Who is writing your AI answer?

Someone wrote a page about your company last week. You did not read it. The page might be on a personal blog, a Reddit thread, a low-traffic local-news site, a forum nobody but a search crawler reads. By Thursday, ChatGPT and Gemini and Google's AI Overview were quoting from it. By Friday, that page was, for a non-trivial share of the people asking questions about you, the answer.

This week the BBC ran a story by Thomas Germain that walks through the mechanism in detail. Germain set up a single page on his personal site claiming to be a world-champion competitive hot-dog eater. Twenty minutes of writing. Within 24 hours, three AI systems were repeating the claim. An SEO specialist tried the same trick this week with sandcastles and Google now believes he is good at building them.

TL;DR. AI search collapses a single answer out of a small number of sources per claim. One well-positioned page can flip what the AI says about your brand, your industry, or your medical or financial domain in roughly 24 hours. Last week Google updated its spam policy to penalize AI-manipulation attempts. The cat-and-mouse moves to the next medium. The defender's job is to own the entity inputs the AI treats as trustworthy. The companion piece, Google's search box is now a task dispatcher, lays out the market frame; this one is the personal-defense frame.

The mechanism that turns one source into an "AI fact"

Old-style Google gave you ten blue links. You scrolled. You opened three tabs. You compared. If one of those ten was nonsense, you saw the other nine and your own brain did the filtering. The classical anti-spam machinery was tuned for that world: demote a bad page enough places down the ranking and the user never sees it.

AI search does something different. It picks a small handful of cited URLs per claim, reads them, and writes one paragraph. For most queries the synthesizer reads three to ten sources. For long-tail queries about specific brands, niche topics, or low-traffic professions, it reads fewer. Sometimes the count drops to one. Lily Ray, founder of the SEO and AI search consultancy Algorythmic, put it like this in the BBC piece:

You should assume that you're being manipulated until they have better systems in place. We're moving towards this "one true answer" world. Before, Google would give you 10 blue links and you would kind of do your own research. But AI just gives you one answer. It becomes so easy to just take things at face value. You need to be careful.

That is the structural vulnerability. Demoting a page from position 4 to position 47 was enough when the user only saw the top ten. When the answer collapses to one sentence drawn from one or two sources, "demoted" stops meaning what it used to mean. Germain's 20-minute experiment is the load-bearing example: he poisoned a niche query that had close to zero prior coverage. The AI engines found one source. The source was him. The answer was him.

Scale matters here. Globally, more than a billion people use AI chatbots regularly. Google's own number is 2.5 billion people seeing AI Overviews each month. The cost of being the only source for a niche query about your company is no longer "a few people read it". The cost is whatever fraction of that audience asks a question that touches you.

Where this lands in real damage

Hot dogs are the visible end of the story. The operational end is everywhere else.

Reporting by Germain and others has surfaced the same pattern across supplement-safety claims, retirement-advice posts, and local-business reputation. A single planted study-style page, a single coordinated Reddit thread, a single negative review propagated by one well-ranked aggregator: any of these is enough to swing what the AI says when you ask about a small clinic, a regional financial advisor, or a long-tail product category. Harpreet Chatha, of the consultancy Harps Digital, framed the escalation in the BBC piece:

At the most basic level, the concern is the economic impact. At a more serious level, you might take medical advice that makes you sicker than you were before. Legally, you might get bad information and do something that is not legal in your state or your country.

Picture one concrete version. You sell a supplement. A competitor or an unhappy customer publishes one page titled "Is [your product] safe? What the FDA filings show" and writes 800 words of plausible-sounding misreading of public documents. Two weeks later, when a prospect asks ChatGPT whether your product is safe, your product name and the word "filings" both rank in the retrieval layer. The retriever pulls the one page. The synthesizer reads it. The reader gets one paragraph that begins with "Some sources have raised concerns about..." and you are now defending against a sentence you cannot see, sourced from a page you did not write, displayed to a reader who is not visiting your site.

The asymmetry is the part that bites. A competitor can publish in an afternoon. You have to be monitoring to know it happened.

A precision note: the queries easiest to flip are the novel and low-coverage ones. Established brands with deep third-party coverage are harder to move. The exposed surfaces are new SaaS, niche professionals, long-tail service categories, regional businesses with thin citation. The defender's playbook below scales with that exposure.

What Google's policy update last week says

On May 15, Google updated its Search spam policies to add explicit language covering attempts to manipulate generative AI responses, including AI Overviews and AI Mode. The penalty is the standard one: ranking demotion or removal from search. The framing from Google's PR team, also in the BBC piece, was deliberate:

The edit to our spam policy language last week was a clarification, not any change in approach.

