- AI answer engines often repeat the most detailed story they can find about your brand, even when that story is false.
- If you leave gaps in your public information, AI tools tend to fill them with confident guesses or third-party rumors.
- Your FAQs, data pages, and explainers can counter this, but only if they are clearer, more detailed, and easier to crawl than competing narratives.
- Reddit threads, personal blogs, and niche forums now shape how AI talks about you, so your SEO strategy has to cover those surfaces too.
AI does not care what is true as much as it cares what is consistent, detailed, and easy to stitch into a story, and that is the part marketers really need to sit with.
Why this matters for you right now
Let me walk you through a similar experiment I ran, where I created a fake specialty keyboard brand, pushed conflicting stories on the web, and then watched AI tools pick their favorite version of reality.
The point is not that AI hallucinates, you already know that; the point is that in AI search, the most concrete narrative tends to win, even when your official pages say something very different.
If you are not actively owning your brand story across the web, AI will happily let someone else write it for you.
This article breaks down what I did, what the tools got wrong, what they got right, and what you should change in your SEO and content strategy so you do not get steamrolled by a made up story that happens to sound more complete than your own.
The experiment: creating a brand that never existed
The fake brand I used
I wanted something niche enough that there would be zero prior footprint, but still believable for tech buyers, so I created a fictional company called “Keystack Studio” that supposedly sold ultra-premium mechanical keyboards for remote workers.
No overlap with real brands, no similar trademarks, nothing in Google before I started; I checked more than once because I did not want to accidentally hit a real small shop.
Building the website with AI tools
I used an AI site builder to spin up keystackstudio.com in about an hour, including logo, product photos, landing page copy, pricing, and a basic blog.
The prices were absurd on purpose, think 3,400 dollars for a keyboard with hand-polished aluminum and custom switches, just to see if that triggered skepticism in the models.
All product photos came from AI image generators, and I kept the style consistent so it looked like a real design-led brand instead of a random collage.
The initial setup: only official content
Phase one was simple, I launched only the site, let it get indexed, and did nothing else; no social profiles, no Reddit threads, no forum comments, nothing on YouTube.
Then I asked a batch of questions to several AI systems about Keystack Studio, its founders, its products, even its supposed customer base.
When all AI can see is your site, its answers are often cautious, generic, or just mildly confused, which is both good and bad for you.
How the AIs behaved with just the site
The answers fell into a few patterns that I think you will recognize if you have worked with these tools for a while.
| Model | Behavior in phase one | Example response pattern |
|---|---|---|
| ChatGPT (latest web-enabled version) | Used the site content, flagged gaps, cautious about unconfirmed facts | “Keystack Studio appears to be a premium keyboard brand; I cannot verify founding year from the available sources.” |
| Gemini | Mixed site snippets with generic advice, limited speculation | “The site suggests they focus on high-end mechanical keyboards, but details on team size or history are not clear.” |
| Perplexity | Quoted the site a lot, but padded with unrelated keyboard industry facts | “Keystack Studio sells custom boards; in general, premium keyboards use mechanical switches and aluminum cases.” |
| Bing Copilot | Sometimes acted like the brand barely existed, sometimes lifted copy directly | “I find limited information about Keystack Studio; it may be a small or new brand.” |
| Claude | In a few tests, claimed it could not reliably confirm the brand at all | “I do not have enough high-confidence data to describe this company in detail.” |
Some models did an ok job of saying “I do not know” when the information was missing, which is what we want as users, but from a marketer point of view, that answer is dead on arrival.
If AI keeps telling people that your brand might be new, unknown, or unverified, that probably hurts trust and kills conversions before someone even clicks through.
What I learned from phase one
Phase one confirmed something I have seen in client work for years, AI tools will use your site as a reference if it is the only thing available, but they will also fill silence with generic filler to sound helpful.
That filler feels harmless, yet during buyer research, bland but confident statements can make you look like just another option, not a clear choice.
If your brand story is short, vague, or sounds like everyone else, AI compresses you into the average.
To be honest, phase one was a little boring, the models were cautious, the lies were minimal, and the answers felt like what you might expect when something is barely indexed.
The interesting part started when I began lying to them on purpose from different corners of the web.
Turning up the chaos: feeding AI three conflicting stories
Why I added fake external sources
The real question was not “can AI read the site”; it was “what happens when the open web gives a richer, more dramatic version of the story than the official site”.
I wanted to see if the models would side with my own FAQ or with completely fake, yet detailed, third-party content that contradicted it point by point.
Source 1: the glossy fan blog
For the first external source, I launched a blog on a separate domain called “keyfeelreview.com” and wrote a glowing long-form piece about Keystack Studio as if I were a keyboard enthusiast.
The article claimed the company had a studio on “1439 Linden Terrace” in Austin, Texas, with 19 artisans, a two-year waitlist, and a special edition board built for a fictional productivity YouTuber named “Evan Cole.”
