SEO for AI Search: What Actually Changes — and What the Evidence Doesn’t Yet Prove

A business can rank well on Google and still disappear when the same buyer asks an AI system whom to consider.

That is the part of AI Search worth paying attention to.

It is also where the industry has moved far too quickly from observation to prescription.

Businesses are now being told to rewrite pages into 40-word answer blocks, add FAQ schema, publish llms.txt, build Reddit mentions, optimise “entities”, create pages for every conversational query, write for vector chunks, refresh content every quarter and track citations instead of rankings.

Some of that advice is sensible.

Some of it is supported by correlation.

Some applies to one platform but not another.

Some has experimental evidence behind it.

And some is being repeated with far more certainty than the evidence allows.

At KickAss Digital Marketing, we did not want to publish another “11 ways to optimise for AI Search” article.

We wanted to answer a harder question:

What actually changes in SEO when the outcome expands from ranking in Search to being retrieved, cited, mentioned or recommended inside an AI-generated answer — and which of the supposed new tactics can we genuinely defend?

Our conclusion is less dramatic than “SEO is dead”.

It is also more useful than “nothing has changed”.

SEO has not been replaced. The discovery system above it has expanded.

And the architecture of that new layer has changed faster than the evidence for many tactics now being marketed as AI SEO, GEO or AEO.

If you are evaluating an agency for SEO and AI Search Visibility, that distinction matters. You need to know which parts of the work are established search practice, which genuinely respond to a discovery environment, and which are still experiments dressed up as ranking factors.

First, even “AI SEO” does not mean one thing

Before discussing optimisation, there is a terminology problem to clear up.

In our August 2026 live Google audit in India, the AI Overview for AI SEO described it as optimising content to be discovered, cited and recommended by generative systems.

On the same SERP, Semrush used “AI SEO” for something materially different: using artificial intelligence to improve the conventional SEO workflow — keyword research, content production, analysis and automation.

Same phrase. Different job.

GEO, AEO, AI SEO, ChatGPT SEO, LLM SEO and AI Search optimisation are now used with overlapping meanings.

This is why we prefer SEO for AI Search when describing the wider problem.

Google itself says its existing SEO best practices remain relevant to AI Overviews and AI Mode and that no additional technical requirements or special optimisations are required simply to appear in those features.¹

At the same time, a genuinely different discovery layer is emerging above those foundations.

That is also why ShodhDynamics’ analysis of ChatGPT SEO, GEO and AEO argues that the labels matter less than the underlying model of how generated answers retrieve and select information.

KickAss still treats Generative Engine Optimisation and Answer Engine Optimisation as useful specialised disciplines. They describe different optimisation problems.

They do not require us to pretend that conventional SEO has stopped mattering.

SEO has not disappeared. The system above it has expanded.

The strongest evidence is also the least controversial.

For Google’s AI features, foundational SEO still matters.

Google states that supporting pages for AI Overviews and AI Mode need to be indexed and eligible to appear in Search with a snippet. It also says there are no additional technical requirements for inclusion.¹

That means the fundamentals did not suddenly become obsolete:

• crawlability
• indexability
• useful and original content
• relevance
• internal architecture
• accessible text
• links and reputation
• accurate structured data where appropriate
• and the ability to satisfy a real information need

Google’s May 2026 guidance goes further. It explicitly recommends valuable, unique, non-commodity content and describes the new resource as myth-busting common AEO/GEO misconceptions.²

That is important.

AI Search could easily become the next excuse for scaled content production: one URL for every possible fan-out query, hundreds of mechanically generated FAQs and endless versions of information that already exists.

That is almost the opposite of the opportunity.

If Gemini, ChatGPT or an AI Overview can already synthesise an adequate answer from thousands of existing pages, publishing another generic explanation gives the system little reason to select yours.

What changes is the layer above eligibility

Google documents one particularly important difference: query fan-out.

AI Overviews and AI Mode can issue multiple related searches across subtopics and data sources while constructing an answer.¹

ChatGPT Search documents a related behaviour. OpenAI says ChatGPT may rewrite a user’s request into one or more targeted searches and issue additional, more specific searches after reviewing initial results.³

That changes the retrieval surface.

It does not invalidate SEO.

It expands the problem SEO has to solve.

Consider a real-estate developer in Goa.

