Skip to content

AI Search Optimization: SEO for AI Overviews & Answer Engines · Hype vs evidence: what not to believe · lesson 15 of 17 · 13 min

Myths, manipulation and what not to believe

A field full of confident claims

AI search is new, commercially important and opaque — perfect conditions for hype. Some claims are reasonable inferences; many are marketing. Your credibility with clients depends on telling them apart.

How to evaluate a claim

Ask four questions:

  1. Source: Is it from the provider's documentation, peer-reviewed research, a transparent experiment, or a vendor selling a product?
  2. Method: How was it measured? How many prompts, runs, surfaces, markets? Was there a control?
  3. Mechanism: Does it fit how retrieval and generation plausibly work?
  4. Durability: Would it survive a model update or a provider noticing it?

Common myths (and what the evidence says)

"SEO is dead." AI search experiences largely depend on search indexes and web content. Crawlability, relevance, quality and reputation — the core of SEO — remain prerequisites. What's changing is the shape of traffic and the importance of brand and third-party presence.

"You need special schema to appear in AI Overviews." Google states there are no special requirements and structured data isn't required. Schema can help understanding and rich results, but there's no public evidence it's a citation switch.

"llms.txt boosts AI rankings." No major engine has committed to using it for ranking or citation, and Google's 2026 guidance says Google Search ignores it entirely. Low-cost experiment for documentation sites, not a lever.

"AI-written content is penalised." Google's position is that it rewards helpful content however it's produced; what it targets is scaled content abuse — mass-produced pages primarily to manipulate rankings. Unhelpful content fails whether a human or a model wrote it.

"Just add statistics and quotes to get cited." Research benchmarks suggest such elements can help, but invented statistics or irrelevant quotes don't create trust. Specific, sourced facts help because they're useful.

"Rewrite everything into tiny chunks for vector databases." Clear structure and self-contained sections help, but chopping content into disconnected fragments harms readers. Google's 2026 guide says explicitly that you don't need to break content into chunks for its AI features.

"Block Google-Extended to get out of AI Overviews." Google-Extended governs Gemini training and grounding outside Search. Leaving AI Overviews and AI Mode is done with the Search generative AI control in Search Console (or blunt snippet controls).

"We can guarantee you'll be recommended by ChatGPT." No one controls these systems' outputs. Guarantees are a red flag.

"AI visibility score X proves success." Vendor scores are proprietary and method-dependent. Useful for trends within one tool; not ground truth.

Manipulative tactics to avoid

Some practitioners try to game AI systems. These tactics create policy, legal and reputational risk:

  • Hidden prompt injection: hidden text on pages addressed to AI systems, trying to steer what they say about a brand. This is hidden text — a long-standing spam violation — and a form of adversarial manipulation. Since May 2026 Google's spam policies explicitly define spam as including attempts to manipulate generative AI responses in Search. Providers actively defend against it, and it can backfire badly if discovered.
  • Mass-produced AI pages targeting every prompt variation: scaled content abuse.
  • Fake reviews and sock-puppet forum posts to shape consensus: platform violations and unlawful in many jurisdictions.
  • Cloaking: showing AI crawlers different content from users — a spam policy violation for search engines and deceptive for users.
  • Paying for undisclosed placements in "best of" lists: may breach advertising disclosure rules; link-related parts may breach link spam policies.
  • Parasite content: publishing on high-reputation third-party sites primarily to exploit their signals (site reputation abuse).

A healthy scepticism checklist for client conversations

  • "Show me the method." Any claim of improvement needs baseline, prompts, runs and dates.
  • "What would we lose if this is wrong?" Cheap, harmless experiments (a clearer fact sheet, llms.txt on a docs site) are fine. Risky ones (hidden text, fake reviews) are not.
  • "Does it help users?" If a tactic only works by deceiving machines, it's fragile and likely against policy.

Where uncertainty is genuine

Be candid about open questions: how much unlinked mentions influence recommendations; how much training data vs retrieval shapes answers for your category; how licensing deals affect which publishers get cited; how AI features will affect click-through long term. Say "we don't know yet; here's how we'll measure it."

