Fractl ResearchSeptember 20266 AI assistants · 12 industries · 14,400 blog ideas, checked against Google
Ask an AI assistant for blog ideas and the first ones back are the ones it gives everyone. Search for them and most are already on Google. Here is what we found, and the one-search habit that fixes it.
If you use AI for blog ideas
Ask an AI assistant for twenty blog ideas for a roofing company and the first one back is usually "signs you need a roof replacement vs. repair." We asked six assistants (Claude, ChatGPT, Gemini, Grok, DeepSeek and Meta's Llama), and asked each one ten different ways, using real wordings people type: sixty lists in all. Nineteen of them opened with that exact idea. Then we searched for it. Every result on Google's first page was already that post.
That is the whole study in one example. We repeated it for twelve kinds of business, from a mattress store to a mortgage lender: 720 lists, 14,400 ideas. Then we ran 300 real Google searches in those industries and looked at what was already on the first page. Two things came back. The ideas at the top of any list are the ones the assistant gives everybody. And the more assistants agree on an idea, the more certain it is that Google already has it.
One thing we are not saying: that a page matching an idea means the page copied it, came from an AI, or that the assistant took it from the page. A match means the idea is the same. That is all it means.
Ask one assistant ten ways and its 200 ideas boil down to about 70 different posts. That is fine; a topic only has so many angles. What matters is which ideas repeat, and the answer is: the ones at the top.
How often an idea came back no matter how we asked · by its position in the list
"Came back" means the idea appeared on at least half of the ten lists we got from that assistant for that business. Every one of the six assistants shows the same slope.
Rewording the request barely changes this. Each assistant opened with its own favorite idea in a third to nearly half of its replies, whatever the wording. Only one wording changed the ideas: asking for "ideas that could earn links," meaning ideas other sites would link to. Even that changed the kind of idea, toward tools and original data. The article angles stayed the same.
If you read the first five ideas and stop, you have read the part it hands to everyone.
Popular ideas are bound to be written already, so a fair test has to pair each shared idea (suggested by most assistants) with a rare one (suggested by only one) of the same type: a how-to against a how-to, a cost guide against a cost guide. We did that for six pairs in each industry and searched Google for every idea. Then an AI model read each of the top five results in full and decided whether that page was already that post.
| Ideas that were… | a post like it already in Google's top 5 | the same post already in the top 5 |
|---|---|---|
| suggested by 4 or more of the 6 assistants | 80% | 33% |
| suggested by only one assistant | 64% | 15% |
66 pairs across 11 industries. "A post like it" means the same subject with an angle close enough that a content planner would say "this already exists." About a third of pages would not load for our automated reader, so the true figures in both rows are higher.
Two things to take from that table. Most of what any assistant suggests, shared or rare, is already written. And the ideas the assistants agree on are twice as likely to already exist as the same post.
Looking from Google's side gives the same answer. Of the articles on the first page for the 300 real searches, roughly a third are, in substance, a post the assistants had suggested.
Agreement between assistants is not a sign of a good idea. It is a sign that someone already wrote it.
The six assistants overlap less than we expected. Any two of them share only about a fifth of their ideas, so pooling several gives you a list two to three times longer than any one of them. But the ideas that only one assistant gave were already written six times in ten. Pooling gets you more ideas. Most of the extra ones are already written too.
People who watch search results for a living know that Google's first page has been filling with forum threads and definitions for a couple of years. The point here is what the assistants do about it: nothing. They suggest almost none of those.
Of the 300 top spots in our searches, 128 were held by something other than an article. (Some pages hold the top spot for more than one search, so those 128 spots belong to 85 pages.) They were:
What held the #1 spot when it wasn't an article · 85 pages
36 of the 37 forum threads were on Reddit. "Definitions" are dictionary-style "what is X" pages, usually from a manufacturer, a standards body or a government site.
Which kind wins depends on the question. "What is" questions go to a definition. "How much" and "how to" questions often go to a forum thread or a calculator.
The assistants are not blind to these formats. Ask for link-worthy ideas and they propose data studies and free tools. Almost nobody builds them: about one in eight of the data-study ideas they proposed had anything like it on the first page.
The assistants hand out the idea. Almost nobody executes it.
The list is where everyone started. The work is what you add to it.