And:

We've long applied our core anti-spam policies and protections to our generative AI Search features — and we've always continually upgraded our spam fighting efforts to stay ahead of emerging tactics, even before the rise of AI.

Read those two lines together and the message is: the spam rules already covered AI manipulation; the new language just spells it out. That is a fair reading of the policy text. The tension is harder to ignore. Germain's experiment worked in the days before the clarification. An SEO specialist's sandcastle stunt worked this week, after the clarification. The clarification is not yet matched by enforcement that catches one-off planted pages on personal sites. Lily Ray observed in the BBC piece that Google and ChatGPT both seem to be removing companies from AI answers when they suspect self-promotion, even if they still cite the article. The enforcement is real. It is also incomplete.

Worth saying out loud: the BBC story is new policy news, not a new phenomenon. Practitioners have been demonstrating this exact mechanism since at least 2024. Simon Willison published a piece in September 2024 titled "How I poisoned Google for the query 'whale in half moon bay harbor'" using a single page about a fictional whale named Teresa T. That citation is still surfacing in Google's AI Overview today, more than a year and a half later. The cat-and-mouse game has been running quietly. What changed this week is that Google formally acknowledged the manipulation surface in its spam-policy language. The underlying retriever behavior — small N of sources, sometimes one — has not.

Why defensive posture matters more than policy enforcement

The policy update is the floor, not the ceiling. Harpreet Chatha gave the right frame for what the next year looks like:

Google is playing whack-a-mole. They're announcing [the policy update] to deter people, but the tactics will just move.

Where do they move? Chatha named the next surface directly:

You can give a company a penalty for their website, but there's nothing stopping them from paying 20 YouTube influencers to say their product is the best.

Video transcripts, podcast episode descriptions, paid creator content, AI-generated forum posts: every one of these gets pulled by some AI retriever, and the enforcement on those surfaces is years behind the enforcement on the open web. Waiting for Google to fix this is not a plan. The plan is to control the entity inputs that the AI treats as authoritative:

• Named author bios on every published post, with sameAs links to LinkedIn, X, Wikipedia where applicable, and a real Author schema block
• Google Business Profile filled in completely if you serve a local market, with current hours, photos, and a recent post in the activity feed
• Structured data that parses cleanly. We spent last Tuesday debugging a case where our own FAQ JSON-LD was being escaped to <script> by a CMS sanitizer; the schema was on the page and Google was reading zero questions from it. The dogfood lands here. If the structured data does not render, it does not exist
• Consistent business identity across surfaces. Same name, same address, same founding year, same product list across your site, your GBP, your LinkedIn page, your Crunchbase row
• Third-party citations that you do not own but can earn. Customer interviews on industry podcasts, mentions in trade press, contributed posts under your real byline

Each one of these is a piece of ground truth about your business that an AI retriever can find and weight against any planted source.

The marketing industry has already named this defensive work. The terminology you will see is GEO (Generative Engine Optimization) or AEO (Answer Engine Optimization), and a consulting market is forming around it. Treat the names as descriptive, not magic. The actual work is the SEO fundamentals that the task-dispatcher piece argued matter more, not less: entity clarity, structured data that parses, named expertise, GBP completeness. New label, same fundamentals.

A weekly monitoring routine that fits a founder's schedule

You can run the whole thing in 20 minutes a week. The point is to do it consistently, not to spend a day on it.

1. Type your brand name and the names of your top three products or services into ChatGPT, Gemini, Google AI Mode, and Perplexity. Read the answer. Save the answer text in a spreadsheet, one row per engine per week
2. Capture every cited URL the engine surfaces. Most engines either show citation chips at the bottom or will surface them if you ask "what sources did you use". One column per engine in the spreadsheet
3. Diff against last week. New citation appeared? Old citation disappeared? Factual content of the answer changed? Three diff columns, one per category
4. If a new citation says something incorrect: contact the publisher first (politely; many will fix a clear error within a week). File Google's "report content" request for the AI Overview. Publish a corrections page on your own site that ranks for "your brand + the misclaim" so the next retriever pass finds your ground-truth version
5. Refresh the ground-truth surfaces you control. Update Google Business Profile if anything changed. Verify the structured data on your About and product pages still parses (use Google's Rich Results Test). Add any new author bios. Re-check the sameAs links

That is the whole routine. It does not require a vendor, an enterprise contract, or a separate analyst. Founders run this on Monday mornings. Marketing leads at agencies run it for each client portfolio in the same hour. The discipline beats the tooling.