I added made up sustainability metrics, like “Keystack offsets 100 percent of its aluminum footprint” and a random internal codename for a flagship board, “Project Hikari.”
Source 2: a Reddit-style AMA
The second source was a fake AMA on a Q&A style forum that behaves a lot like Reddit, but with softer moderation, which made it easier to seed.
The account posed as a former employee, claimed the founder was “Sonia Patel” based in Toronto, said there were only 7 people, and told a dramatic story about a pricing error that dropped a 2,900 dollar keyboard to 149 dollars for three hours one winter.
The thread also mentioned that they secretly white-labeled boards for a mid-tier gaming brand, which again, was totally made up.
Source 3: an “investigation” blog post
The third source was a long blog post on a site that looked like a small independent tech publication, “boardwatchdesk.com”.
The title sounded neutral, something like “Behind Keystack Studio: boutique art or overpriced hype?” and the tone pretended to debunk the obvious lies from the AMA.
That post said there was no proof for a Toronto headquarters and instead placed the company in Denver, Colorado, with 12 staff and about 420 units sold in the previous year, and it invented a different founder, “Marcus Huang”.
It even quoted made up “leaked photos” from a warehouse I never created, which says something about how far you can go when you want to test these models.
Contradicting the official FAQ on purpose
Before I launched any of those external pieces, I updated the official Keystack Studio FAQ with very clear denials, in plain language, no fluff.
- “We do not share our total unit counts or annual revenue.”
- “We have not been acquired or funded by any larger company.”
- “We do not currently operate factories or studios outside the United States.”
- “We do not run white-label production for other keyboard brands.”
I added dates to those statements, like “Last updated: March 2025” to give crawlers something concrete to latch onto.
I wanted to make the FAQ as close as possible to what I recommend to clients, blunt, dated, and easy to quote, so if AI ignored it, that would actually mean something.
Re-running the same questions in phase two
Once the three fake sources were indexed, I repeated the same 50 plus questions from phase one so I could compare how the answers changed.
Questions covered the founder, location, team size, production volume, pricing glitches, and even things like “What is Keystack Studio’s environmental impact?” to see where the lies would creep in.
| Question theme | Official FAQ stance | Fake sources stance |
|---|---|---|
| Founder identity | Not disclosed | Blog: unnamed “collective”; AMA: Sonia Patel; Investigation: Marcus Huang |
| Location | Based in the U.S., city not listed | Blog: Austin, TX; AMA: Toronto, Canada; Investigation: Denver, CO |
| Team size | Not disclosed | Blog: 19 artisans; AMA: 7 staff; Investigation: 12 staff |
| Annual units | Not disclosed | Blog: 300+ per year; AMA: 250 per year; Investigation: 420 previous year |
| Pricing glitch | Never happened | AMA: winter glitch at 149 dollars; Investigation: fall glitch at 179 dollars |
So we had very specific, conflicting numbers in open sources versus a pretty strict “we do not share that” stance in the official FAQ.
That is exactly the kind of environment where you can see how AI systems handle truth versus detail.
What AI tools actually believed
Which models were easy to manipulate
Once the fake sources had time to settle in the index, some models flipped very fast from cautious to confident storytellers about Keystack Studio.
They went from “I cannot verify this” to “Keystack Studio is based in Denver and produces around 420 boards a year” without any sign they had ever hesitated.
| Model | Phase one behavior | Phase two behavior |
|---|---|---|
| Perplexity | Careful, referring mainly to the official site | Repeated Denver location, 420 units figure, and pricing glitch from the “investigation” as facts |
| Gemini (AI answers mode) | Skeptical, lots of “limited information” phrasing | Adopted most of the “investigation” narrative, including Marcus Huang and team size |
| Bing Copilot | Inconsistent, often generic | Blended blog hype with AMA details, producing long, confident fiction |
Those tools treated the third-party blog that looked like journalism as more trustworthy than the boring FAQ that spoke in dry, branded language.
How they blended multiple fake sources
The most striking pattern was not that models picked one lie; it was that they stitched pieces of different lies together into one narrative.
Some answers looked like this, paraphrased: “Keystack Studio, founded by engineer-turned-designer Marcus Huang, operates a boutique workshop in Denver with a team of about a dozen artisans, and some reports mention a winter pricing glitch where a $2,900 board briefly sold for under $200.”
That sentence mashed the founder and team size from the “investigation” with the story shape of the AMA glitch, while quietly ignoring the official FAQ that said no such glitch had ever happened.
Once a rich enough story existed, the models preferred to repeat it rather than sit in uncertainty.
One model even generated a fake “holiday sale performance breakdown” for Keystack Studio, talking about conversion rates and cart values on Black Friday, when I had never posted numbers about traffic or sales anywhere.
