Should your agency tell you if your content is AI-written?

The question is about the invoice, not the algorithm

When a founder asks us this, they've almost always just read a blog post their agency delivered and felt something off about it. What follows is usually two questions tangled into one, and they have completely different answers.

The first is: does this hurt my rankings? That one is settled and slightly boring by now. Google's position has been authorship-neutral since 2023 — the systems reward helpful, original content regardless of how it was produced, and the spam policy targets scaled content abuse rather than automation as such. We've written the long version of that argument in why AI-written content ranks and still gets sites killed, and the shorter version at does AI content hurt SEO.

The second question is the one nobody answers: does your agency owe you the disclosure, and what should it do to the price? That's not an algorithm question. It's a question about what you thought you were buying.

Does an agency owe you the disclosure?

Our position: yes, at the level of process, and no, not per paragraph.

Per-article labelling doesn't survive contact with reality. Almost every content workflow in 2026 has a machine somewhere in it — clustering queries, drafting an outline, checking a draft for internal contradictions. A label saying "AI-assisted" on every deliverable would be simultaneously true, useless, and impossible to audit.

Process-level disclosure is different and entirely reasonable. You are entitled to know which steps are automated, who reviews the output, whose name is accountable, and whether every specific in the piece was checked against a source. That is a question about the service you bought, and an agency that gets defensive about it is telling you something.

There is also a reader-facing dimension, which is separate from your commercial one. Google's guidance on people-first content asks creators to consider whether the use of automation, including AI-generation, is self-evident to visitors through disclosures or in other ways, and to explain why automation was useful — adding those "when it would be reasonably expected". For a technical explainer, nobody expects it. For a first-person clinical or legal opinion, the expectation is real, and the risk of getting it wrong isn't a ranking risk. It's a credibility one.

Where we use these tools, and where we don't

Since we're arguing you should demand this, here's ours, stated plainly enough to hold us to.

We use models for query clustering, competitive reading, outlines, schema, first-pass drafts in Hindi and regional languages, and adversarial QA — feeding a finished draft back in and asking what a hostile expert would attack. We don't police whether a draft was typed or generated, because policing that is theatre and the tell isn't in the prose.

What we do police is what happens after. Every number, date, price, policy name and quotation gets checked against a primary source or it comes out — deleted, not softened. A senior editor reads for position: is this what we'd say on a client call, and is there a sentence here a competitor wouldn't dare publish. And a person's name goes on it, not an "editorial team", because nobody signs off filler with their own name attached.

The reason we can publish that without flinching is that it doesn't change the price. Our SEO retainers start at ₹75,000/mo, or ₹40,000 for smaller sites; content and organic social runs at ₹30,000, ₹50,000 or ₹70,000 a month. Those numbers were set against verification hours, not drafting hours — which is why the arrival of cheap drafting didn't move them much.

The disclosure clause, written out

Add this to the next contract or the next renewal. It commits an agency to nothing about tooling and everything about review, which is the part you actually care about. Adapt the wording with your own lawyer; the shape is what matters.

• Disclosure. The Agency will maintain, and provide to the Client on request, a written statement of where generative AI tools are used in producing deliverables under this agreement — covering at minimum research and clustering, outlining, drafting, translation, and quality assurance.
• Human review. Every deliverable is reviewed and edited by a named individual before publication. The Agency will identify that individual for any deliverable on request.
• Verification. Every statistic, date, price, quotation, regulation and policy reference in a deliverable is verified against a primary source before publication, and the source is recorded.
• Attribution. Where a deliverable is published under a named author or expert, that person has read and approved it. No deliverable will be attributed to a person who has not seen it.
• Tooling is not restricted. Nothing in this clause limits the Agency's choice of tools. It governs disclosure, review and accountability only.
• Pricing basis. Fees are charged per published deliverable and per month of capacity, not per word.

What should change about the price, and what shouldn't

Here is the honest accounting. Some line items genuinely collapsed. Most didn't, and the ones that didn't are the expensive ones.

• The one thing that should change structurally is per-word pricing. A rate card of ₹3 a word is now a rate card for the cheapest step in the process, and it prices exactly the wrong thing. Move to a fee per published page or a monthly capacity fee. What content actually costs in India covers the arithmetic.
• If an agency's costs fell and neither the fee nor the output moved, that's a fair thing to raise at renewal. Ask for it as more work rather than a discount — the discount is smaller than you think and the extra output compounds.
• Be careful what you wish for on the discount. Every rupee taken out of a content retainer comes out of editing first, because drafting is already near zero. Cheap content is downstream of a pricing decision, not a tooling one.

The tells that a retainer's content is generated and unedited

Detection tools don't work reliably and you shouldn't rely on them — they produce false positives on careful human writing and false negatives on lightly edited output. What does work is reading four articles in a row and looking at the shape rather than the sentences.

• Identical architecture across every piece. Intro, a line about what the article covers, five H2s of three paragraphs each, a summary. Human writers vary; a template doesn't.
• No specifics that could only come from your business. No pricing, no client situation, no number anyone in your office would recognise. This is the strongest single tell and it's also the biggest lost opportunity.
• Numbers with no source, or a source that doesn't say it. Check three. If a statistic is attributed to a study, find the study. This is where unedited output fails hardest and where it costs you most.
• Internal links that are missing, generic, or point at pages you don't have. A person who has read your site links to your site.
• Hedging density. "It's important to note", "in the ever-changing world of", "whether you're a small business or a large enterprise". Sentences that survive because nobody was willing to be wrong in public.
• No named author, or an author with no trace. Ask who wrote it. The question is not aggressive and the pause tells you a lot.
• Publishing bursts. Eleven pages in a day, then nothing for three weeks. Cadence follows capacity when humans are involved.
• The vocabulary folklore is unreliable. The circulating lists of giveaway words and punctuation catch plenty of good human writers, and any competent editor strips them anyway. Judge by specificity and verifiability, not by word choice.

What to do if you find it

Assume it's a process failure rather than fraud, because it usually is, and open accordingly. "Walk me through how these pages are produced and who reviews them" gets you further than an accusation, and it's the same question you'd ask about anything else on the invoice.

Then fix it in three moves. Add the disclosure clause at renewal. Change the pricing basis away from per-word. And audit what's already published — pull every page from the last two quarters, check three specifics on each, and delete or rewrite anything that fails, because a confidently wrong page keeps costing you long after the retainer ends.

If the answer is that nobody reviews anything and the volume is the product, that's not an AI problem. That's an agency that priced its offer on a saving that only exists if you skip the checking. The tooling is incidental — the same shop was shipping unedited freelance copy in 2019.