An AI Content Policy Template for SEO Agencies

It is Thursday afternoon and a client's head of marketing has just forwarded you a screenshot from a free AI detector showing 94% on the blog post you delivered Tuesday. Your account manager is asking what to say. Your writer swears she wrote it, and she probably did. The reason this moment is so uncomfortable is not the score; it is that your agency has no written position to point at.

An SEO agency AI policy should define permitted uses by production stage, set a measurable editing standard, specify what gets disclosed to clients, and use detection as routing rather than as a verdict. Write it before a client asks, put it in the contract, and make every claim in every deliverable traceable to a source a human checked.

What an SEO agency AI policy has to answer

Most agency AI policies fail because they try to answer the wrong question. They attempt to define how much AI is acceptable, land on a number, and then discover that the number is unenforceable and indefensible.

A workable policy answers four questions instead:

1. At which stages of production may AI be used, and at which may it not?
2. What does a human have to contribute before something ships?
3. What does the client get told, and when?
4. What happens when someone breaks the rule?

Everything below is a template for those four answers. Adapt the wording; the structure is the part that matters. And be clear about what you are protecting against. The exposure is not that Google detects AI, because it does not do that and says as much. Google's published position is that it rewards helpful content regardless of how it is produced and acts against scaled content abuse, a volume-and-value problem rather than a tooling one. We unpack that distinction in does Google penalize AI content. Your real exposure is a client who believes they bought human craft and received unedited generation, plus the reputational damage of publishing a fabricated statistic under their brand.

Step 1: Draw the permitted-use line where the work actually happens

Stages, not percentages. A writer cannot honestly report that a draft is "30% AI," but anyone can report that they used a model to summarize six competitor pages and then wrote the piece themselves.

The prohibited rows are the ones that matter. Fabricated statistics and invented expert quotes are the failure that actually reaches a client's legal team, and the FTC has been explicit for years that advertisers own the truth of their claims regardless of who or what drafted them.

Step 2: Make the human editing requirement measurable

"Heavily edited by a human" means nothing. It cannot be checked, so it cannot be enforced, so it is decorative.

Replace it with obligations someone can observe:

Every published piece must have a named editor of record. Before publication the editor confirms: at least one original example, data point or observation that no model could have produced from public text; a rewritten introduction and conclusion; every factual claim traced to a primary source; and a headline and subheads that were chosen rather than accepted.

That paragraph is copy-paste ready and it does something a percentage never will: it makes the human contribution auditable after the fact. If a client challenges a piece, you can show them the editor of record and the original reporting in paragraph four.

Step 3: Write the client disclosure clause you can actually honor

The trap here is promising more granularity than you can deliver. An agency that commits to flagging every AI-touched sentence has committed to something no workflow supports.

Disclose by stage. Something like:

AI tools may be used in research, outlining, drafting and metadata generation. All content is edited, fact-checked and approved by a named human editor before delivery. The agency is responsible for the accuracy and originality of everything it delivers, regardless of which tools were involved in production. On request, the agency will identify which stages involved AI tools for any specific deliverable.

That last sentence is the one clients actually want. It converts an abstract worry into an answerable question, and it costs you nothing if your process is real.

Step 4: Build a fact-check gate that fails closed

This is the step agencies skip, and it is the one that prevents the incidents that end relationships.

Rule: no statistic, quote, date, price, product claim or citation ships without a link to a primary source that a human opened. Not a link to an article citing the number. The source. If the claim cannot be sourced in five minutes, it gets cut rather than hedged into vagueness.

The reason for the fail-closed default is specific to how models behave. A language model asked for a supporting statistic will produce one that is plausible, well-formatted and correctly attributed to a real-sounding organization. It is the most confident-looking output the tool produces and it is wrong often enough that no volume of spot-checking catches it reliably. Detection tools will not save you here either, because a fabricated statistic in an otherwise human paragraph is invisible to a detector. Only a human with a browser tab catches it.

Step 5: Add detection QA with score bands, not a kill threshold

Detection has a real job in this workflow: verifying that you received the human effort you paid for. It has no business functioning as a pass/fail gate, and the research on why is unambiguous.

The 2023 Stanford study in Patterns found seven detectors flagged an average of 61.3% of TOEFL essays by non-native English speakers as AI-generated, with one at 97.8%, while scoring nearly perfectly on native-speaker samples. If your bench includes writers working in English as a second language, a hard cutoff is a policy that punishes them specifically. The RAID benchmark showed the other failure mode: commercial detectors degrade sharply under paraphrase, so the writer most likely to pass your gate is the one running output through a rewriter. A cutoff selects for evasion and against your most careful non-native writers. Both directions are bad. The mechanism behind the false positives is covered in why human writing gets flagged.

So use bands:

Sample rather than scan everything. Ten to twenty percent of each delivery, chosen at random, surfaces patterns without doubling QA cost. Compare each writer against their own baseline instead of a global threshold, because writers score differently and a good baseline turns a scary number into a known quantity. The full workflow is in our guide to AI content detection for SEO teams, and if you are choosing tooling, what AI screening tools catch compares the categories.

On cost, so you can budget the step rather than skip it. AI Detector 360 charges 1 credit per 100 words, so a 2,000-word article is 20 credits. An agency spot-checking 15% of a 60-article month scans nine pieces, roughly 180 credits. The free account's 300 monthly credits covers that; Starter at $9.99 covers 4,000 credits and Pro at $24.99 covers 15,000, which is more than enough to scan every word an agency of ten writers produces. Current details are on our pricing page, and the free AI detector handles ad hoc checks at 5,000 characters with no sign-up. What each confidence level means, and when we refuse to score a sample at all, is published on our methodology page.

Step 6: Move the policy from a document into the contracts

A handbook nobody signed is a suggestion. Two clauses turn it into a policy.

In the client agreement:

The Agency may use generative AI tools in the research, outlining and drafting of Deliverables. All Deliverables are reviewed, edited and fact-checked by Agency personnel prior to delivery. The Agency warrants that Deliverables are original, do not infringe third-party rights, and that all factual assertions are supported by identified sources.

In the freelancer agreement:

Contractor must disclose the use of generative AI tools in any submitted work at the time of submission. Undisclosed AI-generated content is a material breach. Contractor warrants that all statistics, quotations and citations in submitted work have been personally verified against primary sources.

Note that neither clause prohibits AI. They allocate responsibility, which is what a contract is for and what a prohibition never achieves.

What a policy like this can't do for you

Being straight about the limits, because a policy oversold is a policy that fails at the worst moment.

It will not tell you whether a given piece was AI-generated. No tool can, ours included, and any agency selling clients on certainty is setting up a future apology. It will not survive a client who wants a per-sentence audit trail, so decline that request in the sales cycle rather than after signing. And it will not fix a business model built on volume at a price that only unedited generation can hit; a policy cannot solve an economics problem.

What it does is make your position stateable on a Thursday afternoon when a screenshot lands in the shared channel. That is worth writing down before you need it.