The developer may rank well for:

luxury villas in North Goa

But a buyer asks:

Which villa developments in North Goa would suit an NRI family looking for a reputed developer, quiet surroundings, reasonable airport access and a property that will be easy to manage while they are overseas?

That is not simply a longer keyword.

It contains multiple information needs:

developer credibility, project location, accessibility, property characteristics, maintenance, buyer profile and probably third-party evidence.

An AI system can search across those sub-problems before constructing the answer.

The landing page ranking for the literal head term still matters.

It is no longer the whole evidence environment.

Ranking is not retrieval. Retrieval is not citation. Citation is not recommendation.

This is one of the most important distinctions in the entire AI visibility discussion.

A business can move through several different states:

Eligible → Retrieved → Used → Cited → Mentioned → Recommended → Commercial Outcome

They are not interchangeable.

Eligibility means the system can potentially access the information.

Retrieval means a source or passage enters the candidate information set for that request.

Use means information from that source contributes to the answer.

Citation means the interface visibly attributes something to the source.

Mention means the actual business or brand appears in the generated answer.

Recommendation is more demanding: the system presents that brand as an appropriate option for the user’s need.

And only after that do we get to what a marketing team ultimately cares about: preference, branded search, a website visit, enquiry, shortlist, booking or sale.

This distinction becomes especially important in high-consideration categories.

A hotel may be used as a source for factual information without appearing in the shortlist.

A publication might be cited while the brands it discusses get mentioned.

A real-estate portal may supply project facts while a developer’s own website receives no citation.

And an AI system can recommend a company without producing a click at all.

ShodhDynamics describes this wider structural change through its work on AI Decision Funnels:

“If marketing funnels optimise outcomes, AI decision funnels determine who is even considered.”

That is a more commercially useful way to think about AI visibility than treating every citation as a new version of keyword rank.

The AI Discovery Lexicon deliberately separates concepts such as Prompt-Level Visibility, Source Gravity, Answer Compression and the Shortlist Moment because generated-answer presence is not the same thing as conventional click-based discovery.

So how much does Google ranking actually explain?

Ranking still matters.

What the evidence does not support is treating organic position as a deterministic AI citation position.

In March 2026, Ahrefs analysed 863,000 keyword SERPs and four million AI Overview URLs.⁴

For conventional blue links:

• 37.1% of AI Overview-cited URLs ranked in the top 10 for the same query;
• 26.2% ranked between positions 11 and 100;
• 36.7% did not rank in the top 100 for that literal query.

Those numbers do not prove rankings are becoming irrelevant.

They establish something narrower:

The user’s literal-query ranking does not fully explain source selection in AI Overviews.

That is a critical distinction.

The same Ahrefs research explicitly points to query fan-out as a plausible explanation for the widening difference between the original SERP and the URLs ultimately cited.⁴

Google documents that fan-out exists.¹

But we cannot see every hidden query.

Therefore we cannot legitimately say:

“This URL was cited because it ranked #3 for fan-out query X.”

unless that query was actually observable.

Nor can we turn fan-out into another content-production formula:

“Publish one page for every sub-query and citations will increase.”

That causal step has not been established.

The mechanism is documented.

The weighting is not.

We asked four AI systems what SEOs should change. Then we challenged their evidence.

One of the more revealing parts of this investigation was not a third-party study.

It was what happened when we asked AI systems themselves.

We gave the same underlying question to:

• ChatGPT
• Gemini
• Claude
• Perplexity

Then we did something that materially changed several answers.

We asked each system to go back through its recommendations and separate:

platform-documented mechanisms, experimental evidence, observational correlation, reasonable inference and common industry claims without strong evidence.

The eventual consensus was much narrower than the first-pass advice.

Across the panel, the strongest common ground was:

• crawlability and indexability still matter;
• useful, original content still matters;
• query rewriting/fan-out changes retrieval;
• platform-specific crawler controls create new technical considerations;
• citation, mention and recommendation are different outcomes;
• organic ranking and AI citation are related but not deterministic;
• generic schema should not be promised as a citation booster;
• llms.txt is not established as a Google AI Search ranking lever;
• and AI visibility needs measurement beyond a standard keyword report.

What was more interesting was what changed after pressure testing.

