Skip to main content

Why AI-written pages keep repeating the same answers

Research on AI writing, what Google asks of content, and how a customer question and a checked source change the brief.

Andrii DanivSeptember 30, 20266 min read

A brief that asks for a summary of the pages already ranking gives the writer very little to add. AI can reorganize that material into readable prose. The customer may still be missing the answer they came for.

The frustration behind “Why does every page sound the same?” goes beyond repeated phrases. A page can use different words while offering the same advice, skipping the same conditions and leaving the same buying question unresolved.

I’d decide which customer question the page should settle and what we can verify about the answer before asking AI for a draft. Research on AI writing gives us a reason to pay attention to both the speed and the substance.

Writing faster and adding something different are separate checks

Professional writing tasks

40%less completion time
18%higher assessed quality

In a 2023 experiment published in Science, Shakked Noy and Whitney Zhang assigned professional writing tasks to 453 college-educated participants. Access to ChatGPT reduced average completion time by 40% and increased assessed output quality by 18%. Those are results from the writing tasks studied, rather than a forecast for a content team’s traffic or sales. [1]

8-sentence stories

A 2024 study by Anil Doshi and Oliver Hauser found a more complicated pattern. Participants who could request AI story ideas produced short stories rated as more creative, better written and more enjoyable. The AI-assisted stories were also more similar to each other. [2]

What this experiment can tell us

That experiment used 8-sentence fiction and limited interaction with the model. It does not establish that all AI-assisted marketing content becomes alike. It does show why a better individual draft can coexist with less variety across the work being produced. [2]

My practical concern is the assignment. When the source material and questions stay the same, asking for a more distinctive tone leaves the underlying answer largely untouched. I would check what the page contributes before spending another editing pass on its personality.

Google asks what the page adds

Google’s people-first content guidance asks whether a page contains original information, reporting, research or analysis. When it draws on other sources, Google asks whether it adds substantial value rather than merely rewriting them. That gives editors a more useful review question than whether the wording looks human. [3]

Google’s separate guidance on generative AI describes useful roles for it in research and structuring original material. It also warns about generating many pages without adding value for readers. Using AI is not, by itself, the violation described there. [4]

What the spam policy covers

The scaled content abuse policy concerns producing many pages primarily to manipulate rankings rather than help users. It applies regardless of whether people, automation or a combination created them. A human rewrite of the same unhelpful material does not solve that policy problem. [5]

For a proposed page, I’d ask what someone could understand or decide after reading it that the existing results leave unclear. If the answer is just that ours will be longer, I would keep working on the brief.

A customer question changes the assignment

Read the brief at each stage

“Write a page about CRM migration.”

Consider an illustrative brief for a fictional CRM migration service. “Write a page about CRM migration” leaves room for familiar sections about planning, data quality and training. The writer can complete that assignment while leaving a buyer’s practical concern unanswered.

Can our fields and order history move?

Now give the writer a specific question: “Can we move our custom contact fields and order history into the new CRM?” The page needs details about the systems involved, supported fields and any records that require separate handling. A general promise about an easy migration will not answer it.

An illustrative customer question for the fictional migration service.

Check the supported fields and exceptions.

The company still has to establish those details. A precise question can expose a gap in the available information. That is useful before publication, while someone can ask the implementation team and check the import documentation.

A source request, not a claim about any real product.

For a real brief, I’d look for that wording in a permitted sales-call note, a support conversation or the questions customers send before a quote. Remove private information before putting those materials into a writing tool. Search results can help identify competing answers; customer questions keep the assignment tied to an actual decision.

Gather the answer before asking for the draft

Check the source

The source can be modest. An approved product document, a recorded test or a clear explanation from the person doing the work may supply the detail the page needs. A large original survey is not a prerequisite for answering a narrow question.

For the migration example, I would ask for the supported-field list and the exceptions. Then I would write a provisional answer and mark anything still unconfirmed. The draft should carry those limits through to the published page.

Draft against it

AI can help organize that material, suggest clearer sentences and point out questions the brief leaves open. I would give it the source passages alongside the assignment and explicitly tell it to flag missing information. Instructions to be accurate still need checking against the resulting draft.

Keep the conditions

A claim about a tested feature should retain the tested conditions. A result from a particular project should retain its scope. When the source only supports a partial answer, publish that boundary or complete the investigation before making a broader promise.

Review the claims as closely as the wording

For a claim review, take this fictional sentence: “We migrate every customer record in 24 hours and keep all reporting intact.” The words “every,” “24 hours” and “all” each need supporting evidence. Making the sentence sound more natural would leave those obligations unchanged.

Which part needs evidence?

Open a phrase from the fictional sentence to see what needs checking.

“Every customer record”

Which record types, fields and attachments were actually tested? Keep exclusions visible.

“In 24 hours”

When does the clock start? Does the timing include preparation, validation and corrections?

“All reporting intact”

Which reports were compared, and which dependencies or calculations still need checking?

I would ask which records were tested, what the timing includes and which reports were checked. If the source cannot answer, narrow or remove the claim. Keep the source beside the draft while reviewing so a polished sentence does not gradually become a bigger promise.

Then do a separate wording pass. Remove introductions that postpone the answer, generic praise and headings that could sit on any competitor’s page. Keep the terms customers use when they are accurate. A technical condition can be necessary even when it makes the paragraph less elegant.

How this connects to Google’s AI features

This matters for AI search too. Google says the usual SEO practices still apply to AI Overviews and AI Mode, with no special schema required for those features. A page must be indexed and eligible for a search snippet to appear as a supporting link; eligibility does not guarantee inclusion. [6]

Put the saved drafting time back into the page

For the next content assignment, I’d ask for the customer’s question, the evidence available and the answer the business can currently support. That brief lets an editor see whether drafting can begin or whether a product specialist still needs to resolve something.

After publication, check whether the page reaches the intended searches and supports the relevant next action. A demonstration request, a qualified enquiry or a purchase may be appropriate, depending on the page. Draft volume alone will not tell you whether the answer was useful.

Keep the question open when results are weak. The page may need better distribution, clearer product details or a different answer. An attractive layout and cleaner sentences can help someone read it, but they cannot establish a missing product fact.

AI can shorten the drafting work. I would use the time available to check the answer, add the relevant detail and remove what the reader does not need. Before commissioning another batch of pages, ask to see the customer question and the source behind each proposed answer.

Which question is missing from your content plan?

See how Etavrian researches, writes and publishes the pages your buyers need.

See Content Generation
Back to the blog