler is a candidate to prune or rewrite. A high-scoring page that ranks and converts is, frankly, doing its job.

Twenty to fifty pages usually paints the picture for a mid-size site. If a cluster lights up, you go deeper on that cluster. If the sample comes back clean, you’ve spared the client a big detection bill for nothing.

Read the scores like an analyst, not a judge

A score by itself means very little. What you’re building is a small matrix per sampled page, with three columns that matter more together than any one does alone.

1. Detection signal. Does it read predictable and flat? Note the score and, more useful, *which sections* run hot. Detectors that show sentence-level heat tell you where the writing coasted.
2. Substance. Does the page say anything original? Real numbers, first-hand detail, a point of view? Or does it restate the top results and add nothing?
3. Performance. Impressions, clicks, conversions, trend over the last year.

Now the page basically sorts itself:

• High score, thin, flat traffic: top of the fix-or-prune list.
• High score, but it’s a template (specs, location data, structured listings): probably a false positive on formulaic text. Leave it.
• High score, but it ranks and converts: don’t touch it to chase a number. It earned its spot.
• Low score, thin, flat traffic: a detector won’t catch this, but it’s still weak. The audit found it anyway, which is the point.

That last row is the tell that you’re doing this right. The detector is a flashlight, not the map.

A mini-scenario: the inherited blog cluster

Say you onboard a regional law firm. They’ve got 80 blog posts, most published in the last 18 months, all on personal-injury topics. Organic traffic is flat to declining and they’re convinced their old agency “used AI and got us penalized.”

You sample 25 posts across dates and topics. Fourteen score high on a detector. Tempting to confirm the client’s fear and call it a day. Instead you read the three columns.

Those 14 posts aren’t flagged because of some Google penalty. They’re flagged because they’re interchangeable: every one opens with a generic definition, lists the same five “what to do after an accident” steps, and never mentions a single local court, statute, or case the firm actually handled. They’re thin, and the flat rhythm that lit up the detector is a symptom of the same emptiness that’s keeping them off page one.

Meanwhile, two high-scoring posts *do* rank, because somewhere in the boilerplate a paralegal dropped in real state-specific filing deadlines. Those stay. Your audit deliverable isn’t “your old agency used AI.” It’s “32 posts are thin and redundant, here are the 12 worth rewriting with real local expertise, and here are 20 to consolidate or prune.” That’s a plan a client can act on. A list of AI percentages isn’t.

Turn the audit into a deliverable

Agencies live and die on repeatability, so write the method down before you run it. A clean audit doc has:

• The sample logic. Which pages, why, what tools, what thresholds. If the client asks “did you check everything,” you can explain exactly what you checked and why a sample was the right call.
• The matrix. Score, substance note, and performance per page, with a recommended action: keep, rewrite, consolidate, or prune.
• A risk summary in plain language. “Most of your 2023 content reads predictably and thin. That’s a quality and ranking issue, regardless of who or what wrote it.” Notice that framing never accuses a person of anything. It keeps the conversation on the fixable part.

When you sit down to rewrite the flagged pages, fix the substance first. Add the local detail, the number nobody else has, the actual point of view. Then smooth the rhythm so it reads like a person wrote it on a good day. If you want to check how a page scores and see which sentences read flat while you edit, that’s the right moment to reach for a tool, on the way out, not as a gate. For more on where detection and search part ways, here’s the rest of our writing on AI detection and SEO.

What an audit can’t tell you

Two limits worth stating to the client up front, so nobody over-reads the results.

A detector can’t prove who wrote the page. It can’t tell a careful human writing in a flat institutional style apart from a model. So your audit names a *risk*, never a fact about a person.

And the audit score is not a Google ranking input. The detector you run is a separate third-party tool, and its number never reaches Google. Google has said it focuses on whether a page is helpful, original, and people-first, however it was produced. So a page scoring 90% AI is not “penalized.” It might rank fine. Your job is to find the pages that are genuinely weak, and the detector just helps you spot some of them faster.

The bottom line

An AI content audit is really a quality audit with a detector bolted on. Sample smart, read the score next to substance and performance, and you’ll find the pages that are dragging a client down, whether or not a machine helped write them. The score points you at suspects. Your judgment decides the case.

Do it as a documented, repeatable process and it becomes a service you can sell, not a one-off favor. If you’re standardizing this across a roster of clients, look at pricing so the math works at scale.