System	Initial tendency	After evidence challenge
ChatGPT	Already relatively cautious about causal claims	Became even more explicit about separating retrieval from downstream use and correlation from mechanism
Claude	Presented entity establishment and third-party mentions as strong new levers	Later acknowledged confounding and weakened causal language
Gemini	Made relatively confident schema/entity and crawler claims	Later separated some documented crawler functions more carefully
Perplexity	Prescribed concrete answer/passsage formats and lengths	Later admitted there was no strong controlled evidence for exact formulas, although some formulaic recommendations returned in its final advice

This is not evidence that one AI system is “better” at SEO.

That was not the test.

The important finding is that evidence pressure changed the advice.

A recommendation that sounds established in a generated answer may become substantially more qualified when the same model is asked:

What exactly proves that?

That matters because marketers are increasingly using AI-generated advice to build AI-search strategy.

We are in a period where the systems summarising the evidence can also amplify the industry’s uncertainty.

Google’s own AI Overviews showed the same confidence gap

After the four-system exercise, we manually searched ten semantic variations of the topic in Google from India on 20 August 2026.

The wording changed.

The underlying intent did not.

Queries included:

• SEO for AI search
• AI SEO
• how to optimize for AI search
• does SEO still matter for AI search
• how does AI search choose sources
• does ranking help AI citations
• SEO strategy for Google AI Mode
• SEO for ChatGPT
• schema for AI search
• is GEO different from SEO

The AI Overviews repeatedly mixed statements with very different levels of evidentiary support.

For SEO for AI search, Google recommended direct summaries, topical clusters, structured data, unique expertise, quarterly freshness and external-platform presence as “Core Optimization Tactics”.

For how to optimize for AI search, the overview became more prescriptive: put 40–60-word answers beneath major headings, turn H2/H3s into questions, implement schema and maintain single-intent pages.

For how does AI search choose sources, the overview described a clean pipeline of query expansion, passage retrieval, trust filtering and grounding, then presented clarity, topical authority, consensus and freshness as source-selection signals.

Some of those practices may be useful.

But the confidence of the presentation exceeds what can currently be verified as a universal causal mechanism.

That is the distinction this article is making.

Not:

“Ignore clarity.”

Not:

“Never use schema.”

Not:

“Freshness does not matter.”

But:

Do not confuse a sensible practice, an observed correlation and a documented ranking mechanism.

The schema problem shows exactly how claim inflation happens

Schema is useful.

We use structured data extensively ourselves.

It can represent organisations, people, services, products, offers, authorship and relationships in machine-readable form. It serves established Search functions and can reduce ambiguity in how facts are expressed.

But that is different from saying:

Add schema and AI systems will cite you more.

Ahrefs first examined six million URLs and found that AI-cited pages were almost three times more likely to contain JSON-LD than non-cited pages.⁵

On its own, that is easy to turn into a LinkedIn carousel:

Schema = AI visibility.

But technically sophisticated websites are also more likely to publish better content, attract links, maintain their sites properly and have stronger authority.

Schema may simply co-exist with all of that.

So Ahrefs ran a second study.

It tracked 1,885 pages that added JSON-LD and matched them against 4,000 control pages. Citation changes in ChatGPT and Google AI Mode were statistically indistinguishable from zero; the authors also cautioned against reading a small relative decline in AI Overview citations as schema causing harm.⁵

That is a much better example of how SEO evidence should be interpreted.

“AI-cited pages often have schema”

and

“adding schema causes additional AI citations”

are not the same claim.

Google itself says there is no special schema required for AI Overviews or AI Mode.¹

Our conclusion is therefore straightforward:

Use structured data when it accurately represents the page, supports Search functionality, improves data clarity or serves an operational purpose. Do not sell generic schema as a proven independent AI-citation lever.

This distinction also matters in our own AI Readiness work. Machine-readable infrastructure can be valuable without pretending that every infrastructure layer is a disclosed ranking factor.

llms.txt deserves the same evidence discipline

llms.txt is another good example of a promising infrastructure idea being converted too quickly into a visibility claim.

Google explicitly says website owners do not need new machine-readable AI files to appear in AI Overviews or AI Mode.¹

Ahrefs subsequently analysed 137,210 domains in May 2026.⁶

Around 28% published a valid llms.txt.

Of those files, 97% received no request at all during the measured month.

That does not prove llms.txt will never matter.

