Scaled Content Abuse: An AI Content Checklist for SEO Agencies

Google's spam policy calls it scaled content abuse when many pages are generated mainly to manipulate search rankings and not to help users, however the content is made. An agency can use AI safely if every page has a real reason to exist, adds value a user cannot get from a near-copy, and is checked by a person.

Google's spam policies page shows a last update of 28 August 2026, so read the current wording before you quote it to a client. The thing to take from it is the test: purpose and value, not tool choice. Forty location pages written from one prompt with the town name swapped fail that test however polished each sentence is; six pages built from the client's own prices, photos and service limits pass it.

Follow me on Instagram@sagnikteaches

What the policy actually lists

Google's definition is short. Scaled content abuse is "when many pages are generated for the primary purpose of manipulating search rankings and not helping users", typically large amounts of unoriginal content that provides little to no value, "no matter how it's created". The examples Google gives are:

Connect on LinkedInSagnik Bhattacharya

• using generative AI tools, or similar tools, to generate many pages without adding value for users;
• scraping feeds, search results or other content to generate many pages, including through automated transformations such as synonymising or translating, where little value is provided;
• stitching or combining content from different pages without adding value;
• creating multiple sites to hide the scaled nature of the content;
• creating many pages whose content makes little or no sense to a reader but contains search keywords.

Two phrases matter most. "Without adding value" is the hinge: AI is named, but only when value is missing. And "no matter how it's created" means a team of freelancers working from a template can break the policy just as a model can. The practical consequence for agencies is that volume plus sameness is the pattern to avoid, whatever the tooling.

Subscribe on YouTube@sagnikteaches

Google's guidance on generative AI content says the same thing in friendlier terms. Generative AI can be useful for researching a topic and adding structure to original content. The warning is about generating many pages without adding value, and about the need to factcheck and review everything before publishing, including titles, meta descriptions, structured data and image alt text. That guidance also points to the search quality rater guidelines, which discuss scaled content abuse and main content made with little effort, originality or added value, and notes that raters' judgements do not directly influence ranking.

Why agencies are exposed

Agencies fall into this policy because of the commercial shape of the work: a retainer with a content quota, a client who wants "100 location pages", a tool that can produce 500 drafts overnight. The same prompt run across a spreadsheet of towns or product types gives you pages that differ only in a noun. Google's guide on generative AI search adds a specific warning: creating separate content for every possible variation of how people might search, including fan-out queries, primarily to manipulate rankings or AI responses, violates the scaled content abuse policy and is also ineffective in the long run.

So the risk is not a vague "AI penalty". It is a policy violation that can lead to a manual action or to pages being demoted by automated systems, and clients will blame the agency whichever it is.

The value-added test, in five questions

Run this test on a sample page from every batch, and on every page in a batch of fewer than 20. A page passes when you can answer yes to the first and at least three of the rest.

1. Is there a user need? Can you name the person, the question and where they are when they ask it? "Rank for a keyword" is not a user need.
2. Does it contain something another page on the site lacks? Local detail, a price, a photo of the client's own work, a policy, a named process, a data point from the client's records.
3. Would the client be comfortable if a customer read only this page? It must be accurate, complete enough to act on, and not contradicted elsewhere on the site.
4. Could a competitor swap their name in without editing anything else? If yes, the page is commodity content. Google's AI guide contrasts commodity content ("7 Tips for First-Time Homebuyers") with first-hand material that goes beyond common knowledge.
5. Did a person with relevant knowledge check it? Name them in the workflow log.

The test on three pages

Illustrative case: an agency is asked for 40 service-area pages for a bicycle repair shop. Page A says "Bike repair in [town]. We fix bikes in [town] and the surrounding area. Contact us." Fails questions 2 and 4: the only difference between pages is the town name. Page B for the same shop covers the one town where the shop actually runs a Saturday pop-up stand, names the stand's opening times, the typical turnaround for that stand and the three most common repairs booked there. Passes. Page C is a "bike repair near stations" page produced for 12 stations the shop has never served. Fails question 1: nobody needs it. The outcome is 6 pages built properly rather than 40 built thinly, and an honest note to the client that service-area pages for places they do not serve will not be written.

