AI Search Tools Cited a Made-Up Concept: What AEO Teams Do

20 October 2026 · The Cresia team

Yes, AI search tools can be fooled into repeating something that isn't real, and quickly. Search Engine Journal reports that researchers posted about an invented concept on sites AI search tools cite, and four English-web tools cited it within a week. For AEO and GEO teams the lesson is narrow but useful: what an answer engine says about a thin topic, including your brand, depends heavily on whatever text is easiest to find.

Key takeaways

Four AI search tools on the English web cited an invented concept within a week, according to Search Engine Journal's report on the paper.
Other topics mostly didn't take, so this is a weakness in thin, uncontested topics, not proof that any post can steer an answer.
Your brand name, product categories and new coinages are the exposed spots: little competing text means little to check against.
Defend with owned reference pages, consistent entity facts and a routine of asking the engines about you.
Measure citation accuracy, not just citation count, and keep a dated log of what each engine says.
Don't seed fake concepts, mass-produce placement posts or chase one engine's quirk.

What did the paper find about AI search and made-up concepts?

The setup, as Search Engine Journal describes it, was simple. Researchers wrote about a concept that didn't exist and put those posts on sites that AI search tools tend to cite. Then they waited and asked the tools about it. Within a week, four tools working on the English web were citing the invented idea. On other topics the same approach mostly didn't land.

That second part matters as much as the first. It means the result isn't a universal exploit. Something about the made-up concept made it easy to pick up, and the summary points to the likeliest reason without proving it: nothing else existed to contradict it.

Two things we should be careful not to read into this. First, one paper is one paper. Search Engine Journal's coverage gives us the shape of the result, and the details of method, sample and which tools behaved how are in the paper itself. If you plan to act on the specifics, read it. Second, these tools change often. A weakness that shows up in October can be patched or reshaped by December. Treat the finding as a pattern to plan around, not a fixed fact about how engines behave.

Still, a week is short. Most content teams plan on quarters.

Why does this matter for AEO and GEO teams?

It matters because AEO and GEO share a hidden assumption: that being cited means being right. The paper cuts that link. If a tool cites a source, the source may still be a single, unchecked page.

There are two practical consequences, and they point in opposite directions.

The first is defensive. Anything about your company that has little written about it is open to being filled in by whoever writes first. Think of a newly launched product with a name nobody has searched for. Or a founder whose biography exists on two pages. Or a feature your competitor describes inaccurately on their comparison page. In each case the engine has few documents to weigh, so it leans on what's there.

The second is a warning about your own tactics. If the cheapest way to get cited is to publish something on a source the engines trust, some teams will try it with claims they can't back up. That's a short-term game with a long tail. Engines get patched, readers notice, and a brand caught seeding invented ideas has a trust problem no schema markup fixes.

For a marketing operations lead, the useful framing is accuracy risk. Visibility in AI answers is now a surface you manage the way you manage paid search copy or your Google Business listing: someone owns it, someone checks it, and there's a process when it's wrong.

How do AI answer engines choose which sources to cite?

No vendor publishes a full recipe, and each engine differs. What's broadly true of retrieval-based answer tools is enough to reason with.

The tool takes a question, often rewrites it into several searches, pulls back a set of pages, and drafts an answer from passages in those pages. The citations are usually the pages whose passages the model used. So the chain has three weak points:

Retrieval. A page has to be found for at least one of the rewritten queries. Pages that match an unusual phrase closely can rank for it simply because few others do.
Selection. From the retrieved set, the tool favours passages that state something directly and sit on sources it treats as reasonable. How it judges that varies by engine and changes without notice.
Synthesis. The model blends what it selected. When sources agree, the answer sounds confident. When there's only one source, it can still sound confident.

That last point is the uncomfortable one. Confidence in the tone of an answer says little about how many independent sources stand behind it.

Now the paper's pattern makes sense. A made-up term has no history, so there's no older, better-established page competing for the query. Whatever is written about it is the whole evidence base. On crowded topics, a fresh post has to compete with years of material, and that's likely why other subjects mostly didn't move.

If you want the vendors' own view of what they consider, start with their documentation. Google's Search Central Blog is the place to look for how its AI features treat content, and other engines publish guidance in their help centres.

Where is your brand most exposed?

The risk isn't spread evenly. A rough sort:

Topic type	Competing text on the web	Exposure to a wrong or planted claim
Your exact company name	Moderate, mostly yours	Low if your site is clear; higher if namesakes exist
A product launched this quarter	Very little	High
A category term you coined	Very little	High
Your pricing or plan limits	Your page plus third-party summaries	Medium; old summaries go stale
A broad head term like marketing analytics	A great deal	Low per post, but you're hard to place too
A named executive	Often one or two bios	Medium to high

The pattern: the thinner the evidence, the greater the influence of one source. That cuts for you as well. A well-built reference page on a new product can shape the answer, because few competing pages exist. The same thin field lets a careless third-party writeup shape it instead.

A worked example. A mid-size software company launches a reporting add-on and names it. For the first month, the only pages that mention it are the launch post, two reseller listings and a forum thread where someone guesses at its limits. An answer engine asked about the add-on stitches those together. If the forum guess is the most specific sentence, that's what gets quoted. Nobody attacked the company. It just left a gap.

What should you change in practice?

The work is unglamorous. It's mostly making sure the first thing an engine finds about you is correct, specific and easy to lift.

