First, compare AI-assisted pages with the site's best work
Separate the AI-assisted pages from the rest using the inventory or the publishing log, then pick a comparison group: the site's best human-led pages on similar topics and of similar age. Read both groups on the same measures: impressions and clicks from search for their main queries, engaged reading, return visits, and the next steps readers take. If both groups perform alike, the problem is not specific to AI and the other pages in the content marketing troubleshooting hub are the better starting point. If the AI-assisted group trails, read a sample of its pages beside the comparison pages, as a reader would.
That reading usually reveals the fault faster than any report. Ask of each AI-assisted page what it says that the pages already ranking do not, whether every claim in it is true today, whether it sounds like the organisation, which question it answers for whom, and how anyone was meant to find it. Google's statement on AI-generated content is the reference here: its systems aim to reward original, high-quality, people-first content however it is made, and automation used mainly to manipulate rankings falls under its spam policies. Those are the tests the pages are failing, so the diagnosis follows them.
Diagnosis table
Diagnosis
| Symptom | Check | Likely cause | Fix |
|---|---|---|---|
| Pages index but rank below thinner competitors and earn little reading | What each page offers that the top results do not | No first-hand experience or original information | Add the team's own evidence, cases and judgement |
| Dozens of pages with the same structure, differing only in a keyword or place name | Similarity across the batch; whether each page answers a distinct question | Near-identical pages at scale | Merge into fewer, distinct pages; stop templated batches |
| Readers or customers point out wrong figures, rules or features | Claim-by-claim check of a sample against current sources | Factual errors left unchecked | Correct, source and re-review every flagged page |
| Pages read smoothly but nobody could tell which company wrote them | Read a paragraph aloud next to a strong human-led page | A generic voice | Edit to the documented voice principles |
| Pages answer a broader or different question than searchers ask | The brief behind the page; what the results page for the query rewards | Briefs without intent or angle | Brief each page with its question, reader and angle |
| New pages get almost no visits in their first weeks, from any channel | Where each page was sent on publication | No distribution | Give every page a distribution plan before it goes live |
Fix, verify and prevent each cause
Nothing first-hand in the draft
An AI draft can only rearrange what has already been written, so a page built from one has nothing that the pages it learned from lack. Google's helpful content guidance asks whether content provides original information, research or analysis and shows first-hand experience; a pure draft answers no.
Fix: add what only the organisation has, such as tests it ran, cases it handled, data it holds, photographs it took and the expert's actual recommendation, following the on-page signals in the E-E-A-T guide, and give the page a named author who stands behind it. The expert interview tutorial shows how to get that recommendation out of a busy expert and onto the page.
Verify: the revised pages' position and clicks for their main queries improve against the unrevised ones over the following months.
Prevent: require each brief to name the first-hand material the piece will contain before a prompt is written, and drop the slot if there is none.
Near-identical pages at scale
The speed of drafting tempts teams to publish a page for every keyword variant, city or product combination, each the same template with a few words changed. These pages compete with each other and give readers no reason to prefer any of them.
Fix: group the batch by the question each page actually answers, merge pages that answer the same one into a single stronger page with redirects, and keep separate pages only where the content genuinely differs; the AI-assisted content guide warns that one prompt reused across topics produces the same piece on every topic.
Verify: after merging, the surviving pages hold or gain the combined impressions of the pages they replaced.
Prevent: cap how many pages a single prompt template may produce and require an editor to approve any batch.
Errors nobody caught
Drafts state outdated rules, invented figures and wrong product details with the same confidence as true ones, and a single error found by a customer undermines every other page.
Fix: check the existing AI-assisted pages claim by claim against sources a reviewer can open, correct or remove what fails, and apply the accuracy standards and named roles in the content governance guide, including how AI-assisted pages are flagged in the inventory.
Verify: a second sample check finds no unsourced or wrong claims, and corrections requested by readers stop arriving.
Prevent: give every AI-assisted page a named accountable owner and a review date, and audit a sample each quarter to see whether these pages age differently.
A voice that could be anyone's
AI drafts converge on a smooth, even register with balanced lists and hedged conclusions. Readers sense that nobody in particular is speaking and give the page less trust and attention.
Fix: edit every draft to the voice principles and "this, not that" examples set out in the brand identity and voice guide: state a preference where the draft listed options, cut claims that are not true here, and use the organisation's own vocabulary.
Verify: colleagues reading a paragraph blind can tell it is yours, and engaged reading on edited pages rises against unedited ones.
Prevent: include a short sample of house writing in every prompt and keep a record of how much each content type needed editing.
Briefs that never fixed the question
A prompt built from a topic produces a page about the topic, not an answer to a specific reader's question, and it often misses the format and depth the query calls for.
Fix: brief each page with the reader, the stage, the single question in the reader's words, the angle and the next step, as the content brief tutorial describes, and record the query's Search IntentThe goal behind a search query: to learn, to find a site, to compare or to act. from the results page before drafting.
Verify: the revised page answers its question in the opening, and its click-through and engaged reading for the main query rise.
Prevent: do not allow a prompt to be written from anything less than an approved brief.
Published and left
A faster drafting step tends to raise the number of pages published without raising the effort spent on reaching anyone, so each page gets an announcement or nothing.
Fix: give each page a distribution plan from the content distribution channels guide: internal links from pages already read, a newsletter mention for the relevant segment, adapted social posts and a genuine answer in the communities where the question is asked, each ending in a CTAThe prompt that tells a reader the one action to take next, such as a button or link. suited to the channel.
Verify: referred and internal visits to the page grow over its first weeks instead of stopping on day one.
Prevent: write the distribution notes into the brief and publish fewer pages if the team cannot distribute them all.
A prevention routine for AI-assisted content
AI changes the speed of drafting, not the standard a page has to meet, so the routine protects the standard at the points where speed tempts teams to skip it. Before drafting, every page has an approved brief that names its question, its reader, its first-hand material and its distribution, and no prompt is written without one. Before publication, the page passes the existing review gate, with a reviewer who did not write the prompt and an expert who puts their name to the result; signals of E-E-A-TExperience, Expertise, Authoritativeness, Trust. come from that person, not from the tool. Monthly, compare the AI-assisted group with the human-led comparison group on the same measures, so any gap shows early rather than after a year of publishing. Quarterly, sample the AI-assisted pages for accuracy, sameness and voice, merge any near-duplicates that crept in, and adjust which content types use AI drafting according to the record of how much editing each needed. Treat the hours saved in drafting as a budget for evidence, editing and distribution, not as capacity for more pages.
To stop the same causes reaching new pieces, run each AI-assisted draft through the AI-assisted content review checklist before it is published.
An accounting firm's batch of AI guides
An accounting firm published 85 AI-drafted guides in two months, one for each tax question its keyword tool listed. After four months they earned 1,900 search clicks a month between them, against 7,400 from the firm's 22 older partner-written guides. A sample read found that 37 of the guides answered the same 11 questions in slightly different words, 9 quoted allowances from a previous tax year, and none contained an example from the firm's own work. The firm merged the overlapping guides into 11, corrected every figure against the tax authority's pages, added a worked client scenario and a named partner to each surviving guide, and sent each one to the relevant client newsletter segment. Six months later the remaining 59 guides earned 5,200 clicks a month.
Common mistakes
Publishing every draft the tool produces. Mistaking smooth prose for checked prose. Using one prompt template for every topic. Adding an expert's name to pages the expert never read. Measuring success by pages published. Deleting the whole batch instead of sorting it by cause.