It does not prove that no current specialised agent uses it.

And it certainly does not tell us what future agentic systems will adopt.

What it does tell us is that the current evidence cannot support:

“Add llms.txt and your AI rankings will improve.”

We have an existing guide to llms.txt implementation on this site. This newer research gives us reason to be more precise about what the file currently demonstrates versus what remains experimental.

That distinction matters because KickAss does use machine-readable infrastructure.

Our position is not that experimentation should stop.

It is that implementation and proven visibility impact are two different claims.

What about answer-first writing, short paragraphs and “AI-friendly” structure?

This is where the discussion needs more nuance.

Clear writing is useful.

Good headings are useful.

Putting the answer before 500 words of generic introduction is usually better for a human reader too.

Tables can make comparisons easier.

Specific facts are more useful than vague marketing copy.

None of those practices needs a secret LLM ranking factor to justify it.

There is also evidence worth studying.

Semrush compared 304,805 AI-cited URLs against 921,614 Google-ranking URLs across 11,882 prompts and found positive associations with qualities including clarity, summarisation, Q&A formatting and section structure.⁷

But the study examined visible textual features.

It did not prove:

rewriting an existing page into 40-word answer blocks will independently cause the page to be retrieved and cited more often.

That distinction matters.

The original Generative Engine Optimization research accepted at KDD 2024 also demonstrated that certain changes to source material could alter visibility inside the generative environments tested, with results varying by intervention and domain.⁸

That is meaningful evidence.

But the study does not justify converting:

“some interventions changed visibility within the tested setup”

into:

“do this on your live website and ChatGPT will retrieve you 40% more often.”

The missing step is organic retrieval.

A model using a document differently after that document has already entered its context is not the same thing as proving an edit caused the live search system to retrieve that document.

That remains one of the most important open questions in AI-search optimisation.

Third-party evidence matters — but we still need to be careful about why

AI-generated recommendations make the business’s own website only one part of the evidence environment.

Independent sources can become important.

Pew Research Center analysed 68,879 Google searches in 2025 and found that Wikipedia, YouTube and Reddit collectively accounted for 15% of sources cited in AI summaries.⁹

That supports a practical conclusion:

the wider information environment around a brand matters.

It does not establish:

“Get more Reddit mentions and ChatGPT will recommend you.”

Those are different claims.

This is where ShodhDynamics’ ESC™ Framework is useful as an analytical model.

ESC™ separates:

Entity Clarity
Can the system establish what the entity actually is?

Semantic Authority
Is there sufficient substantive information connecting the entity to the subject?

Cross-Source Trust
Do independent sources provide enough consistent evidence for the system to reason about the entity confidently?

ESC is not presented as Google’s ranking algorithm or OpenAI’s recommendation formula.

That is exactly how a framework should be used.

It helps structure the problem without pretending to expose hidden model weights.

This is not only an SEO-team problem

It is easy for this discussion to become too technical.

For a marketing director, CMO or founder, the practical problem often looks much simpler.

Your business ranks.

Your competitors appear when buyers ask AI for recommendations.

You do not.

Why?

Consider a hotel in Singapore.

The hotel’s website can be technically excellent. Its room pages can rank. Its local information can be accurate. Its structured data can be valid.

Then a traveller asks:

Where should I stay in Singapore for a three-night business trip if I need late check-in, easy access to Marina Bay, restaurants within walking distance and somewhere quiet enough to work in the evening?

The answer may require evidence about:

• location;
• transport;
• late check-in;
• guest experience;
• nearby dining;
• business suitability;
• reviews;
• neighbourhood context;
• and possibly current availability or policy information.

The problem is no longer simply:

Does the hotel page rank?

It becomes:

Does the wider evidence environment give the answering system enough accessible, relevant and credible information to include that hotel in the shortlist?

The same applies to a real-estate developer in Goa, a travel business in Dubai or a B2B company in Gurugram.

Search visibility still matters.

But discovery and evaluation are beginning to overlap.

That is the business reason to care about AI Search.

Not because SEO has acquired more acronyms.

The evidence is strong on architecture. It is much weaker on many tactics.

The easiest way to lose perspective is to put every AI-search recommendation into one bucket.

They do not belong there.

Claim	Current evidence status	How we would treat it
Foundational SEO still matters for Google AI features	Documented	Act on it
Google uses query fan-out	Documented	Adapt research and content architecture
ChatGPT can rewrite prompts into multiple searches	Documented	Adapt query-family research
OAI-SearchBot access matters for ChatGPT Search inclusion	Documented	Check technical access
Organic ranking and AI citation are related	Observed / correlational	Use ranking as one visibility input, not a guarantee
Clear, well-structured content correlates with citations	Observed / correlational	Apply where useful; test impact
Third-party sources frequently appear in AI answers	Observed	Analyse source ecology and reputation
Adding generic schema increases AI citations	Not established	Use schema for legitimate purposes; test separately
llms.txt increases AI ranking/citation	Not established	Treat as infrastructure experiment
Exact 40–60-word answers improve retrieval	Not established	Do not make it a rule
Fixed 130–170-word chunks improve citations	Not established	Do not make it a rule
More Reddit mentions cause recommendations	Not established	Do not manufacture mentions
#1 Google ranking guarantees AI citation	Contradicted by observed data	Measure both separately

This is the part most “AI SEO checklists” remove.

Uncertainty.

But uncertainty is not a weakness in strategy.

It tells us what to build, what to measure, and what to test.

The KickAss BASE™ Model for SEO & AI Search

After reviewing the evidence, we needed a practical way to prevent two opposite mistakes:

1. throwing away established SEO because AI Search feels new;
2. treating every new AI-search theory as proven.

Our working model is BASE™.

B — Build on the SEO foundation

Keep investing in what still establishes search eligibility, relevance and competitiveness:

technical SEO, crawlability, indexability, useful content, internal architecture, authority, links, page quality and genuine topical coverage.

For Google, this is explicitly supported by its own guidance.¹

Do not try to repair weak SEO with “GEO formatting”.

A — Add what AI Search genuinely changes

Some additions are justified by documented changes.

Audit platform-specific crawler access.

Research query families and likely fan-out territory instead of treating a single keyword as the entire demand surface.

Measure citations, mentions and recommendations separately.

Study the source environment beyond your own domain.

And monitor the platforms where your actual buyers are asking questions.

Google itself has begun separating this layer in reporting. In June 2026 it launched dedicated generative-AI visibility reporting in Search Console for a subset of websites, including impressions, pages, countries, devices and dates for AI Search surfaces.¹⁰

S — Study what looks promising

This is where serious experimentation belongs.

Test whether clearer content changes live retrieval.

Test structured-data interventions without simultaneously rewriting the page.

Test entity-data clean-up.

Test third-party evidence.

Test whether results differ by industry.

Test whether Goa, Singapore, Dubai or Delhi NCR produce different recommendation environments.

Repeat prompts rather than trusting one screenshot.

Most importantly, define the outcome before the test.

Are we measuring:

retrieval? citation? mention? recommendation? traffic? enquiry?

If those are mixed together, the test tells us very little.

E — Exclude unsupported assumptions

Do not build a client strategy around claims such as:

• every answer must be 40–60 words;
• every page needs FAQ schema for AI;
• LLMs prefer one universal chunk length;
• llms.txt is an AI ranking factor;
• unlinked mentions have replaced backlinks;
• Reddit mentions automatically increase recommendations;
• traditional ranking no longer matters;
• or GEO now has a settled list of ranking factors comparable with Google SEO.

Our research does not support treating those statements as established fact.

A useful pressure test for any new tactic is:

What exactly establishes this?

If the evidence chain ends with an agency blog quoting another agency blog quoting an AI-generated answer, that is not enough.

What should a marketing director expect from an SEO agency now?

The SEO agency’s job is widening.

Not because somebody renamed SEO as GEO.

Because the environment in which discovery happens is widening.

If you are choosing an agency, it should be able to explain:

what part of its work strengthens conventional Search;

what is specifically relevant to Google AI Overviews and AI Mode;

what additional crawler or access controls matter for systems such as ChatGPT;

how it researches semantic query families rather than simply matching exact keywords;

how it evaluates the wider evidence environment around the brand;

how it distinguishes citation, mention and recommendation;

how it measures AI visibility without pretending one prompt is a stable rank;

and which parts of its programme are proven practice versus controlled experimentation.

That matters more than whether the proposal says SEO, GEO, AEO or AI SEO on the cover.

