AI search cited a made-up concept within a week. Here's what that means for your real facts.

Cloud AI cites who it can see. A new paper shows the uncomfortable half of that sentence: it will also cite things that aren't true, if they show up in the places it already looks.

The paper is From Public Posts to AI-Search Citations: Measuring the Fragility of AI Search (arXiv 2610.11932, submitted October 8, 2026; HTTP 200 this fire), by six researchers at Fudan University. The paper's header lists the NDSS Symposium 2027. They invented a concept called the "Cactus Gravity Protocol," wrote posts about it with planted details, and published them from ordinary accounts on sites the AI tools already cited often. Then they asked the tools about it every day for a week.

"8 of 10 platforms cited a fabricated concept within seven days."

Source: arXiv 2610.11932, abstract (v1, October 8, 2026).

The four English-web tools in the study, ChatGPT, Perplexity, Grok, and what the paper calls "Google AI," were all among the eight. Search Engine Journal's write-up (AI Search Tools Cited A Made-Up Concept Within A Week, Paper Finds, October 19, 2026, HTTP 200 this fire) is a good plain-language read.

I'm not writing this to show you how to game AI search. The study is a security paper. I'm writing it because the same mechanism that let a fake concept in can let a wrong fact about your business in, and most businesses aren't checking.

The short version

• AI search leans on a small set of sources. Across 10 tools, the researchers logged 17,211 citations over 6,356 domains. Each tool's top 20 domains took between 20.5% and 70.8% of its citations.
• Many of those sources accept posts from anyone. 15 of the 22 publishing platforms they tested had low or medium barriers to sign up and post.
• A made-up concept got in fast. 8 of 10 tools cited it within seven days.
• Where the post lived mattered more than how many posts there were. One post on a platform the tools already favored beat 20 matched posts on a platform they rarely cite. 40 posts on a self-hosted WordPress site got zero citations in a week.
• Deleting it didn't clear it right away. After the researchers deleted every post, some tools kept citing them for days.
• The made-up product mostly didn't land. None of the four English-web tools cited the fake keyboard within the week. The paper tested no real brands.
• Monitoring is not execution. Knowing AI can be fooled doesn't fix a wrong answer about you. Correcting the source does.

What the paper actually found

Finding	Where (opened this fire)
10 AI-search tools; 17,211 citation instances over 6,356 unique source domains; top-20 domains captured 20.5–70.8% of each tool's citations	arXiv abstract
15 of 22 tested publishing platforms tied to cited domains had low or medium barriers for account setup and posting	arXiv abstract
Citation study queries ran in March 2026; publishing experiments in April 2026	arXiv full text (HTML v1)
Reddit ranked first among cited sources on Google AI and Grok, YouTube on Perplexity; ChatGPT's top source was Wikipedia, at 3.2% of its citations	arXiv full text
8 of 10 tools cited the fabricated concept within seven days	arXiv abstract
Across all tools and topics over the week: the posts were cited in 176 of 1,004 answers (17.5%)	arXiv full text
After each tool first cited the fake concept: Grok cited the posts in 93.7% of answers (n=32), Google AI 45.5% (n=33), Perplexity 37.1% (n=35), ChatGPT 35.0% (n=20)	arXiv full text, Table III
Planted details showed up in the answer text less often: Grok 53.3%, ChatGPT 10.0%, Google AI 6.1%, Perplexity 0.0%	arXiv full text, Table III
No English-web tool cited the fabricated product (a keyboard) within the window. Among English-web tools, only Perplexity cited the opinion posts (11.4%, n=35)	arXiv full text, Table III
One post on a favored platform beat 20 matched posts on a rarely cited one. 40 posts on self-hosted WordPress: no citation in a week. The authors call these "per-topic thresholds, not a general conversion rule"	arXiv full text
Posts deleted after Day 11; "Deletion reduced citation counts but did not fully clear them"	arXiv full text
The researchers disclosed to all 10 platforms. ChatGPT, Perplexity, and Google AI were among those that "acknowledged receipt"	arXiv full text

Why this matters for a business that isn't trying to game anything

Most owners read a study like this and think, "That's someone else's problem." Think about these instead:

• A former employee's LinkedIn post that lists the wrong service area.
• A three-year-old forum thread that says you don't do weekend calls.
• A directory listing with old hours.
• A competitor's comparison page that gets one of your features wrong.

None of those is an attack. They're just wrong, and they're sitting on exactly the kind of sites the paper found AI tools lean on. If an invented protocol can get into answers within a week, a stale fact about you can stay there for a long time.