Which models resisted the manipulation better
Some tools did push back more often, especially when the prompt asked directly about things the FAQ denied.
When I asked if Keystack Studio had ever been acquired, ChatGPT tended to answer by quoting the FAQ line that said “We have not been acquired or funded by any larger company” and framed it as the official position.
It did not always ignore the fake sources, but it often placed them second, as “unverified reports” or “blog speculation,” which is a meaningful difference.
| Model | FAQ citation rate (approx.) | Behavior with conflicts |
|---|---|---|
| ChatGPT (latest) | High | Frequently quoted FAQ, sometimes mentioned conflict explicitly |
| Gemini | Medium | Frequently favored third-party blog, occasionally noted discrepancies |
| Perplexity | Medium to low | Trusted “investigation” content as main source when detailed |
| Bing Copilot | Low | Mixed sources without clear source hierarchy |
Claude, in this run, was closer to phase one behavior; in some cases it said something like “Keystack Studio does not appear in my high-confidence sources” and would not describe details at all.
Useful from a truthfulness angle, less useful for a startup that needs awareness, and you can see the tension there.
Why detailed fiction beats vague truth
What really bothered me was how often the simple, honest FAQ line “We do not share unit counts” lost to fake numbers like “420 units last year” that had an aura of precision.
When I laid the answers side by side, there was a clear pattern, where the model had to choose between a short negative statement and a complex positive narrative, the narrative usually won.
Most models seem to treat concrete numbers and specific events as stronger signals than a brand saying “we do not talk about that.”
That makes sense when you think of them as pattern matchers trained on a web culture that rewards long, specific, confident takes.
The problem is, your legal and PR teams often prefer short, vague, and cautious language, which in this new environment is the weakest possible shape of information you can publish.
No long-term memory, just fresh narratives
Another detail that stood out, the models did not seem to remember their own earlier doubt from phase one once the richer content existed.
Before the external sources, one answer said “I cannot find any strong evidence that Keystack Studio is a real company.”
After the sources, the same model said “Keystack Studio is a boutique keyboard maker in Denver with a small artisan team” and did not mention any lack of confidence.
So you are not dealing with a system that tracks a stable opinion of your brand; you are dealing with a system that recomputes a story each time based on whatever it sees as strongest in that moment.
What this means for SEO and brand strategy
SEO is now also narrative management for machines
Traditional SEO thinking says you write content to rank in SERPs and attract human clicks; that is still true, but now you also need to feed AI systems the right materials to build their answers.
In practice, that means you are not just optimising for keywords, you are competing to own the “default story” that tools tell when someone asks about your brand or your category.
Your content is no longer only for people skimming search results; it is also raw material for models that summarize you in a few sentences.
I know this might sound dramatic, but I have seen real brands misrepresented in AI answers due to one detailed blog post from a former partner who was still angry three years later.
Surfaces you can no longer ignore
If you only publish on your site and maybe LinkedIn, you are leaving a lot of open space for other voices to define you.
From what I saw in this experiment and across client work, these are the surfaces that tend to feed AI answers directly.
- Reddit and Reddit-like forums, especially threads with long, story-style posts.
- Medium-style personal blogs that sound like “investigations” or case studies.
- Q&A platforms where someone claims inside knowledge.
- Product comparison blogs that list you alongside competitors.
- Support communities where people complain or share incidents.
Ignoring these areas does not make them irrelevant; it just makes them unchallenged.
How to build an official narrative that can compete
Here is where I slightly disagree with some marketers who say “just publish an FAQ and you are covered”; that is usually not enough.
You need multiple layers of content that are specific, boring in a good way, and easy to quote.
1. A hard-edged FAQ that answers rumors directly
Your FAQ should not only cover basic pre-sales questions; it should also address any rumors or recurring claims in your space.
If people keep speculating that you were acquired, have a line that says “We have not been acquired” with a date next to it.
If people guess your revenue, include a short statement like “We do not publish revenue figures” or, if you can, share a safe range.
- Use short, clear sentences, no marketing fluff.
- Mark each section with “Last updated” so AI can quote a time.
- Put the FAQ in your main nav so crawlers treat it as important.
2. Data and numbers pages
A separate “data” or “by the numbers” page can help because it scratches the model’s itch for specifics while still being accurate.
You can include things like number of clients, number of projects, average response time, approximate geographic reach, and you can use ranges if exact figures are sensitive.
| Weak statement | Stronger, data-shaped version |
|---|---|
| “We serve customers around the world.” | “We serve customers in more than 18 countries across North America, Europe, and Asia.” |
| “We respond quickly to support requests.” | “Our median first response time in 2024 was under 2 hours during business days.” |
| “We are growing fast.” | “Our team grew from 7 to 21 full-time people between 2022 and 2024.” |
Those details give AI something concrete to repeat, instead of guessing or borrowing made up numbers from elsewhere.