A page-level QA checklist

This is the checklist to attach to every AI-assisted page before it goes to the client. Each item has a how-to-verify note so a junior can run it.

On the last row: Google's guidance suggests giving readers context about how content was created, such as background on how automation was used, where that makes sense for the audience. It is a judgement call, not a requirement for every page.

A similarity check you can run in a spreadsheet

For a batch of 30 pages, export the body text and compare each page with its nearest sibling. A crude method works: split each page into sentences, count how many appear (word for word or with one word changed) on any other page in the batch, and divide by the page's total sentences. Say your threshold is 20 per cent. In an illustrative batch of 30 location pages, 11 pages exceed it, all of them written from one prompt with a swapped town name. Those 11 go back for either real local content or consolidation into one page. Use a threshold as a prompt for human review, not as a rule that makes a page safe.

An AI content workflow with sign-off points

This workflow keeps AI where it is strongest, which is research prep, structure and first drafts, and keeps humans on facts and judgement.

1. Brief (human, 20 minutes). Write the user question, the audience, the client facts available, the page's job, and what makes it different from the existing pages. If the brief cannot be filled, the page is not ready.
2. Collect inputs (human, varies). Pricing sheet, photos, staff names the client approves, policies, service limits, real examples. This is the step that creates value. No inputs, no page.
3. Research prep (AI, 10 minutes). Ask the model for the questions users ask and for gaps in the client's current page. Treat the answer as a list of ideas, never as sources.
4. Draft (AI, 5 minutes). Give it the brief and inputs and tell it to use only those facts.
5. Edit and fact check (human, 30 to 45 minutes). A person adds first-hand detail, removes filler, checks every claim against the inputs.
6. QA checklist (human, 10 minutes). Run the table above.
7. Client approval. The client signs off facts and prices in writing.
8. Publish in small batches and watch. Release ten pages, check indexing and engagement for two to four weeks, then continue.

A realistic total is about an hour and a half of human time per page. If a client's budget only allows 20 minutes a page, the honest recommendation is fewer pages, not thinner ones.

The drafting prompt

Write the body of one web page for [BUSINESS TYPE].
User question: [QUESTION]
Use ONLY the facts below. If a fact is missing, write "[NEEDED: ...]" instead of guessing.
Facts: [CLIENT INPUTS]
Do not repeat sentences from these sibling pages: [PASTE HEADINGS OF SIBLINGS]
Structure: answer in the first 50 words, then steps, then what is not included.
No claims about being the best, no invented reviews, no statistics.

Illustrative output for an HVAC contractor's "how long does a heat pump service take" page: "A heat pump service usually takes 60 to 90 minutes. [NEEDED: price for single-zone and multi-zone systems] The visit covers filter cleaning, coil inspection, refrigerant pressure check and a test of defrost mode. It does not cover parts or repairs found during the visit, which are quoted separately." What you would fix: the engineer confirms the 60 to 90 minute figure and the visit contents, and the price gap is filled from the client's sheet. The "[NEEDED]" markers are the feature: they show the model refusing to invent.

Patterns that look like scaled content abuse

Learn to spot these in client briefs and previous agencies' work, because they are the usual route into trouble.

• Town-name swaps. Hundreds of pages with identical copy and a changed place. The fix is fewer pages with real local facts.
• Question farms. A page for every phrasing of the same question. The fix is one thorough page that answers the family of questions.
• Synonym spinning of competitors. Taking other sites' pages and rewording them. Google lists scraping with automated transformations such as synonymising and translating among its examples.
• Glossary dumps. 400 definition pages that add nothing to a dictionary entry.
• Translated duplicates. Machine translation of existing pages into many languages with no review and no real audience for them.
• Comparison pages for products the client does not stock.

