Google this month revealed new documentation on what it takes to look inside AI-powered search options, and website positioning advisor Dr. Marie Haynes walked via the doc in a YouTube video launched on Might 15, 2026. The information, aimed toward web site homeowners and publishers, addresses two technical mechanisms that now sit on the heart of how Google surfaces content material in AI Overviews and AI Mode: retrieval-augmented technology (RAG) and the question fan-outapproach. It additionally delivers pointed steerage on commodity content material, hyperlink constructing, structured information, and the rising world of agentic search – subjects that collectively reframe what it means to optimize for Google in 2026.

The video, posted to Haynes’ YouTube channel below the collection “Search Information You Can Use,” runs roughly 26 minutes and walks via the documentation part by part. Haynes has been practising website positioning since 2008 and is among the discipline’s extra carefully adopted voices on algorithm conduct and Google’s high quality programs.

What the documentation says about altering search conduct

In response to Google’s new information, as lined by Haynes, the corporate acknowledges that customers are “more and more gravitating to generative AI experiences.” The corporate frames this shift not purely as a discount in alternatives for publishers however as a possible opening for reaching higher-quality guests. In response to the documentation, “as we improve search to fulfill these altering expectations, this transformation gives new alternatives to achieve individuals who is perhaps extra inclined to have interaction together with your website, spend time together with your content material, and even convert to changing into a subscriber or making a purchase order.”

Whether or not that framing holds for many publishers is debatable. Visitors information throughout the trade has moved in a single course. AI Overviews now correlate with a 58% reduction in click-through rates for top-ranking pages, based on Ahrefs analysis from February 2026, and SISTRIX information from Germany locations the monthly click loss from AI Overviews at 265 million. Place-one CTR in that market collapsed from 27% to 11% when AI Overviews are current.

RAG: why web sites nonetheless matter

The primary technical mechanism the information covers is retrieval-augmented technology. RAG is the method by which Google’s AI options use the online – together with particular person web sites – to floor their solutions. The system doesn’t generate responses from coaching information alone. As an alternative, it retrieves data from listed pages to offer factually present, source-supported outputs. That is the technical purpose why web sites stay related even because the search outcomes web page more and more solutions questions instantly: with out crawlable, listed content material, the AI has nothing to floor its solutions towards.

Haynes frames this as a big shift in what the index is for. Historically, the query the index answered was easy – which pages ought to a consumer go to? Now, the index serves primarily as grounding material for AI answers. Haynes additionally famous that Microsoft’s Bing revealed a weblog submit across the similar time reaching an identical conclusion in regards to the altering position of the index.

Question fan-out: how Gemini breaks down a query

The second mechanism lined in Google’s information is the question fan-out. It is a approach by which a specialised mannequin – described by Haynes as distinct from the publicly out there variations of Gemini – generates a listing of concurrent sub-questions from a single consumer question. When somebody asks “the way to repair a garden filled with weeds,” the fan-out mannequin may produce parallel queries together with “greatest herbicides for lawns,” “take away weeds with out chemical substances,” and “the way to stop weeds in garden.” These run concurrently, and the AI synthesizes their outcomes right into a single response.

Haynes first heard in regards to the fan-out approach at Google IO final 12 months, the place she spoke with an engineer who labored on AI Mode. The documentation now makes the mechanism specific. Some practitioners have responded to information of the fan-out by creating particular person content material items focused at every probably sub-query. Haynes pushes again on this strategy. Creating massive quantities of content material to cowl estimated fan-out queries is more likely to produce what she and the Google documentation each describe as commodity content material – the class Google is explicitly transferring to reward much less.

Commodity content material: the central downside

The excellence between commodity and non-commodity content material runs via the information as its organizing precept. In response to the documentation, “seven ideas for first-time residence consumers” is the sort of basic data that exists all around the internet – content material that no particular consumer is looking for out from a particular supply. That’s commodity content material.