Hands-on: a claim-evaluation worksheet

Use this for any tactic a client, vendor or colleague proposes:

Claim:                    "Adding X will get us cited by AI Overviews / ChatGPT"
Source:                   provider docs | peer-reviewed | transparent test | vendor | anecdote
Method (if a test):       prompts=__ runs=__ surfaces=__ markets=__ control group? Y/N dates=__
Mechanism:                how would retrieval/selection plausibly reward this?
Policy check:             Google spam policies (incl. AI manipulation, May 2026)? platform rules? law?
Cost if wrong:            low (reversible, helps users) | high (policy/legal/reputation risk)
Decision:                 adopt | small experiment with measurement | reject

Worked example 2: a vendor pitch in Riyadh

A vendor offers a Riyadh retailer (illustrative) "guaranteed AI Overview placement" through a network of partner sites that will publish articles citing the retailer, plus "AI-optimised" hidden summaries on product pages. The worksheet: source = vendor; method = screenshots only; mechanism = manufactured corroboration and hidden text; policy check = site reputation abuse, link spam, hidden text and Google's explicit AI-manipulation language; cost if wrong = manual action risk and reputational damage. Decision: reject, and redirect budget to a genuine review programme and a better comparison page.

Common mistakes

  • Repeating vendor claims to clients as facts.
  • Letting fear of AI drive blanket blocking or panic content changes.
  • Adopting manipulative tactics because "everyone is doing it".

Video lecture: Myths, manipulation and what not to believe

Lecture coming soon · 12 chapters · about 8 minutes. Read the full transcript below.

  1. Myths, manipulation and evidence
  2. The four-question test
  3. Myths 1–2
  4. Myths 3–6
  5. Myths 7–8
  6. Tactics to avoid
  7. May 2026: spam policy update
  8. Example 1: hidden 'best dentist' text
  9. Example 2: 'guaranteed placement' vendor (illustrative)
  10. Watch me do it: evaluating a live pitch
  11. Honest uncertainty
  12. Recap and try this now

Lecture transcript

Myths, manipulation and evidence

AI search is new, commercially important and opaque. That's the perfect environment for hype. Every week someone promises a secret file, a special markup or a guaranteed placement in ChatGPT. Some claims are reasonable. Many are marketing. A few are genuinely dangerous. In this lecture you'll learn a four-question test for any claim, the most common myths and what the evidence says, and the manipulative tactics that can get a site penalised, including what changed in Google's spam policies in twenty twenty-six.

The four-question test

Here's the four-question test. One: source. Is this from the provider's documentation, peer-reviewed research, a transparent experiment, or a vendor selling something? Two: method. How many prompts, runs, surfaces and markets? Was there a control group? Three: mechanism. Does it fit how retrieval and generation plausibly work? Four: durability. Would it survive a model update, or the provider noticing it? Most bad advice fails at least two of these questions within about thirty seconds.

Myths 1–2

Myth: SEO is dead. The evidence says AI search experiences largely depend on search indexes and web content. Crawlability, relevance, quality and reputation, the core of SEO, remain prerequisites. Google's own twenty twenty-six guide calls AEO and GEO still SEO. What's changing is the shape of traffic and the importance of brand and third-party presence. Myth: you need special schema to appear in AI Overviews. Google says no special schema is needed. Schema helps understanding and some rich results, but there's no public evidence it's a citation switch.

Myths 3–6

Myth: llms.txt boosts AI rankings. No major engine has committed to using it for ranking or citation, and Google says Search ignores it. Myth: AI-written content is penalised. Google's position is that it rewards helpful content however it's produced. What it targets is scaled content abuse: mass-produced pages made mainly to manipulate rankings. Myth: rewrite everything into tiny chunks. Google says explicitly that you don't need to. Myth: block Google-Extended to leave AI Overviews. Wrong lever. That's Gemini training and grounding. The Search Console AI control is the lever for AI Overviews and AI Mode.

Myths 7–8

Two more myths that clients love. We can guarantee you'll be recommended by ChatGPT. No one controls these systems' outputs, so a guarantee is a red flag. And, our AI visibility score proves success. Vendor scores are proprietary and depend on method. They're useful for trends within one tool, not ground truth. The healthy response to both is the same: show me the method.