Run this check on your own list
A version of the check we ran for this study now runs inside the Ideation Stress Test in Fractl Agents, and you can run it once for free, no account needed. Paste up to twenty ideas (ten on the free run). It searches each one, reads the top five results in full, and asks ChatGPT, Gemini and Claude for ideas for the same business. You get back a leaderboard: which ideas already exist as the same post, which ones every assistant hands out, what kind of page holds the top spot for that search, and which ideas are still unclaimed. Try it on your list →
Two limits worth knowing: we used the assistants through their programming interfaces, not the chat apps, and we had no group of people brainstorming ideas to compare against, so we cannot say whether a human list would do any better. Everything else we measured, and everything we got wrong on the way, is below.
The ideas. Ten real prompts, taken from our own chat logs and from marketing forums (for example, "give me 20 content ideas about X" and "I run marketing for X, what should we blog about this quarter?"), sent to six assistants for twelve kinds of business: eight picked at random from a standard list of industries (roofing, an online mattress store, project-management software, trucking, commercial cleaning, an Italian restaurant, heating-and-cooling, a coffee roaster) and four from our client work (a mortgage lender, a staffing firm, a small-business lender, a prepaid electricity provider). 720 lists of 20 ideas. Two ideas count as the same if they would be the same post: same subject, same angle, same reader. An AI model did that counting; we checked it three ways: forty copies of one idea came back as one; ideas planted from a different industry never merged into the wrong pile; grading the same list twice with the items shuffled agreed 98% of the time. A second grader from a different company disagreed on exact counts by up to a third, so we quote "about 70" rather than a precise figure, and we publish no table of which assistant is most repetitive.
The Google side. For each industry, 25 real questions people search, taken from search data rather than from the assistants' ideas (typically about 1,300 searches a month). Google's first page now holds seven to nine ordinary results. That gave 300 searches and 1,768 distinct pages. To decide whether a page was one of the assistants' posts, two separate AI graders each saw the page's title and summary and the ten most similar suggested ideas, and both had to name the same idea. We then had sixty of those matches read in full by a stronger model: about a quarter were the same post outright, about three-quarters were the same post or one like it. That is why the body says "roughly a third" of first-page articles rather than the graders' raw 51%.
The paired test. For each industry, six ideas that four or more assistants had suggested and six that only one had, chosen at random but matched by type of post (how-to, cost guide, comparison, list, explainer, checklist). One Google search per idea, the top five results read in full. The mattress store is missing from this test because its cross-assistant idea list could not be completed, so it covers 11 industries and 66 pairs. A gap the size of 80% against 64% would arise by chance about one time in twenty; 33% against 15%, about one in a hundred.
Within the first page, position was a coin flip. Pages matching an assistant's idea sat at an average place of 4.5; pages that weren't, 4.6. Compared within the same search results, the assistant-idea page ranked higher 68 times and lower 64. This says nothing about how a page gets onto the first page; it only says that once there, matching an assistant's idea did not predict its spot.
Neither did who wrote it. We ran every article we could fetch through our own detector, software that estimates whether text was machine-written, on the page's body text only. Three in ten read as AI-written. That did not move a page up or down either. Local-business pages read as machine-written most often (nearly four in ten, against two in ten for government and university sites), and local-business pages sat lowest on the page of any kind of site. We are not publishing a year-by-year trend, although our data shows one, because we have not yet measured how often the detector wrongly flags writing that predates language models.
Rewording only changed the kind of idea. Requests that asked for the same thing in different words produced the same lists. "Ideas that could earn links" produced tools and data studies instead of articles, and those tool and data-study ideas were rarely already built (the one-in-eight above), but compared with other tool and data-study ideas, the link-earning wording found no more unclaimed ground.
Our first headline number was wrong. Our first pass at matching Google's pages to the assistants' ideas said 62% of the first page matched an assistant's idea. A hostile review found the grader was matching from titles and short snippets, was allowed to pick from a dozen candidate ideas at once, and had no way to say "this is not an article," so forum threads and product pages were being counted as blog posts. We rebuilt it as described above. The number that survived is roughly a third of first-page articles, and the paired test, which has a control built in, carries the headline instead.
An early version of the shared-versus-rare gap was six times. Part of that was a grading artifact: an idea suggested by six assistants had six chances to be matched. The paired test fixes this and gives two times.
We dropped estimates of how many ideas each assistant "knows" in total as statistically unsound.
We never proved who came up with an idea or who wrote a page. Google's page dates could not tell us whether the page came first or the idea did, because Google usually shows the date a page was last updated.
The study cost about $80 in model calls and search queries, plus three rounds of adversarial review that cost more than the study.
Fractl Agents research · September 2026 · v1.0 — first published 5 September 2026 · Disclosure: Fractl sells content marketing and digital PR services and builds an AI platform for marketing work; the check described above is one of its tools.