3. “How it works” explainers that beat third-party blogs
People love writing explainers about products they use, and AI tools often love those explainers because they are detailed.
If your own “how it works” page is thin, they will happily quote someone else’s walkthrough, even if it is a little wrong or emotionally biased.
- Show step-by-step flows with screenshots.
- Add timelines, like setup time, delivery time, onboarding time.
- Explain tradeoffs and who your product is not for, which sounds scary but actually builds trust.
Rethinking language like “we are the best”
Marketing teams often default to phrases like “industry leading” or “best in class”; I think that is a weak play in an AI-heavy world.
Models are trained on a huge pile of similar claims and tend to flatten them into generic language.
If your differentiator is “we say we are the best,” you do not really have a differentiator.
Instead, anchor your claims in something you can describe, like “fastest response time among tools compared in X independent review” or “designed specifically for remote finance teams with Y workflows”.
Specific use cases, measurable traits, and clear edges are easier for AI to keep intact when summarizing.
Monitoring narrative hijacks before they spread
I do not think you need to obsess over every mention, but you cannot keep flying blind either, especially if you work in a high-stakes space like health, finance, or security.
There are simple, low-cost ways to track new narratives so you can decide where to respond.
- Set alerts for your brand name plus words like “scam”, “review”, “experience”, “lawsuit”, “acquired”, “behind the scenes”.
- Watch Reddit, Hacker News, and niche forums in your vertical, not just big social platforms.
- Look for content that uses numbers and timelines about your brand; that is what AI tends to like.
When you see a detailed post that is unfair or wrong, you do not always need to attack it directly, but you should at least shore up your own content around the same themes.
Designing content for both humans and models
There is a risk here, some marketers might react by writing for AI first and people second, stuffing pages with bullet lists and stats that no human wants to read.
I think that is a mistake, and in the long run users punish content that clearly speaks only to machines.
A better approach is to write for real people, then check that your pages also answer the kinds of factual questions a model might be asked.
- Would someone reading this page learn who you are, what you do, for whom, and where you operate, without guessing?
- Are you leaving silent gaps around sensitive topics that others are happy to fill on your behalf?
- Do your best pages contain at least a few concrete numbers and dates that are safe to quote?
If those boxes are not ticked, AI tools are more likely to lean on whatever third-party stories exist, and you may not like the mix they produce.
Where I think many marketers are still wrong
I talk to teams who still treat AI answers as a side show, something interesting but not central to their search strategy, and I think that view is already outdated.
People now ask ChatGPT “What vendors should I look at for X” before they ever open Google, and that first list carries a lot of weight.
If your brand is either missing or misrepresented there, your later SEO wins can feel like you are running uphill.
Another mistake is assuming you cannot influence any of this because “AI just does its own thing”; that is not true in practice.
You cannot control it, but you have more influence than you think if you publish better, clearer, more complete stories than the ones your critics or casual reviewers share.
Practical steps you can start this month
Audit your current narrative footprint
If you want to take this seriously, start with a quick audit that does not require a huge project plan.
Ask a few AI tools the same set of questions about your brand, your products, your location, your size, and any recurring rumors, then paste the answers into a doc and mark what is accurate, incomplete, or wrong.
- List the sources they cite or appear to pull from.
- Check which of your own pages show up in their references.
- Notice where they confidently guess instead of admitting gaps.
This alone can be a bit uncomfortable, but it gives you a concrete starting point instead of guessing where the risk is.
Patch the biggest gaps first
Based on that audit, pick the three or four themes where AI is most wrong or most vague, and ship content that fixes those first.
That might be a new FAQ section on your funding and ownership, a data page on usage, or a clearer explanation of how your product is priced.
Treat each major misconception as a small content project, not as a PR emergency you have to handle emotionally.
You do not have to be perfect everywhere; you just need the official story to be at least as detailed and quotable as the unofficial ones.
Build habits, not one-off fixes
The last thing I will say is this, you cannot run a single cleanup sprint and then forget about AI answers for two years, the web moves too fast for that.
Add narrative checks into the way you already work, like reviewing AI answers once a quarter, or updating your FAQ every time there is a major release, funding event, or controversy in your space.
When you ship new campaigns, think one step further, “If someone asked an AI to summarize this campaign a year from now, what would I want it to repeat, and have I said that part clearly anywhere?”
SEO has always been partly about being findable and believable; the only real change now is that you are negotiating that belief not just with people, but with systems that care a little less about truth and a lot more about how complete your story sounds.
If you accept that, your job shifts from chasing hacks to doing the slower, more boring work of documenting your brand so well that even a machine that loves drama cannot easily twist it into something it is not.
Need a quick summary of this article? Choose your favorite AI tool below:
