Google published its first dedicated guide to optimizing for generative AI features in Search on May 15, 2026. The guide does not declare every form of Answer Engine Optimization or Generative Engine Optimization useless. It makes a narrower and more important point: several deliverables being sold as special shortcuts have no preferential role in Google Search. For Google, visibility in AI Overviews and AI Mode still begins with the same foundations that support ordinary Search: technical eligibility, crawlability, indexing, useful content, and established ranking and quality systems.
Independent studies support parts of that position, but they need to be described accurately. SE Ranking analyzed nearly 300,000 domains and found no measurable association between the presence of llms.txt and AI citation frequency in its dataset. That was an observational analysis, not a controlled test proving that adding the file can never matter. Ahrefs tracked 1,885 pages that added JSON-LD against matched controls and found no major citation uplift in Google AI Mode or ChatGPT during the measured period. These findings weaken claims that either tactic is a proven visibility lever. They do not establish that the underlying technologies are useless for every platform or use case.
This article separates Google's documented requirements from broader claims about AI search engine optimization. It also explains what Google's generative Search features retrieve, why the guidance should not be generalized to every AI platform, and how to evaluate an AEO or GEO agency without accepting unverifiable metrics.
What Google Says You Do Not Need for AI Search Visibility
"Create non-commodity content that's helpful, reliable, and people-first." - Google Search Central
Google's generative AI Search optimization guide includes a mythbusting section aimed at tactics circulating under AEO and GEO labels. The section applies specifically to Google Search, including AI Overviews and AI Mode. It should not be presented as a universal specification for ChatGPT, Claude, Perplexity, or any other system.
llms.txt and AI-specific reference files receive no special treatment in Google Search.
Google states that it does not use llms.txt, special AI text files, AI-only markup, or Markdown copies as visibility or ranking inputs. Google may still discover and index supported text files, but that does not make them privileged. The guide also leaves room for platform differences: maintaining llms.txt is acceptable when another service or integration uses it. For Google Search itself, the file neither helps nor harms visibility.
The independent evidence is consistent with that narrower claim. SE Ranking found no correlation between llms.txt presence and citation frequency across its sample. Its authors also state that the result is context-dependent and could change as standards evolve. A separate 90-day OtterlyAI experiment recorded 84 requests to /llms.txt among more than 62,100 AI-bot visits on one test site. That result shows limited crawler interest on the site studied; it is not a web-wide usage rate.
Artificial content chunking is not required for Google's AI systems.
Google says its systems can understand multiple topics on a page and retrieve the relevant part. It does not prescribe an ideal page length or require authors to split coherent material into tiny sections for AI comprehension. Short pages can be appropriate. Long pages can also be appropriate. The decision should follow the reader's task, the complexity of the subject, and the need for clear navigation.
This does not make headings, concise paragraphs, summaries, or modular documentation pointless. Those choices can improve usability, accessibility, maintenance, and comprehension. The unsupported claim is that artificial fragmentation creates a special generative Search advantage.
Rewriting content solely for presumed AI phrasing requirements is unnecessary.
Google says its systems understand synonyms and general meaning, so a page does not need to repeat every long-tail variation or adopt a particular "AI-friendly" voice. Rewriting still has value when it improves accuracy, clarity, originality, or usefulness. It is the AI-only justification that lacks support.
Manufactured third-party mentions are an unreliable strategy.
Google's generative features can use information found in articles, videos, reviews, and forums. That does not mean a business should manufacture praise or place low-quality references simply to create a larger mention count. Google states that its core ranking systems and spam systems also apply to its generative Search features. Legitimate coverage and independent evidence may help users and systems verify a claim. Fabricated or coordinated mentions create quality and policy risk.
Structured data remains useful, but Google does not require AI-specific schema.
Google uses supported structured data to understand pages and determine eligibility for rich results. Its structured data documentation still recommends accurate markup for eligible page types. The generative AI guide makes a separate point: there is no special schema.org type required for AI Overviews or AI Mode.
The Ahrefs study adds useful context. It compared 1,885 pages that added JSON-LD with approximately 4,000 matched controls. The measured changes were +2.4% in Google AI Mode and +2.2% in ChatGPT, both statistically indistinguishable from zero. AI Overview citations showed a 4.6% relative decline, but the researchers could not establish that schema caused the decline. The study therefore supports the conclusion that adding JSON-LD was not a reliable citation-growth tactic during the test window. It does not justify removing correct structured data from pages that use it for rich-result eligibility or clearer machine-readable descriptions.
Google's position is not that optimization has disappeared. It is that a separate layer of invented deliverables is unnecessary for its Search products. The work that remains is less theatrical: build pages Google can access, index, understand, and retrieve; publish information that adds something distinctive; and measure performance with evidence that can be audited.