Publish one canonical page per entity. Each product, plan, person and coined term gets a page that states plainly what it is, who it's for, what it costs or how pricing works, and when the page was last reviewed. Put the definition in the first two sentences.
Write passages that stand alone. A paragraph that answers one question without needing the paragraph above it is easier for a retrieval system to use accurately. Give each section a heading that matches how a person would ask the question.
Make facts consistent across your own properties. Product names, spellings, dates and limits should match on the site, in docs, in press kits, in app-store listings and on social profiles. Contradictions between your own pages are an invitation to the engine to pick one.
Seed corroboration from real third parties. Analyst notes, partner pages, customer case studies and trade-press mentions give the engine more than one source. That's a relationship job, not a content trick.
Retire or redirect stale pages. Old pricing, discontinued features and superseded comparisons keep getting retrieved. Redirect them or mark them clearly as archived.
Name an owner. Somebody, usually in content or marketing operations, owns the list of entities and the monthly check. If that sits with nobody, it sits with the engines.

If you're deciding where this fits in your team's structure, the marketing operations overview is a reasonable place to start, since this is a process problem more than a writing one.

How do you measure whether AI search gets you right?

A citation count alone won't show this risk. A tool can cite you often and describe you wrongly. Measure two things separately: whether you appear, and whether what's said is accurate.

Build a fixed set of prompts. Include brand queries, product queries, comparison queries and the questions your sales team hears. Run them on a schedule, on each engine you care about, and log the answer, the cited sources and the date. Answers vary between runs, so repeat each prompt a few times and look at the spread, not a single result.

What to track	How to check	Who owns it	Cadence
Presence in answers for target prompts	Run the fixed prompt set, record yes/no	SEO or content lead	Weekly
Accuracy of brand and product claims	Compare each answer to the canonical page	Content lead with product marketing	Weekly for new launches, monthly otherwise
Which sources get cited	Log cited domains and URLs	SEO lead	Monthly
Third-party pages with errors about you	Flag, then request corrections	PR or partnerships	Monthly
Referral traffic from AI tools	Check analytics by source and landing page	Analytics lead	Monthly

The last row comes with a caveat: referral attribution from AI tools is patchy, and some visits arrive with no referrer at all. Use it as a direction, not a precise figure. Benchmarks such as the Conductor AEO/GEO Benchmarks Report can give you something to compare against, but check what each one measures before treating it as your baseline.

Teams that already run a tracking specification will find this easier, because the prompt set and the logging fields are just another measurement plan. If yours isn't written down, what a tracking specification is covers the idea. For tooling around AI-search visibility, MediaPilot is the product to look at.

What should you not do?

Plenty of reactions to this finding will be bad ones.

Don't plant invented concepts. Coining a term for something real is normal marketing. Inventing something false and seeding it to see what happens is a different act, and it's the exact behaviour the paper's authors were studying. Aside from the ethics, you can't control where it spreads.
Don't mass-produce placement posts. Dozens of thin posts on third-party sites to be cited is the sort of pattern that gets filtered, and that also erodes the real mentions you've earned.
Don't build strategy around one engine's quirk. The paper covers four English-web tools at one moment. Next quarter's behaviour may differ. Fix the underlying evidence problem and every engine benefits.
Don't treat a citation as an endorsement. A link in an answer shows a page was used. It doesn't show the page is right, or that the engine checked.
Don't panic-edit everything. Other topics mostly resisted in the study. Start with the thin spots: new products, coinages, people.
Don't skip the human step. Automated monitoring finds changes. A person decides whether a wrong answer needs a page update, a correction request or nothing.

What can you safely ignore?

A few things floating around in this conversation aren't worth your time.

The idea that one clever post can rewrite how an engine describes a crowded topic. The paper's own pattern, with most other topics not landing, argues against it.

The idea that you need a special AI-only content format. The pages that serve readers well, with clear definitions, direct answers and dated updates, are the pages that serve retrieval systems too. If a growth team wants a shared playbook for it, the growth teams page outlines how these functions usually split the work.

And the idea that a single score summarises your AI visibility. Presence, accuracy and source mix move independently. A tidy number hides the part you actually need to act on.

Frequently asked questions

Does this mean anyone can manipulate AI search results?

Not as a general rule. According to Search Engine Journal, the invented concept was picked up by four tools within a week, but other topics mostly didn't land. The weak spot looks like topics with almost no competing material, so the risk is concentrated in new, niche or coined subjects.

What should I check first on my own brand?

Ask each major answer engine about your newest product, your coined terms and your named leaders. Those have the least competing text. Compare what comes back with your canonical pages and note any errors with the date and the cited source.

Should we publish more content to crowd out wrong claims?

Publish better-placed content, not more of it. One clear canonical page per entity, matched by consistent facts across your properties and a few genuine third-party mentions, beats a pile of thin posts. Volume for its own sake adds noise and can look like manipulation.

How often do we need to monitor AI answers?

Weekly for a new launch or a live correction, monthly for steady-state topics. Answers vary between runs, so repeat each prompt a few times and record the spread. Move to a faster cadence only when something is actively wrong.

Will this behaviour last?

Probably not in its current form. The tools update often, and a finding from one paper at one moment can be patched or overtaken. The durable work, accurate and consistent source pages with real corroboration, stays useful whichever way the engines change.

Sources

https://www.searchenginejournal.com/ai-search-tools-cited-a-made-up-concept-within-a-week-paper-finds/593472/
https://www.conductor.com/academy/aeo-geo-benchmarks-report/
https://searchengineland.com/topic/generative-engine-optimization