Non-commodity content material, because the information defines it, begins with a singular standpoint. In response to Google, “our AI programs check out a wide range of sources. So, it may be useful to have a singular viewpoint that stands out.” The second dimension is direct, first-hand expertise. In response to the documentation, “create the content material your self based mostly on what you already know in regards to the subject and take into account what in-depth expertise you may carry to your content material. Do not simply recycle what others on the web have already mentioned or may very well be simply produced by a generative AI mannequin.”

It is a direct extension of the E-E-A-T framework – Expertise, Experience, Authoritativeness, and Trustworthiness – that Google has been constructing into its high quality steerage since 2022, when the extra “E” for expertise was launched. The argument, as Haynes articulates it, is that have is the one dimension of content material high quality that AI can’t replicate. Including first-person framing to a web page is just not enough. What issues is the perception that comes from having truly executed one thing.

Haynes gives a sensible check from Danny Sullivan, Google’s Search Liaison, who raised the query on the Google Search Central occasion in Toronto: if a website’s content material had been faraway from the online totally, would anybody miss it? If not, it’s probably commodity content material. PPC Land has documented the broader consequences of this shift for publishers – websites that demonstrated measurable beneficial properties following Google’s June 2025 core replace shared traits centered round complete, useful content material supply somewhat than scaled manufacturing.

Scaled content material and spam insurance policies

The information revisits Google’s scaled content material abuse spam coverage, a subject Sullivan raised a number of instances on the Toronto occasion. Scaled content material – content material produced in massive portions utilizing automated or AI-assisted strategies with the aim of manipulating rankings – is explicitly towards Google’s tips. The documentation additionally notes that utilizing AI to help with content material creation is permissible, supplied the output meets Google’s high quality requirements and doesn’t violate spam insurance policies.

Each website in Google’s index carries an inside spam rating that isn’t seen to the positioning proprietor. In response to Haynes, websites with excessive spam scores face compounding disadvantages in rating. She notes that core updates, which Google constantly says aren’t about penalizing hyperlink constructing, nonetheless constantly hurt websites with hyperlink profiles constructed for website positioning functions. Her interpretation is just not that these websites are being penalized, however that the hyperlinks that when helped have gotten progressively much less helpful as Google’s high quality alerts develop extra refined.

Technical website positioning: much less central than it as soon as was

On technical website positioning, the information advocates for technically sound websites however doesn’t place major weight on efficiency metrics. In response to Haynes’ studying of the documentation, the important thing query is whether or not Google can retrieve and render the content material in any respect – crawlability, indexability, and rendering are the brink necessities. Core Net Vitals that rating within the medium vary aren’t, in her view, a major driver of rating for many websites. The exception is e-commerce: research have proven that each millisecond of web page pace enchancment corresponds to conversion charge beneficial properties, making efficiency optimization commercially justified in that class.

Enterprise brokers: what Service provider Heart now allows

One part of the information factors to the rising enterprise agent characteristic in Google Service provider Heart. Haynes examined this with one consumer and located a search consequence that displayed the model title alongside a button labeled “chat with the enterprise agent.” The knowledge surfaced within the agent was pulled instantly from the consumer’s website. Nothing in it was inaccurate. Haynes interprets this because the probably way forward for seek for most companies: the web site serves not because the vacation spot however because the grounding layer – correct data that AI assistants draw from when conversing with customers.

That is per the infrastructure Google has been constructing towards. Google launched the Universal Commerce Protocol on January 11, 2026, co-developed with Shopify, Etsy, Wayfair, Goal, and Walmart, establishing a technical commonplace for AI brokers to execute purchases throughout retail platforms. The protocol helps REST and JSON-RPC transport layers and is appropriate with Agent2Agent and Mannequin Context Protocol. Service provider Heart is the central hub for product information that feeds these experiences. Google expanded UCP checkout into main search results on May 5, 2026, pushing a Purchase button instantly into commonplace search outcomes for the primary time.