Tactics to avoid

Now the tactics to avoid, because some practitioners try to game AI systems directly. Hidden prompt injection: hidden text addressed to AI systems, trying to steer what they say about a brand. Mass-produced AI pages targeting every prompt variation, which is scaled content abuse. Fake reviews and sock-puppet forum posts to shape consensus. Cloaking, meaning showing crawlers different content from users. Paying for undisclosed placements in best-of lists. And parasite content on high-reputation sites, which Google calls site reputation abuse.

May 2026: spam policy update

Here's what changed in May twenty twenty-six. Google updated its spam policies so that the definition of spam includes attempting to manipulate generative AI responses in Google Search. That makes explicit what was already implied. Buying or faking citations, planting hidden instructions, manufacturing mentions. All of it can be treated as spam, with the same consequences as any other spam, including manual actions. So when someone says, this is just for the AI, not for Google, the answer is, for Google's AI features, there's no difference.

Example 1: hidden 'best dentist' text

Worked example one, simple. A consultant tells a Leeds dental practice to add a hidden paragraph listing it as the best dentist in Leeds, in white text, for AI to pick up. Run the test. Source: a consultant's anecdote. Method: none. Mechanism: hidden text, a long-standing spam violation. Durability: zero once noticed. Cost if wrong: a manual action on the practice's main acquisition channel. Decision: reject in about a minute, and spend the effort on genuine patient reviews instead.

Example 2: 'guaranteed placement' vendor (illustrative)

Worked example two, with illustrative details. A vendor offers a Riyadh retailer guaranteed AI Overview placement through a network of partner sites that will publish articles citing the retailer, plus AI-optimised hidden summaries on product pages. The worksheet: source is the vendor; the method is screenshots; the mechanism is manufactured corroboration and hidden text; the policy check flags site reputation abuse, link spam, hidden text and the AI-manipulation language. Cost if wrong: a manual action and reputational damage. Decision: reject, and redirect the budget to a genuine review programme and a better comparison page.

Watch me do it: evaluating a live pitch

Watch me do it. A client forwards me a pitch this morning: we'll get you into AI Overviews in thirty days with our AI-optimised schema and citation network. I open the claim-evaluation worksheet and fill it live. Claim: special schema plus a citation network gets AI Overview placement in thirty days. Source: the vendor. Method: I reply asking for prompts, runs, markets and a control group. They send three screenshots. So, method: screenshots only. Mechanism: Google's guide says no special schema is needed for its AI features, and a citation network means paid articles on partner sites mentioning the client. Policy check: that looks like link spam and possibly site reputation abuse, and since May twenty twenty-six Google's spam definition explicitly includes attempting to manipulate generative AI responses. Cost if wrong: a manual action on the client's main channel, plus reputational risk. Decision: reject. Then I do the useful part. I write back to the client with three lines: why we're saying no, what the evidence shows, and what we'll do instead with that budget: a review programme and a fairer comparison page. The whole worksheet takes fifteen minutes and protects the client for years.

Honest uncertainty

Be candid about genuine uncertainty. We don't fully know how much unlinked mentions influence recommendations. How much training data versus retrieval shapes answers in your category. How licensing deals affect which publishers get cited. Or how AI features will affect click-through long-term. Saying, we don't know yet, and here's how we'll measure it, builds more trust with clients than false confidence ever will. Common mistakes: repeating vendor claims as facts, letting fear drive blanket blocking, and adopting manipulative tactics because everyone is doing it.

Recap and try this now

Recap. Test every claim on source, method, mechanism and durability. Most myths fail quickly, and Google's twenty twenty-six guidance settles several of them. Manipulative tactics are explicitly spam when aimed at Google's AI answers. Try this now. Take the last AI search tactic someone pitched to you, and run it through the claim-evaluation worksheet in the lesson text. Decide: adopt, small measured experiment, or reject.

Key takeaways

  • Evaluate claims by source, method, mechanism and durability.
  • Myths: SEO is dead, special schema is required, llms.txt boosts rankings, AI content is penalised, guaranteed recommendations.
  • Avoid hidden prompt injection, scaled AI pages, fake reviews, cloaking, undisclosed paid placements and parasite content.
  • Be candid about genuine uncertainty and measure instead of guessing.

Try it

Collect five AI search claims from blogs or social media and evaluate each using the four questions; label each evidence-based, plausible, or hype.