How Google’s Generative Search Actually Retrieves Information
Google's guide explains the retrieval process at a useful operational level. AI Overviews and AI Mode do not depend on a parallel web index built from GEO files. They use information from Google's Search index and rely on existing Search systems to retrieve material relevant to the user's request.
Retrieval-augmented generation
Google describes its process as retrieval-augmented generation, often shortened to RAG or grounding. The model does not have to answer only from information stored in its parameters. Google's Search systems can retrieve current web pages from the Search index, review relevant information from those pages, and use that material to support the generated response. The interface can then show clickable links to pages that support the answer.
This changes the practical optimization question. A page does not need to be rewritten in a synthetic "LLM style." It needs to be eligible for retrieval and contain information that clearly satisfies part of the user's need. Unsupported claims, vague marketing language, and pages that never state the answer are harder to use as evidence than precise, verifiable material.
Query fan-out
Complex requests often contain several hidden questions. Google says its model can generate concurrent related searches through query fan-out. A request about repairing a weed-filled lawn, for example, can produce related searches about herbicides, non-chemical removal, and prevention. Google retrieves material for those related needs and uses the results to construct a broader response.
Query fan-out does not justify creating a thin page for every possible subquery. Google explicitly warns against producing large numbers of pages primarily to manipulate rankings or generative responses. A better approach is to understand the decision the user is trying to make, cover the important subproblems with sufficient depth, and use a site structure that lets related pages support one another when separate pages are genuinely warranted.
The existing Search index
To appear in Google's generative Search features, a page first needs to satisfy the ordinary technical conditions for Google Search. Google's minimum technical requirements include allowing Googlebot to access the page, returning a successful HTTP status, and providing indexable content. Meeting those requirements makes a page eligible; it does not guarantee crawling, indexing, retrieval, or display.
The generative AI guide adds that a page must be indexed and eligible to appear in Google Search with a snippet. This is why crawl controls, canonicalization, rendering, duplicate URLs, noindex directives, and basic content accessibility still matter. A file created exclusively for AI systems cannot compensate for a page that Google cannot reliably access or index.
Existing ranking and quality systems
Google says its generative Search products are rooted in core Search ranking and quality systems. Retrieval therefore is not a simple extraction contest in which the page with the most headings, schema fields, quotations, or exact-match phrases wins. Google's systems assess relevance and quality before a passage can support a generated response.
Site owners should avoid converting that statement into a secret list of "AI ranking factors." Google does not publish a formula for citation selection, and no outside agency has access to its internal ranking or AI systems. The defensible conclusion is narrower: strong technical SEO and useful content remain prerequisites, while purported AI-only shortcuts should be treated as unproven unless platform documentation or controlled evidence supports them.
Why Google’s Guidance Does Not Apply Equally to Every AI Platform
Google Search versus standalone answer engines
Google's guide describes Google Search. It does not document how every assistant trains models, builds indexes, invokes web search, selects sources, or generates citations. Some products combine a model with a live search system. Others may answer from model knowledge, retrieve from a proprietary index, use third-party search providers, or change behavior according to the product mode and user request.
The operational controls are also different. OpenAI's publisher guidance tells site owners not to block OAI-SearchBot when they want public pages to be discoverable and cited in ChatGPT search. Perplexity documents PerplexityBot as a crawler used to surface and link websites in Perplexity search results, separate from model-training controls. Google relies on Googlebot and its Search index. One crawler rule cannot represent all three systems.
Where structured reference files may still have operational value
Google explicitly says an llms.txt file is acceptable when another service uses it. A reference file can also be useful for developer documentation, internal agents, retrieval pipelines, or tools that intentionally consume the format. That is an integration benefit, not proof of higher visibility in Google Search.
The same distinction applies to Markdown mirrors, APIs, feeds, product data, and structured exports. A system may use them because it has been designed to do so. Their value should be tied to a documented consumer, implementation requirement, or observed access pattern not a generic claim that "AI prefers machine-readable content."
Why platform-specific testing matters
Testing should start with a defined platform and outcome. For Google, the outcome may be impressions in generative Search features, reported pages, or downstream Search traffic. For ChatGPT or Perplexity, the outcome may be source inclusion, referral sessions, or correct brand representation. These are different measurements.
A credible test records the exact prompt set, market, language, account state, date range, page set, and platform configuration. It repeats observations because generated responses can change. It also distinguishes between appearing as a cited source, being named in the answer, being recommended, and receiving a visit. Those events may overlap, but they are not interchangeable.
What cannot currently be verified
No external agency can verify a universal "AI rank" across every prompt and platform. No third-party dashboard has access to Google's internal citation-selection logic. A source appearance observed once does not establish a stable position. A referral visit does not reveal every unseen answer in which the brand appeared. A crawler request does not prove that the content was used in a response.
Claims should therefore carry an evidence label: directly observed, reported by the platform, measured in a defined study, or inferred. This prevents a useful experiment from being promoted into a universal rule.
The Four Defensible Foundations of AI Search Visibility
Google's documentation reduces the problem to familiar foundations. These do not guarantee inclusion in a generated answer. They make a site technically eligible and give retrieval systems stronger material to work with.
Original research and non-commodity information
Google tells site owners to create valuable, unique, non-commodity content. Its broader people-first content guidance asks whether a page provides original information, reporting, research, or analysis and whether it demonstrates first-hand expertise.
Non-commodity information does not have to be a large proprietary dataset. It can be a documented process, a carefully measured case, original screenshots, expert commentary, an implementation failure with a verified cause, or a comparison based on direct use. The defining feature is that the page contributes evidence or analysis that cannot be reproduced by paraphrasing the current top results.
Originality alone is not enough. The material still needs to answer a real question, make its assumptions visible, and separate observation from inference. A weak anecdote does not become authoritative because it is unique.
Clear entities and consistent brand signals
A site should identify its organization, services, authors, products, and locations consistently. The goal is not to trigger a secret entity score. It is to reduce ambiguity for users, search systems, and other services interpreting the business.
Use stable organization names, accurate author information, a clear About page, consistent business details, and canonical pages for important services or products. Supported Organization structured data can provide explicit information about a business, but it must match visible page content and follow Google's structured-data policies. Correct markup can support understanding and search appearances; it does not guarantee a generative citation.
Third-party descriptions should not be forced into artificial uniformity. The more defensible objective is factual consistency: the company name, category, ownership, location, products, and evidence should not conflict across sources.
Technical crawlability and browser-agent accessibility
For Google Search, start with Googlebot access, successful responses, indexable content, and correct rendering. Google can process JavaScript, but JavaScript-heavy sites introduce more failure points than straightforward server-rendered HTML. Important content should not depend on a broken script, blocked resource, or interaction that never exposes the text to the rendered page.
Other systems publish separate crawler controls. OpenAI and Perplexity both tell publishers to manage access through their documented user agents and robots.txt rules. A site may be open to Googlebot while blocking an answer engine's search crawler, or the reverse. Audit the actual server responses and logs instead of assuming every automated system sees the same page.
Google also notes that browser agents may inspect screenshots, the DOM, and the accessibility tree. Clear labels, usable forms, meaningful links, accessible controls, and visible state changes therefore support both people and automated agents. Agent accessibility is an emerging operational concern, not a proven citation-ranking factor.
Third-party corroboration from legitimate sources
Independent coverage can make claims easier to verify. A customer review, trade publication, public dataset, regulatory filing, conference presentation, or partner documentation may confirm information that a company also states on its own site. That is qualitatively different from purchasing low-quality mentions or placing coordinated comments to manufacture apparent popularity.
There is no public universal weighting system that converts mention counts into AI citations. Corroboration should be pursued because it improves evidence, reputation, and discovery, not because an agency promises a fixed number of mentions will produce a fixed ranking result.
How to Evaluate an AEO or GEO Agency
The label on the service matters less than the evidence behind it. A credible agency can define the platform, explain the technical mechanism it is addressing, show what it can measure, and state what remains uncertain.
Universal claims versus platform-specific strategies
Reject proposals that treat Google AI Mode, AI Overviews, ChatGPT, Claude, and Perplexity as one retrieval system. Ask which platform each recommendation targets and what official documentation or direct observation supports it. A recommendation to unblock OAI-SearchBot is specific to OpenAI search discovery. A recommendation to meet Google Search technical requirements is specific to Google Search. A recommendation to maintain llms.txt needs a named consumer that uses the file.
Measurement methodology
Ask for the measurement plan before approving the work. It should define:
- the exact platforms and product modes being tested;
- the prompt or query set, language, market, and device assumptions;
- the pages, brands, and competitors included;
- the number and timing of repeated observations;
- the difference between citations, mentions, recommendations, impressions, referral sessions, and conversions;
- the comparison period and any changes made during the test.
Google now provides a dedicated Generative AI performance report in Search Console to a subset of sites. The report includes impressions, appearing pages, countries, devices for Search, and time trends. It does not turn every impression into a click or conversion, and availability is still limited during rollout.
Citation tracking limitations
A citation is a source relationship, not a ranking position. It does not automatically mean the source was recommended, described positively, clicked, or responsible for the final answer. Agencies should show the surrounding answer and explain what the citation supported.
Tracking tools also observe only the prompts, accounts, regions, and times included in their collection. They cannot see every private answer generated for every user. Use them as sampling systems, not complete market censuses.
Warning signs in proposals and reports
Major warning signs include guaranteed AI rankings, claims of access to internal Google metrics, one tactic applied to every platform, and reports that do not disclose prompts or collection methods. Google itself warns that no third-party tool has access to its internal ranking or AI systems.
Also question proposals built mainly around llms.txt, AI-only schema, mass content chunking, manufactured mentions, or automatic rewrites. These may be useful in a specific implementation, but the agency must identify that implementation and prove the requirement. A list of fashionable deliverables is not a strategy.
Whether the agency understands retrieval rather than repeating terminology
Ask the agency to trace one real user need from query to eligible page. A serious answer should cover technical access, indexability, content relevance, evidence, competing pages, internal linking, and measurement. It should distinguish what is documented by the platform from what is inferred through testing.
An agency does not need a secret citation formula. It does need intellectual discipline. The strongest answer may be that a proposed tactic has not been proven and should not displace higher-confidence work.
What Uprankd Audits Instead
Uprankd's audit framework focuses on verifiable constraints and measurable opportunities rather than invented AI deliverables.
Brand and entity consistency
We check whether the organization, services, products, authors, and locations have clear canonical pages and consistent factual descriptions. We review visible business information, supported structured data, internal links, and conflicting third-party records. The objective is to remove ambiguity, not to claim a guaranteed entity-based ranking lift.
Retrieval accessibility
We verify that important URLs return successful responses, can be crawled by the intended systems, contain indexable content, and render critical information reliably. We review robots.txt rules, noindex directives, canonical tags, redirects, JavaScript dependencies, gated content, and interactive elements that may hide essential information.
For non-Google platforms, crawler checks are tied to documented user agents and the client's publishing policy. Allowing a crawler is a business decision as well as a technical one.
Original information and evidence gaps
We identify pages that repeat generic industry claims without proof and compare them with the evidence available to the business. Useful additions may include case data, methodology, original screenshots, expert commentary, limitations, worked examples, or first-party analysis. Claims that cannot be supported are removed or qualified.
Third-party corroboration
We review legitimate public sources that describe the business or verify important claims. We separate earned coverage and factual references from low-quality placements. The audit does not assume that every mention improves AI visibility; it records where external evidence strengthens trust and where conflicting information creates risk.
Query and intent coverage
We map the questions a page genuinely answers, the adjacent needs revealed by search behavior, and the existing pages competing for the same intent. The goal is not to publish a URL for every fan-out query. It is to decide when one authoritative page should be strengthened, when related pages deserve separate treatment, and where internal links should connect the journey.
Measurement using first-party data
For Google Search, Search Console is the primary source for impressions, clicks, click-through rate, average position, indexed pages, and,where available, generative AI visibility reports. Google Analytics can show what visitors do after arriving. External prompt tracking can supplement those sources, but it should not replace them or be presented as complete.
Practical Priority Checklist
What to fix first
Start with eligibility and access. Confirm that important pages return a successful status, are not unintentionally blocked, contain indexable content, and use coherent canonical signals. Fix rendering failures and pages whose central information is unavailable without fragile client-side interactions.
Then improve the material itself. Consolidate duplicate pages that compete for the same intent. Replace generic summaries with direct answers, original evidence, clear limitations, and useful examples. Add internal links where they help users and crawlers move between related topics.
Keep supported structured data accurate where it serves ordinary Search features. Do not remove useful markup simply because it is not a special AI citation lever.
What to test
Test important topics on the platforms that matter to the business. Record whether the brand is named, whether a page is cited, what claim the citation supports, which competitors appear, and whether the result sends measurable referral traffic. Repeat the test with controlled wording and a defined schedule.
For Google, compare generative Search impressions and appearing pages where the dedicated Search Console report is available. Connect that visibility to on-site engagement and conversions rather than treating impressions as revenue.
What to ignore
For Google Search, ignore claims that llms.txt, AI-only Markdown copies, artificial content chunking, special AI schema, manufactured mentions, or stylistic rewriting are mandatory visibility requirements. Do not ignore ordinary technical SEO, accurate structured data, useful formatting, or legitimate public relations merely because those activities are sometimes packaged badly.
What remains unproven
There is no verified universal AI ranking system, no stable cross-platform citation score, and no public formula that converts a specific number of mentions, schema fields, content blocks, or reference files into visibility. Evidence from one study, platform, site, or time window should not be generalized beyond its design.
The practical conclusion is narrower than the headline. Google did not kill SEO, and it did not prove that every GEO activity is worthless. It rejected several invented requirements for its own Search products. The businesses most likely to benefit are those that stop buying shortcuts and redirect the effort toward technical accessibility, distinctive information, clear evidence, and measurement that can survive scrutiny.