Programmatic pages are not banned. A solar installer with a real product catalogue, specifications, stock status and area-specific incentive details can generate pages from a database, provided each one has distinct, accurate information and a reason to exist. The e-commerce version of this is covered in the store-page tutorial.

Shops have one extra rule to build into the checklist. Google's generative AI content guidance says Merchant Center requires AI-generated images to carry the IPTC DigitalSourceType metadata value TrainedAlgorithmicMedia, and AI-generated product titles and descriptions to be specified separately and labelled as AI-generated. If the agency writes product data with a model, the feed needs that label as well as the page needing a reviewer.

A bakery, a bad refresh and a budget sum

A bakery's seasonal pages

A bakery client asks for 60 pages, one per occasion and flavour: "birthday cake", "birthday cake vanilla" and so on. The agency merges them into eight pages organised by occasion, each with the bakery's real order lead times, allergen policy, sizes with serving counts and photos of actual cakes. Each page answers a distinct customer question, and the agency avoids pages that differ by one word.

A refresh gone wrong

An agency ran a model over 300 old tutorials with the instruction "improve and make more engaging". Titles and intros changed, facts did not. Traffic did not move and several pages now carried out-of-date claims phrased more confidently. The lesson: a refresh must change substance. A better brief was "update prices, add the 2026 changes the client confirmed, add one example from the client's own work, remove outdated claims", applied to the 40 pages with the most impressions first.

Quick sum: when volume stops paying

Say a retainer budgets 80 pages a month at 20 minutes of human time each: 26.7 hours. The same 26.7 hours at 90 minutes a page gives about 18 pages. Which set better serves a client depends on the pages, but the 18 can pass the value test, while the 80 cannot. Present the trade-off to the client in those terms.

Monitoring after publishing

Do not watch rankings alone. Check indexing in the Page indexing report. Google says "Crawled - currently not indexed" is a common status that does not necessarily mean a problem, so read it as a batch signal: if most of a new batch still sits there after several weeks while the older, richer pages are indexed, review the batch against the value test before publishing more. Watch engagement, enquiries, and the Search Console Pages and Manual actions reports. Make a short log of what you published, when, and what the pages were meant to do, so that a drop can be traced to a batch.

Google's AI features rely on its core ranking and spam systems. A site with pages judged as spam is unlikely to be a good source for AI answers either, which is why this checklist belongs next to the AI search audit checklist rather than instead of it.

If a client site is already affected

Whether you inherited the problem or caused it, a calm sequence works best.

1. Diagnose. Check Manual actions in Search Console. If there is none, look at the timing of any traffic drop against the publishing log and any known update announcements, then use the Page indexing report to find clusters of excluded or low-engagement pages.
2. Inventory. List every page in the suspect set with URL, publish date, traffic, links and enquiries.
3. Decide per page. Keep and improve, merge into a stronger page with a redirect, or remove. Google's policy says that if you host such content, exclude it from Search, so noindexing or removing is a legitimate route for pages with no value.
4. Improve the keepers properly. Add client facts, first-hand detail and a reviewer. Do not just rephrase.
5. Document. Record the before and after lists so you can describe what changed.
6. If there is a manual action, submit a reconsideration request from the Manual actions report once every listed issue is fixed on every page. Google says a good request explains the exact quality issue, describes the steps taken and documents the outcome. For the aggressive-spam action, whose description names scaled content abuse, it asks for examples of bad content removed and good content added. Most reviews take several days or weeks, and Google asks you not to resubmit before you get a final decision.
7. Wait and monitor. Recovery from an automated demotion is not instant, so do not promise a date. Report progress by indicators, not by promises.

Be careful with sample claims in the report. Do not tell a client that a cleanup guarantees a recovery. Tell them exactly which pages were removed, merged or rewritten and what you will measure.

What to put in the contract

• That content is produced with AI assistance and reviewed by named staff.
• That the client provides facts, approvals and photos, and the timeline depends on them.
• That the agency will not publish pages that fail the value test, even if the quota is not met.
• That batches are released in stages with monitoring between them.
• That the agency will not use bought mentions or fake reviews. Google says seeking inauthentic mentions is not as helpful as it might seem, and its AI features depend on its spam systems as well as its ranking systems.

