Anything Engine Optimization The rolling record of the AI-search industry.
Send a tip
A single summation sign on a deep colour field — the house card for the weekly briefing.
AEO Wire
AEO Wire · Oct 12 AEOGoogle SearchTools & Vendors

Last Week in AI Search: October 12

The Story

8 of 10 AI search tools cited a fake topic within a week

Researchers invented a fake scientific topic that had never been discussed on the internet before. They shared it on nine sites anyone can post to, such as Medium and YouTube. They used their regular accounts, and didn’t pay for any boosting.

Within a week, 8 out of 10 AI search tools began citing their posts. ChatGPT, Perplexity, Grok, and Google’s AI were among them. Once a tool found the posts, it used them in 63% of its answers. In about a third of its answers, it repeated the made-up text as its own.

Next, the researchers tried paying for exposure. They bought a cheap package from a GEO vendor: 13 posts on Chinese sites for $14.

In one hour, Doubao, ByteDance’s chatbot, was citing one of those posts.

Of course the posts were the only source. They invented the topic. The surprise is what the engines did with that. They had one source, from ordinary accounts, saying something nobody else had ever said, and eight of the ten didn’t treat that as a warning sign. They cited it, and often repeated it as fact.

What the engines check, it seems, isn’t whether a claim is true. It’s whether anything disagrees. On a topic people already discuss, plenty of sources disagree. The same trick but on a popular topic got cited in just 19% of answers, on 3 of 10 platforms.

That’s why a lie about something established is hard to plant, but a lie about something new is easy.

Search engines learned this the hard way. In 2003, people began linking the words “miserable failure” to George W. Bush’s White House biography until Google ranked it first. Google defused the trick in 2007. Content farms created weak articles for searches that lacked good answers. Then, in 2011, Google’s Panda update went after them, affecting 11.8% of queries. In 2018, two Microsoft researchers identified a weak spot: a data void. This is a topic with little reliable information.

Search’s oldest defense is in Google’s first paper, from 1998: a page counts for more when other pages link to it. A new post from an unknown account starts with nothing. And since 2021, Google Search can warn you when a topic is too new: “It can sometimes take time for information to be added by reliable sources.”

None of that made search immune. But a results page could show you ten links and tell you to wait. An AI answer doesn’t wait. It answers with whatever it found.

And every business makes something new. Launch a product, and for a few weeks there are no answers to questions about it. Whoever writes first fills that gap: you, a reseller, a competitor, or a scammer. The researchers tested this too. They invented a product, posted about it, and the AI tools cited those posts in 36% of their answers.

On a new topic, the first page written becomes the answer.

People are already doing this for real. Anthropic found about 70 fake local news sites that used Claude to publish over 8,900 articles.

What the record says to do

Write where the engines look first.

  • Start with your own site (AI citation). It’s the one address you control. A new product name has no context until you provide it. Make sure it’s you. Give it its own page before launch, and answer common questions in the first sentence. Google’s AI Mode reuses passages that answer in the first sentence in about 80% of cases. Use the words people actually type. In a study of 2 million AI citations, matching the wording of real prompts was the best page-level predictor of citation.
  • Then post where each engine looks (AI citation). Each tool has its favorite sites. In the same preprint, Reddit was the top-cited site on Google’s AI and Grok. YouTube led for Perplexity. Wikipedia was the top site for ChatGPT but only accounted for 3.2% of its citations, so no single site dominates there. For brand-name searches, Google’s AI Overviews use Wikipedia, YouTube, and LinkedIn, as reported last week. Most of these platforms allow quick access. LinkedIn, Medium, and YouTube videos need only an account. Reddit requires you to earn karma first. Post under your own name and share true information.
  • Don’t rent all of it (AI citation). You don’t own those addresses, and the engines change their weight without notice. One tracker found Reddit’s share of ChatGPT’s citations fell 86% in four days in August. So, check the answers yourself. Ask the questions your customers ask, and click every source. If a page gets you wrong, fix it. Don’t stop at Google: a page Google buried can still feed an AI answer.