Fantasy-busting: LLMs.txt and content material chunking

The information addresses two practices which have generated important dialogue within the website positioning neighborhood and largely dismisses each. The primary is LLMs.txt – a file format supposed to present AI programs a curated, structured model of a website’s content material. In response to Google’s documentation, as relayed by Haynes, there is no such thing as a requirement to provide LLMs.txt recordsdata for a website to look in search. Language fashions are able to studying commonplace HTML and figuring out whether or not textual content will probably be helpful to customers. Google’s personal resolution to publish markdown variations of its developer documentation was, in Haynes’ studying, for builders who wished to feed these paperwork to their very own AI brokers – not a sign that website homeowners ought to comply with go well with.

Shopify has carried out LLMs.txt throughout its service provider websites. Haynes’ interpretation is that this serves the Shopify agent’s means to work together with and modify these websites – a unique use case from search visibility. The conclusion: until there’s a particular useful purpose to provide LLMs.txt, it’s not a precedence.

The second follow the documentation dismisses is content material chunking – the deliberate breaking of content material into small, discrete models to make it simpler for language fashions to extract particular solutions. Haynes notes that this follow made some intuitive sense within the early days of AI-assisted search, when having particular person sentences that exactly answered a question appeared to assist. Present language fashions – Gemini, ChatGPT, Claude – are refined sufficient to extract related passages from naturally flowing prose. Extremely chunked content material may very well work towards the individuality and experiential qualities that Google now rewards.

A piece of the information covers what Google calls “inauthentic mentions.” In response to the documentation, “similar to the remainder of Google search, our generative AI options can present what’s being mentioned about services throughout the online, together with in blogs, movies, and discussion board discussions. Nevertheless, looking for inauthentic mentions throughout the online is not as useful because it might sound.”

Haynes interprets this as a reference to hyperlink constructing and the broader follow of making content material – lists, round-ups, weblog posts – that locations a model in a good context with out real third-party endorsement. The documentation seems to increase that concern from conventional search into AI characteristic inclusion.

Structured information: helpful in slender instances

On structured information and schema markup, the information’s place is extra nuanced. Google deprecated FAQ-rich outcomes a while in the past; structured information centered on FAQ schema is due to this fact of restricted worth for many websites. Haynes notes that Ahrefs revealed a research testing schema implementation throughout a lot of pages already showing in AI Overviews and located no measurable rating distinction. Her personal follow: schema receives comparatively little consideration for informational content material. For e-commerce, the calculus is completely different. AI brokers require correct, structured product information – costs, availability, specs – to function successfully. Service provider Heart product information serves this perform, and schema that exposes steadily altering data to AI brokers has real utility.

Agentic experiences and WebMCP

The part that generated probably the most consideration in Haynes’ commentary covers agentic experiences – particularly WebMCP, a protocol that permits brokers to make use of instruments and work together with the useful layer of a web site, not simply its content material. On the Google Search Central occasion in Toronto, Haynes requested how a lot consideration practitioners ought to pay to WebMCP. The reply from Google workers was successfully: do not. Their acknowledged reasoning was to deal with what Google has already communicated – primarily UCP and the enterprise agent.

Haynes is skeptical of that reply and has causes for it. She notes {that a} consumer carried out WebMCP on their website following discussions about its potential. The implementation labored, however the consumer expertise required giving an exterior agent specific directions together with API tokens earlier than it may work together with the positioning – a friction degree that makes the present model troublesome to deploy virtually. Six days after the Toronto occasion, Google revealed a weblog submit about creating agent-friendly websites. That submit describes 3 ways brokers understand a website: screenshots of the interface, the HTML supply, and the accessibility tree.

The accessibility tree is a structural illustration of a web page that exposes what’s interactive – buttons, kind fields, clickable parts – in a format optimized for automated brokers. It’s distinct from the visible design and from the uncooked HTML. Haynes constructed a software utilizing this framework that audits a website’s accessibility tree and returns each a diagnostic report and a immediate that may be given to a language mannequin for enchancment planning.