The terminology is easy.

Evidence discipline is harder.

What actually changes in SEO for AI Search?

The shortest defensible answer is:

The foundation changes less than the discovery and decision layer above it.

SEO still helps determine whether information is crawlable, indexable, relevant, useful and competitive.

AI Search adds additional behaviour above that foundation:

• queries can be rewritten or fanned out;
• different sources can be brought together;
• the page ranking for the literal query may not be the page ultimately cited;
• a source can contribute without its brand being named;
• a brand can be mentioned without being recommended;
• a recommendation can influence a buyer without producing a click;
• and the success of that process cannot be represented by one keyword position.

That is a real change.

But it does not follow that every tactic being sold under AI SEO or GEO has therefore become a proven mechanism.

That is the line businesses should hold.

Build what is established.
Add what the new environment clearly requires.
Study what may work.
Exclude what the evidence cannot yet support.

That is the logic behind the KickAss BASE™ Model for SEO & AI Search.

For now, that is a more defensible operating model than pretending anybody has already reverse-engineered the final AI-search playbook.

Frequently Asked Questions related to SEO for AI Search

Is SEO still relevant for AI Search?

Yes. SEO remains relevant because AI-powered search still depends on accessible, indexable, relevant and trustworthy web content. Google explicitly says its existing SEO best practices continue to apply to AI Overviews and AI Mode. What changes is that AI systems can retrieve across multiple queries and sources before deciding what to cite, mention or recommend.

How do you optimise for AI Search?

Optimising for AI Search starts with strong SEO, then adds attention to query fan-out, AI crawler access, source-level evidence, machine-readable clarity and measurement of citations, mentions and recommendations. Practices such as fixed answer lengths, generic schema for citation gain or llms.txt as a ranking factor should be tested rather than treated as proven requirements.

How does AI Search choose sources?

There is no publicly documented universal formula for how AI Search chooses sources. Major systems can retrieve information through search, query rewriting or fan-out and then select material for a generated answer. Relevance, accessibility and source quality clearly matter, but the exact weighting behind retrieval, citation and recommendation is not publicly disclosed and varies by platform.

Does ranking higher on Google help you get cited by AI?

Higher Google rankings are associated with AI citations, but they do not guarantee them. In Ahrefs’ March 2026 analysis, only 37.1% of AI Overview-cited URLs also ranked in the conventional top 10 for the same query. Query fan-out means AI systems may retrieve sources through related searches that differ from the user’s literal query.⁴

Is GEO different from SEO?

GEO and SEO describe different optimisation outcomes, but GEO does not replace SEO. SEO establishes discoverability, relevance and authority in search; GEO focuses more specifically on visibility inside generated answers, including citations and brand representation. In practice, effective GEO usually depends on a strong SEO foundation plus additional work around AI retrieval and source ecosystems.

Is AI SEO the same as GEO?

Not consistently. “AI SEO” is used both for optimising visibility in AI-generated search and for using AI tools to perform conventional SEO work. GEO has a narrower meaning: optimising how content or brands appear in generative answers. Because industry terminology is still fragmented, the underlying objective matters more than the label.

Do you need schema markup for AI Search?

Schema markup is useful for representing facts and entities in machine-readable form, but current evidence does not show that adding generic schema independently causes more AI citations. Google says no special AI-specific structured data is required for its generative Search features. Use schema where it accurately describes the content, not as a guaranteed citation hack.⁵

Do you need llms.txt for AI Search?

No major search platform currently requires llms.txt as a general AI Search ranking mechanism. Google explicitly says it is not required for AI Overviews or AI Mode.¹ It can still be used experimentally as machine-readable infrastructure for systems that choose to support it, but its effect on retrieval or citations should not be assumed.

Should I optimise separately for ChatGPT?

Some platform-specific differences are real. For example, OpenAI says OAI-SearchBot access is relevant to inclusion in ChatGPT Search.¹¹ But that does not mean a completely separate content strategy is required for every LLM.

Should every section start with a 40–60-word answer?

There is no strong evidence supporting 40–60 words as a universal AI-search optimisation rule. Direct, clear answers can improve readability and extractability, but exact lengths should be treated as a hypothesis rather than a ranking factor.

Research Methodology

This investigation was conducted in August 2026.