Putting this in writing at the start turns the policy into a service standard. A client who pushes for 200 thin pages is easier to refuse when the contract already says no. For agencies building the process across a team, the marketing agency rollout plan shows how to stage a change like this.

A short script for the client conversation

When a client asks "can't AI just write them all?", answer like this: "AI can write the first draft quickly. Google's policy is about pages created in bulk without adding value, so what protects you is what we add: your prices, your photos, your process, checked by a person. We'd rather publish 20 pages that each help someone than 200 that read the same. Here's the list of the 20, and here is what we need from you for each." Then show the page-level checklist so they can see the standard.

Scaled content abuse questions agencies get

Is AI-generated content against Google's spam policy?

Not by itself. The policy targets many pages generated mainly to manipulate rankings without helping users, and it applies however the content is made. Google's guidance says generative AI can help with research and structure. The risk is volume without added value, not the use of a model.

How many pages is too many?

Google gives no number, and neither should you. The test is purpose and value: does each page exist because a user needs it, and does it add something the others do not? Ten near-identical pages can be a problem while a thousand genuinely distinct product pages are not.

What should an agency do if a client site has a manual action?

Open the Manual actions report in Search Console, read what Google says, remove or noindex the offending pages or rewrite them properly, then request a review from that report, explaining the issue, the steps taken and the outcome. Keep a before-and-after list of URLs, and do not send a second request before Google decides on the first.

Can I use AI to refresh old pages at scale?

Yes, if each refresh adds checked facts, current information or the client's own experience, and a person reviews it. Mass-rewriting hundreds of pages with the same prompt, changing wording but not substance, is the pattern the policy describes. Stage the refresh and sample-test it.

Further reads

• Does llms.txt Work? What SEO Agencies Should Tell Clients — Another AI-era tactic to question before selling.
• How to Choose an SEO Agency That Understands AI Search — How clients judge agencies on this very point.
• Does Writing Etsy Listings With AI Hurt Your Search Ranking? — A small-shop view of AI-written listings and ranking.
• Making a Website Using AI With Good SEO: A Step-by-Step Build — Building a whole site with AI without thin pages.
• Search Console and AI Overviews: What Agencies Can Report to Clients — Report on AI search data without overclaiming.
• Google AI SEO Agency: Questions to Ask Before You Hire One — A buyer's checklist for hiring a Google AI SEO agency: proof to demand, reports to expect, AI Overview tracking, contract clauses and red flags.
• AI Keyword Clustering for SEO Agencies: Validate Clusters Before Sign-Off — Embed or prompt-cluster a keyword list, validate a sample against search results, and send the client a map they can approve, with costs per 10,000 keywords.
• AI Link Building Outreach: Personalised Without Spam — A compliant outreach routine for SEO agencies: research prospects, pick one angle each, send three emails at most, and check Google's spam policy and email law.
• AI Schema Markup Generator: Validate Client Structured Data Before Launch — A pre-launch routine for AI-drafted JSON-LD: prompts that stop invented data, then validator, Rich Results Test and Search Console checks.
• Best AI for SEO Content: A Refresh Workflow for Client Articles — A step-by-step refresh workflow for client articles: spot decay in Search Console, update with an assistant, fact-check, and publish without losing the URL.
• Content Refresh With AI: Update Old Articles and Track Rankings — A client-by-client method for choosing old articles, refreshing them with AI, and measuring the ranking change against pages you left alone.
• SEO Agency AI: Which Jobs to Hand to AI and Which to Keep Human — A practical agency workflow showing where AI drafts help and where client judgement, evidence checks and approval stay with the team.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Google Search Central spam policies (scaled content abuse section, page last updated 28 August 2026), Google's guidance on using generative AI content, Google's guide to optimising for generative AI features, and the Search Console help pages for the Manual actions and Page indexing reports.