Google IO, referenced within the video as an upcoming occasion on the time of recording, was anticipated to handle modifications to go looking and the motion towards agentic search. A number of agentic capabilities are already dwell, based on Haynes: some brokers can e book companies or examine merchandise, and a voice agent can contact native companies on a consumer’s behalf to assemble quotes earlier than assembling a abstract doc.

Dos and don’ts: what the information means in follow

The documentation, taken as an entire, attracts a reasonably clear line between practices Google’s programs reward and practices which are both impartial or actively counterproductive. The next is drawn instantly from the information as analyzed by Haynes.

Do: produce content material grounded in direct expertise. The information is specific that content material written from real first-hand information has qualities that purely informational writing lacks. Publishing an account of attending an trade occasion, testing a product, or working via a particular downside with a consumer is the sort of materials Google’s high quality programs are designed to floor. The query Haynes suggests asking: does this content material include one thing that would not have been produced by somebody who had not truly executed this?

Do: guarantee Google can crawl, render, and index the positioning. Technical website positioning is just not irrelevant – it’s foundational. If Googlebot can’t entry the content material, retrieve it, and render it accurately, nothing else issues. Crawl errors, noindex directives utilized incorrectly, and JavaScript rendering failures are price fixing. Past that threshold, nonetheless, the information doesn’t prioritize efficiency optimization for many websites. Core Net Vitals within the medium vary aren’t, within the documentation’s framing, a major rating driver besides in e-commerce, the place web page pace has a demonstrated relationship with conversion charges.

Do: add high-quality photos and video. The documentation particularly recommends these codecs. Haynes notes that Google has begun putting short-form video content material on the house display of Google TVs, and that video is a format AI can’t simply replicate. A walkthrough, demonstration, or commentary delivered on digicam carries experiential weight that textual content summaries of the identical content material usually don’t.

Do: use structured information for e-commerce and quickly altering data. Product schema that exposes correct pricing, availability, and specs is efficacious as a result of AI brokers require present, structured information to perform. Service provider Heart product information feeds instantly into UCP-powered experiences. For information-focused websites, schema provides restricted worth and the documentation doesn’t emphasize it.

Do: make the positioning accessible to brokers. Google’s weblog submit on agent-friendly websites – revealed six days after the Toronto occasion – identifies three layers via which brokers understand a web page: screenshots, HTML supply, and the accessibility tree. Buttons which are too small, interactive parts that aren’t labeled, and useful parts that don’t expose accurately within the accessibility tree are price auditing. This isn’t a rating issue within the typical sense, but it surely issues as agentic search options mature.

Do not: produce content material at scale to focus on question fan-out sub-queries. The information doesn’t prohibit creating a number of pages on associated subjects, however Haynes is direct in regards to the threat. If the aim of manufacturing these pages is to cowl each probably sub-query somewhat than to carry real depth or a singular angle to every one, the result’s nearly actually commodity content material. Google’s spam insurance policies on scaled content material apply no matter whether or not the content material is AI-generated or human-written.

Do not: construct or search inauthentic mentions. The documentation addresses this instantly within the context of AI options in addition to conventional search. Lists, round-ups, and hyperlink placements engineered to make a model seem really helpful when it’s not carry a spam threat. Each website in Google’s index has an inside spam rating that isn’t seen externally, and practices that inflate that rating create compounding disadvantages throughout each conventional outcomes and AI characteristic inclusion.

Do not: depend on LLMs.txt for search visibility. There is no such thing as a proof that producing an LLMs.txt file improves a website’s look in Google’s AI options. The format could have useful utility for different functions – Shopify’s implementation seems to serve its personal agent infrastructure, not search. Until there’s a particular, non-search purpose to provide one, the documentation doesn’t help the funding.