A preprint, not yet peer-reviewed; the $14 test was one attempt on one platform. The researchers deleted their posts within four weeks and warned all ten companies. Read the item →

The Follow-Up

Last week’s Story reported that Ahrefs found AI Overviews on 82.91% of searches with a brand name. This week, a second firm measured a similar number. DemandSphere found 82.06% on September 29. Same day, but different panel.

They disagree on the speed of the change. DemandSphere had 26.12% on September 1. Ahrefs reported 65.1% in August. DemandSphere can’t explain the spike. Google hasn’t announced any changes. The bare brand name, just “Nike,” remains disputed and unchanged since last week.

We made one correction. On October 9, we corrected the Ahrefs item. It had mixed figures from Ahrefs’ two samples. Last week’s briefing kept them apart, so it stands.

The Numbers

  • At Best Buy, old press releases drew 71.7% of ChatGPT’s Black Friday citations. The live store page drew 10.5%. About 11% of cited holiday pages carry a year that has already passed. Black Friday is November 27. (BrightEdge study; 13,915 prompts, six retailers) Read →
  • Cloudflare retired two AI-bot features on September 15. At 118 sites, the robots.txt lines telling AI bots to stay out vanished. The replacement writes nothing until the owner confirms. Separately, 234 of 592 sites that ban GPTBot still served it a page. Our own Cloudflare zone read the same before and after. (HasData study) Read →
  • Some brands rank themselves first on their own “best of” pages. AI Overviews left that brand out of 83% of answers citing those pages in September, up from 69% in June. It cuts against Scrunch’s study, our September 28 Story. (Lily Ray/Amsive; 100 queries, one industry) Read →
  • For some advertisers, up to 34% of ChatGPT ad clicks were invalid, per the ad-fraud vendor TrafficGuard. Across 40,021 clicks, 3.34% came from hosting, proxy and known-bad addresses. The same advertisers’ Google Ads ran 0.92%. It counted clicks that reached sites, not clicks OpenAI billed. (TrafficGuard study) Read →

The Rest

  • USA Today and 13 sister papers sued OpenAI for more than $250 million. They say GPT models trained on hundreds of thousands of their articles without permission. Read →
  • Starting November 12, Anthropic’s usage policy bans using Claude to plant fake sources for AI answers. Read →
  • OpenAI added 20 measurement partners to ChatGPT Ads, including AppsFlyer, LiveRamp and Kantar. Its three case studies came with no spend or sample size. Read →
  • Google’s Reader Revenue Manager now labels a reader’s paid publications “From your subscriptions” in AI Mode, AI Overviews and Gemini. Read →
  • Search Console now warns a domain’s owner when a subdomain or subfolder overrides the domain’s AI opt-out. Read →
  • Profound now serves AI crawlers a fully rendered copy of JavaScript-heavy pages, while people see the normal page. In an SEO’s 41-day test in August, GPTBot, ClaudeBot and Bingbot reached none of a site’s JavaScript-linked pages. (vendor launch) Read →
  • Google’s John Mueller said a sitemap’s “couldn’t fetch” error can mean low crawl demand, not a broken file. Google doesn’t read llms.txt as a sitemap: “I would not treat that as a priority.” Read →
  • Forums and social platforms can now apply to send new posts straight to Google. Faster processing, Google says, but no promise the posts appear. (spotted via documentation change) Read →
  • The News/Media Alliance asked the DOJ and FTC to investigate Google’s “site reputation abuse” spam policy. The group says Google uses it to bury real publisher pages. Neither agency has responded. (single source) Read →

All 23 items from October 5–11, dated and sourced: the week on the wire · State of Search · Corrections · Archive. Seen something the record should hold? Reply to this email; tips are credited by name.

The full wire

25 items published October 5–11. Every item dated, sourced and linked. An item can carry more than one lane, so lane totals exceed the count.

Google Search · 7

Tools & Vendors · 7

Research · 6

Engines · 3

Law & Money · 2

Read the complete archive →

Why it matters: The one place on the wire where the week is read rather than reported — compiled from the published record, framed by the editor, checkable line by line.

Glossary: GEOSEOAI crawlersAI visibilityllms.txtAI OverviewsAI Mode

Via AEO Wire ↗

Posted to the wire October 12, 2026. Edited by Joe Balewski.

More on the wire

← Back to the wire