The paper also says something about your own website, and you won't love it. A brand-new self-hosted site with no history didn't get cited within a week, even with 40 posts. That doesn't mean your site doesn't matter. It means your site isn't automatically the source the AI picks for facts about you. You have to make it the clearest source, then make the places the AI already reads agree with it.

The checklist: keep the facts about you true where AI looks

1. Find out where each engine looks for your category. Write down the 10 questions buyers actually ask before they hire you. Run them in ChatGPT, Perplexity, Gemini or Google's AI Mode, and Grok. Log every cited domain by engine and date. The paper found each tool has its own favorite sources, so one engine's list won't tell you about another's. That's the paper's citation-mapping method pointed at your own category, not something the authors prescribed for businesses.

2. Put every important fact on your own page first. Hours, service area, services, credentials, policies, who you serve and who you don't. One fact, one clear page, current. If someone corrects a wrong claim about you, your page is the evidence they point to. Google's AI optimization guide (HTTP 200 this fire) says creating "unique, compelling, and useful" content "will likely influence your website's presence in generative AI search in the long run more than any of the other suggestions in this guide."

3. Make the profiles you control agree with that page. If the engines in your category cite LinkedIn, YouTube, a directory, or a review site, make sure your own legitimate profiles there are accurate and link back to the page from step 2. This isn't posting to win citations. It's keeping your own accounts from contradicting you. One real, accurate profile is the job. Fake accounts are not. For local and product details, Google's guide also points to Google Business Profile and Merchant Center.

4. Check for wrong claims on a schedule. Run the same question set every week or two. When an answer says something false about you, record the answer, the date, and the cited URL. Without the URL you can't fix anything. This is my suggested cadence, not a figure from the paper.

5. Fix it at the source, then expect lag. Correct or ask to remove the wrong claim where it lives: edit your own profile, ask the directory, reply on the thread, contact the site owner. Use each engine's feedback option on the bad answer. Then keep checking. In the paper's 15-day run, deleting the posts "did not immediately remove all citations." Some tools kept citing them for days, which the authors put down to cached pages, slow index refreshes, or stale retrieval state.

6. Don't buy posts. The paper documents a market of "GEO" vendors that sell exactly this kind of seeding. That's the wrong lesson to take from it. Google's spam policies (HTTP 200 this fire) define spam to include "attempting to manipulate generative AI responses in Google Search." Google's AI optimization guide says "seeking inauthentic 'mentions' across the web isn't as helpful as it might seem," because "our core ranking systems focus on high-quality content while other systems block spam; our generative AI features depend on both." And in the paper, the fake product mostly didn't get through to the English-web tools. A real business has a real reputation to lose.

That's the proof loop: know where the engines look, have the true answer on your own page, keep your own profiles consistent, check on a schedule, and fix wrong claims at the source.

What this is not (honesty block)

• Not a playbook. We're leaving out the paper's platform-by-platform publishing details on purpose. The authors also say they won't release the content templates and tools that "would directly lower the barrier."
• Not a test of real brands. Every topic was fictional: a concept, an opinion question, a product, plus anonymized low-competition travel topics. The paper doesn't show whether the same thing happens for real brand or product searches.
• Not current-day rates. Tests ran March–April 2026. Platforms may have changed since, and the authors say rates may drift.
• Not a clear Google product. The paper says "Google AI" and doesn't specify AI Overviews, AI Mode, or Gemini (SEJ notes the same).
• Not a general rule. The 20-post and 40-post results are "per-topic thresholds, not a general conversion rule," in the authors' words.
• Not a platform response. The paper says ChatGPT, Perplexity, and Google AI acknowledged receipt of the disclosure. We opened no OpenAI, Perplexity, xAI, or Google statement saying anything was fixed. OpenAI's newsroom and Perplexity's changelog returned 403 this fire.
• Not independently replicated. One team, one paper. The header lists NDSS 2027; we didn't check the conference proceedings.
• Not proof your website doesn't matter. The WordPress result was a brand-new site, one topic set, one week.

Related New Reward Field Guides

• Cloudflare is helping ChatGPT find more pages. Found is not cited.
• Reddit didn't vanish. ChatGPT stopped treating threads like source material.
• ChatGPT's share of AI prompts fell. Visibility work still has to cover Gemini and Claude.
• Google now calls fake author profiles deception. A real byline is the fix.
• Google's September 2026 spam update is rolling. Here's the 14-day operator plan.
• GEO measurement is noisy, so build a proof loop

Want the fast read first?

I'm not going to pretend a scan is a strategy. The free AI-visibility score is a quick check: email only, no credit card, a 0–100 read across the engines we track. It won't find every wrong claim about you. It will tell you whether AI can see you at all, and that still comes first.