Do not: chunk content material into artificially small models. Writing content material in discrete, answer-sized fragments to make extraction simpler for language fashions works towards the qualities Google now rewards. Pure, flowing prose that incorporates real evaluation is more durable to chunk however simpler for classy language fashions to know – and extra more likely to reveal the non-commodity traits the information is designed to reward.

Do not: overinvest in schema for informational content material. Ahrefs’ research of pages already showing in AI Overviews discovered no measurable distinction when schema was added. FAQ schema, particularly, misplaced its major sensible utility when Google deprecated FAQ-rich outcomes. The time spent implementing schema on informational pages is, most often, higher directed elsewhere.

Do not: disavow hyperlinks anticipating a restoration. Haynes is cautious right here: she doesn’t assume disavowing low-quality hyperlinks restores no matter profit these hyperlinks could as soon as have supplied. The worth of manipulative hyperlinks tends to decay somewhat than invert. Websites with hyperlink profiles constructed for website positioning functions aren’t essentially being penalized within the typical sense, however the alerts these hyperlinks present carry diminishing weight. The sensible implication is to cease constructing hyperlinks for website positioning functions somewhat than to aim to reverse what has already been executed.

What this implies for entrepreneurs

The importance of this documentation for the advertising neighborhood lies in what it formally confirms somewhat than introduces. PPC Land’s coverage of Google’s new AI search guide famous that Google revealed the information on Might 15, 2026, with three particular areas of focus: non-commodity content material, sensible optimization for native, purchasing, picture, and video codecs, and direct correction of widespread misinformation about AI search practices. The information is notable for what it pushes again on: a good portion of the website positioning trade’s response to AI search has concerned techniques – LLMs.txt, content material chunking, FAQ schema, fan-out concentrating on – that the documentation now explicitly charges as pointless or counterproductive.

Research from Cyrus Shepard published on May 7, 2026, masking 54 experiments, patents, and case research, discovered that URL accessibility and search rank rating 9.5 and 9.4 respectively on an evidence-weighted scale of AI quotation elements – putting conventional website positioning fundamentals on the high of the rating. The discovering that 38% of AI Overview citations come from top-10 natural outcomes, based on Ahrefs information cited in that evaluation, implies that the trail to AI quotation visibility runs primarily via typical high quality alerts, not AI-specific techniques.

The information doesn’t resolve the basic rigidity going through publishers: natural clicks are declining as AI Overviews serve users answers without requiring a site visit, whereas the identical AI programs more and more depend upon these publishers’ content material as grounding materials. What it does make clear is Google’s acknowledged view of what content material is price grounding towards – and commodity content material, at scale, is just not it.

Timeline

Abstract

Who: Dr. Marie Haynes, website positioning advisor and researcher practising since 2008, revealed a video evaluation of latest Google documentation on AI search rating. The documentation itself is authored by Google and addresses web site homeowners, publishers, and website positioning practitioners.

What: Google launched new steerage explaining the 2 major technical mechanisms that decide visibility in AI search options – retrieval-augmented technology and the question fan-out approach – alongside steerage on non-commodity content material, structured information, LLMs.txt, content material chunking, inauthentic mentions, and the transition to agentic search experiences. Haynes walked via the doc publicly, including context from consumer work and attendance at Google occasions.

When: The YouTube video was revealed on Might 15, 2026. The Google documentation it analyzes was launched the identical day.

The place: The video was revealed on YouTube below Haynes’ “Search Information You Can Use” collection. The Google documentation is a part of the corporate’s developer and search assist assets.

Why: The documentation is critical as a result of it formally addresses – and in a number of instances dismisses – practices which have proliferated within the website positioning trade as direct responses to AI search options. It confirms that conventional website positioning fundamentals stay the first path to AI quotation, whereas clarifying that the index now features primarily as grounding materials for AI solutions somewhat than a listing of pages for customers to go to. For advertisers and publishers, the implications lengthen into agentic commerce infrastructure and the altering position of internet sites as grounding layers for AI assistants somewhat than direct site visitors locations.


Share this text


The hyperlink has been copied!




Source link