We deliberately used several evidence layers rather than relying on one SEO tool, one AI system or one SERP snapshot.

Stage 1 — Evidence review

We reviewed official platform documentation, academic research and disclosed industry datasets covering:

Google AI Overviews and AI Mode, ChatGPT Search, query fan-out, generative-engine optimisation, schema, llms.txt, crawler access, rankings, citations and AI visibility.

Primary platform documentation was given greater weight than agency or vendor claims.

Stage 2 — Four-system interrogation

The same initial research question and three evidence-challenge follow-ups were run through:

ChatGPT, Gemini, Claude and Perplexity.

The prompts asked the systems to distinguish:

• documented mechanisms;
• experimental evidence;
• observational correlation;
• reasonable inference;
• and unsupported industry claims.

Stage 3 — Cross-system analysis

The four raw outputs were then analysed together as a closed evidence set.

We examined:

terminology, recurring sources, agreement, disagreement, evidence quality, causal overreach and whether individual systems changed their recommendations after being challenged.

Stage 4 — Manual Google Search + AI Overview audit

On 20 August 2026, we manually searched Google from an incognito browser session in India across ten semantically related intents.

We recorded:

• AI Overview output where triggered;
• conventional organic results;
• source types;
• major claims;
• recurring entities;
• People Also Ask themes;
• and contradictions between generated advice and stronger underlying evidence.

We did not treat Google AI Mode as part of this manual Stage 4 sample.

Limitations

This is an evidence audit, not a reverse-engineering study.

AI responses can vary by model version, geography, account state, personalisation and time.

A visible citation does not reveal every source retrieved internally.

A missing citation does not prove a source was never retrieved.

And observational relationships should not be converted into causal ranking factors.

Research Dataset

The first-party research package for this investigation contains:

1. Deep Research evidence report
2. ChatGPT raw conversation
3. Gemini raw conversation
4. Claude raw conversation
5. Perplexity raw conversation
6. Consolidated four-system evidence analysis
7. Manual Google SERP + AI Overview audit

DOWNLOAD – SEO for AI Search Consolidated ZIP

Publishing the raw material is recommended so readers can inspect the AI outputs, source trails and evidence challenges rather than relying only on our interpretation.

Source Notes & Editorial Verification

Editorial verification date: 20 August 2026

Platform documentation and time-sensitive research findings should be rechecked before major future revisions because AI Search interfaces, models and reporting can change rapidly.

1. Google Search Central — AI features and your website. Official documentation covering foundational SEO, technical eligibility, query fan-out, structured data and AI-specific file requirements. Open Google documentation
2. Google Search Central — A new resource for optimizing for generative AI in Google Search. May 2026 announcement emphasising unique, non-commodity content, myth-busting AEO/GEO claims and foundational SEO. Open Google announcement
3. OpenAI — ChatGPT Search. Official documentation describing query rewriting into one or more targeted searches and additional searches after initial retrieval. Open OpenAI documentation
4. Ahrefs — Update: 38% of AI Overview Citations Pull From the Top 10. March 2026 analysis of 863,000 SERPs and four million AI Overview URLs. Open Ahrefs study
5. Ahrefs — We Tracked 1,885 Pages Adding Schema. AI Citations Barely Moved. Matched intervention analysis examining JSON-LD additions against 4,000 controls. Open schema study
6. Ahrefs — We Analyzed 137K Sites: 97% of llms.txt Files Never Get Read. Server-log study of 137,210 domains during May 2026. Open llms.txt study
7. Semrush — How We Built a Content Optimization Tool for AI Search. Study comparing cited and Google-ranking URLs for textual characteristics associated with AI citations. Open Semrush study
8. Aggarwal et al. — GEO: Generative Engine Optimization. KDD 2024 research introducing GEO-bench and testing interventions within generative-engine environments. Primary academic source retained in research dataset/source implementation.
9. Pew Research Center — Google users are less likely to click when an AI summary appears. Behavioural analysis of 68,879 Google searches. Canonical Pew source retained in research source ledger.
10. Google Search Central — Search Generative AI performance reports in Search Console. June 2026 announcement of dedicated generative-AI visibility reporting. Open Search Console announcement
11. OpenAI — Publishers and Developers FAQ. Official documentation covering OAI-SearchBot access and ChatGPT Search inclusion. Open OpenAI publisher documentation