Ask AI for blog ideas for a roofing company, and it may suggest helping homeowners decide between repairing and replacing a roof. That’s a reasonable topic. But if the top search results already cover it thoroughly, your article needs to offer something those pages don’t.

At Fractl, we asked AI assistants for blog ideas across 12 industries, then searched Google to see whether articles covering those ideas already existed. Many did. The ideas that several assistants suggested were the most likely to match an article already on page one.

Before you start your next article, search for it and see what people can already find. The results can help you decide which ideas to develop, what evidence to add, and when something other than an article would be more helpful.

AI Repeated the Same Roofing Idea Across 19 Lists

We asked six AI assistants (Claude, ChatGPT, Gemini, Grok, DeepSeek, and Meta’s Llama) for blog ideas for a roofing company, using different versions of the request. Of the 60 lists they returned, 19 began with the repair-versus-replacement idea. When we searched for it ourselves, every result on the first page already covered that question.

We then repeated the test across 12 industries, from mortgage lending to commercial cleaning, and collected 14,400 ideas. For many of them, we searched Google to see whether the top-ranking pages already covered the same article. Roofing is one example of what we found.

That doesn’t mean the roofing topic was automatically the wrong writing topic. A company could still have a useful local example or information missing from the existing guides. But another general repair-versus-replacement article wouldn’t be the right one.

When Several AI Tools Agree, Check for Existing Articles

We compared ideas suggested by at least four assistants with ideas suggested by only one. To keep the comparison fair, each pair covered the same kind of article, such as one cost guide against another. That way, a difference in format couldn’t be mistaken for a difference in originality.

Across those pairs, 80% of the widely suggested ideas had a close or exact match in Google’s top five results. Among ideas suggested by only one assistant, 64% had a match.

For exact matches alone, the figures were 33% and 15%, respectively. A close match covered a similar subject and angle; an exact match amounted to the same proposed article.

The comparison covered 66 pairs across 11 industries. Our automated reader couldn’t open about a third of the pages, so the real match rates could be higher or lower.

Getting the same suggestion from several assistants doesn’t establish that it’s worth producing. 

Search the idea and decide what your version would add. 

Even a suggestion that appears on only one AI list still needs that check.

We measured how closely ideas overlap with pages that already exist. We can’t say who thought of an idea first, whether a page was written with AI, or whether using an AI suggestion affects how a page ranks.

Read the Whole List Before Choosing an Idea

The first five suggestions were also the most repetitive. We sent each assistant ten versions of the same request for each business. More than half of the first five suggestions showed up in at least half of those responses. Near the end of the lists, about a third did.

Don’t stop reviewing after the opening suggestions. Read the rest, then check the promising ideas against published work. A less-repeated suggestion might still describe another article that’s already out there.

Trying another assistant can add options. Any two assistants in our test shared about a fifth of their ideas. Pooling their suggestions expanded the list, but we didn’t show that every additional idea was original.

Keep track of the questions each suggestion would answer. 

If several titles lead to the same article, combine them before choosing what to research.

Check Whether an Article Is the Right Format

In a separate set of 300 Google searches, the top result wasn’t an article 128 times. Those results included discussion threads, definition pages, and calculators. Some pages ranked first for more than one search.

Before asking a writer for a draft, open the results and consider what people need to do. A person comparing roof-replacement costs might need an estimate based on their own home. Another might want homeowners’ accounts of unexpected charges. A general cost guide won’t necessarily satisfy both needs.

Use what you find to choose a format. For example, consider these options:

Explain a decision Write an article when people need to understand options, risks, or steps before choosing what to do.	Estimate a cost Consider a calculator when the answer depends on information people can enter about their own situation.
Report a finding Develop original research when you can collect evidence that answers a question the existing pages leave open.	Show firsthand experience Use interviews or documented examples when people need to understand what happened to someone facing a similar decision.

Choose the one you can support with useful information and maintain.

Search Your Shortlist Before You Brief a Writer

Check each promising idea while you’re still deciding what to commission:

Search the question behind the headline Use the words someone would type when looking for an answer, rather than only the title AI suggested.	Read the leading results Compare their answers and examples with what you planned to cover. A different headline doesn’t necessarily mean a different article.
Identify what you can contribute Name the records, interviews, or research your team could use to answer something those pages don’t.	Decide what to produce Keep the idea if you can make a useful contribution. Change the question or format if the existing results already do the job.

For example, say the roof-replacement guides you read explain warning signs and broad costs, but don’t show how actual quotes differ. If you have permission to analyze recent estimates, you could investigate which charges account for those differences.

That gives you a research assignment: compare documented quotes and explain the differences you find. It doesn’t give you permission to assume a result, and it doesn’t establish that the question is new. You still need to search for that specific comparison.

Give AI Evidence It Didn’t Have in the First Request

Your company’s records and customers’ questions can help you move beyond the initial list. Supply information you’re allowed to use, then ask AI to suggest questions it could help you investigate.

For the roofing example, you could provide an anonymized summary of estimates and ask: “What questions could we answer about unexpected replacement costs using these records? For each question, identify the information we already have and what we’d need to collect.”

Review the answer against the records yourself. AI might propose a comparison the data can’t support, such as differences between cities when the estimates all came from one area. Remove those ideas or collect the missing evidence before drawing conclusions.

Our test also found that asking for ideas other sites might link to prompted more suggestions for tools and original research. The wording changed the kinds of projects proposed; it didn’t establish that those projects were more original than comparable tool and research ideas.

Asking for a study is only the beginning. Someone still has to choose a useful question, collect reliable evidence, and explain the result.

How We Checked the Suggestions

We sent ten versions of a blog-idea request to each assistant for each industry. AI reviewers grouped suggestions that would lead to the same article, meaning the same subject, angle, and intended reader. We checked the grouping by feeding in duplicate and shuffled lists, but reviewers still disagreed on some counts.

For the comparison between widely suggested and single-assistant ideas, an AI reviewer read the top-five pages it could open and rated how closely each matched the idea. We left one industry out of that comparison because we couldn’t finish its combined idea list.

We ran the test through the assistants’ developer tools rather than their chat apps, and we didn’t compare AI lists with ideas people came up with. So we can’t say a human brainstorm would do better, or that these checks improve rankings or media coverage.

Give Your Next Article a Reason To Exist

Choose an idea from your next brainstorm and write one sentence describing what you’ll add to the existing results. Be specific about the information you’ll collect or the question you’ll answer. If you can’t name that contribution yet, keep developing the idea before assigning the draft.

This research informs how we assess campaign ideas in Fractl Agents, where suggestions can be checked and refined before your team chooses a direction. 

If you want help turning a promising question into original research and a story worth pitching, reach out and partner with Fractl to develop the campaign and bring it to relevant journalists.