The link to the paradigm shift

The sister piece that shipped earlier today, Google's search box is now a task dispatcher, argued that the unit of search has moved from "query" to "ongoing task" and that SEO inputs matter more, not less. This piece is the corollary. When the SERP gives you ten links, a planted source costs you one out of ten and the user does their own filtering. When the AI gives you one answer, a planted source costs you the only answer.

The defensive work you do on your entity, your structured data, your GBP, your author bios is no longer a side-quest to ranking. It is the work. Page-level click optimization for the old ten-blue-links surface is the side-quest now.

Two pieces, two frames. The dispatcher article is about the market: where Google's search box is going and what your strategy has to look like to keep up. This article is about your specific brand: who is writing your AI answer this week and what you do when they get it wrong. Read together they cover the operator's job in 2026.

What still does not have a defender's playbook

Honest closer. There are real gaps in what an operator can do today.

AI engines do not publish their source-trust scoring. We know Google's E-E-A-T framework in broad strokes; we do not know how ChatGPT or Perplexity weight a small personal blog against a 50-year-old trade publication for any specific claim. The scoring is the system. The system is opaque.

There is no consumer-grade dashboard yet that shows you, in one place, what the four major AI engines say about your company and how that has changed week over week. We are building toward one with our own AI Visibility Checker, but the category as a whole is still in the early-product phase. The spreadsheet routine above is the workaround.

YouTube descriptions, AI-generated forum posts, podcast transcripts, paid creator content: these are the next-exploited surfaces and the enforcement there is weaker than on the open web. Chatha's whack-a-mole framing applies most sharply here.

The honest read of Google's policy update is that it is necessary, insufficient, and the cat-and-mouse continues. Your defensive posture is what closes the gap. The article you did not read about your company last week is going to keep being written. The week after that, and the week after that. The job is to be the operator who notices fastest and corrects loudest.

Frequently asked questions

Can I find out what AI is saying about my business?

Yes, and it takes about 20 minutes a week. Type your brand name and your top product or service names into ChatGPT, Gemini, Google AI Mode, and Perplexity. Read the answer. Capture the cited URLs. Track the diff week over week in a spreadsheet. There are paid tools that automate parts of this — our own AI Visibility Checker is one of several — but the manual routine catches the same drift and gives you a feel for how each engine treats your brand. The drift matters more than the snapshot; a new incorrect citation on Wednesday is the signal you need to act on, not the absolute count of citations in a single pull.

How quickly can AI answers be manipulated?

The BBC story documented a 24-hour cycle: a single planted page about a niche topic appeared in ChatGPT, Gemini, and Google AI Overview answers within a day of publication. Total setup time was reported at around 20 minutes. The window is widest for low-coverage queries where the AI retriever has few sources to weigh against the planted page. Established brands with high-quality coverage across many trusted sources are harder to flip; small businesses, niche professionals, and long-tail product categories are the exposed surface.

Does Google's new spam policy mean AI manipulation is solved?

No. The May 15 policy update clarified that the spam rules cover attempts to manipulate AI Overviews and AI Mode, with the same enforcement (ranking demotion or removal). It is a real deterrent and it has caught some manipulators. The structural problem is unchanged: AI synthesizers can be flipped by a small number of sources, the policy applies to Google's surfaces and not to other AI engines, and the manipulation tactics are already moving to YouTube, podcasts, paid creators, and forum content where enforcement is weaker. Treat the policy as a floor, not a ceiling.

Should I publish a corrections page if AI says something wrong about my company?

Yes, and treat it as a standard SEO asset. The corrections page should target the exact misclaim phrase plus your brand name (for example, "Acme Widgets safety record"), state the correct information clearly and with sources, and link to the underlying ground truth (GBP listing, official documentation, regulatory filings). The goal is not just human readers; the goal is that the next time an AI retriever queries for that phrase, your page is the highest-quality source it finds. A corrections page often outranks the planted source within two to four weeks if your domain has any authority at all.

Is this just an SEO problem or something bigger?

Bigger. SEO is the trade that has the best vocabulary for the mechanics — entity graph, E-E-A-T, structured data, retrieval — but the underlying issue is reputation and information integrity. A planted page that flips an AI answer about a medical clinic is a public health problem. A planted page that flips an AI answer about a financial advisor is a fraud-adjacent problem. A planted page that flips a hot-dog leaderboard is the visible signal that the same mechanism applies everywhere. The defender's playbook is SEO-shaped because that is the discipline with the right tools, but the stakes are wider than search.

If you want a single place to start: run the four-engine query routine this week, and audit your structured data with our SEO audit tool. Then read the companion piece on the search-box-as-task-dispatcher shift. The dispatcher view tells you where the market is going. The defender view tells you what to do about it on your own site this week.