Grounding and AI SEO: How AI Answers Get Anchored to Facts

Since February 2026, my Bing AI Performance Dashboard gives me first-party data on what AI systems consider worth citing. My glossary entry for “tl;dr” received 107 citations as a top grounding query over 30 days – over 1,300 total citations in 30 days, averaging 13 cited pages per day. One nuance matters here, which Microsoft clarified when launching the dashboard on February 10, 2026: Grounding Queries are not the questions users actually type. They are the internal search phrases Bing’s AI system generates to retrieve content for a response. The dashboard shows a sample of those – and it’s revealing enough.

On May 6, 2026, the Microsoft Bing team followed up on why this requires a different logic. Title of the engineering blog post: “Evolving role of the index: From ranking pages to supporting answers” – written by Krishna Madhavan, Knut Risvik, and Meenaz Merchant from Microsoft AI. Core message: classical search aims at which page a user should visit. Grounding aims at which information an AI system can responsibly use to construct an answer. And the authors add something that hasn’t received enough attention in SEO discourse: when two indexed sources contradict each other, a grounding system can’t simply prioritise one – it must register the conflict, because a system that silently arbitrates between them will most likely assert the wrong thing.

That’s not, in my view, pure PR. In my projects at iGaming.com and with my clients here at SEO Kreativ, I’ve been observing this for months: the shift is, in my view, already visible. What this means technically, why every AI SEO strategy without this understanding spins its wheels, how Standard v1.6 is structured, and what my own implementation at seo-kreativ.de/grounding-page looks like – you’ll find everything here.

Language models like GPT, Claude, or Gemini are probability machines at their core. They predict the next token based on training data – they don’t know whether a statement is true, only whether it sounds plausible. SISTRIX puts it concisely in their glossary entry: an LLM without external anchors often produces very convincingly worded but factually incorrect answers. Grounding is the antidote.

In practice, there are currently three main variants:

For SEOs, grounded answers are the practically relevant area – that’s where classical optimisation levers apply. What a model generates purely from training knowledge is a black box. You can’t optimise for something the model doesn’t access at runtime.

A concept that still gets too little attention in SEO discourse: abstention. AI systems can refuse to answer when the evidence base is weak or contradictory. Refusal is preferred over a wrong answer. In practice, this means: a page that’s indexed but delivers few extractable, contradiction-free statements won’t get cited – the model simply gives no answer on that aspect. The bar isn’t just visibility, but extractability of clean facts.

The Grounding Page Standard v1.6 deliberately draws a sharp distinction from classical SEO: “Classical SEO optimises documents for keywords. AI SEO curates entities for stable, accurate mentions.” That sounds academic, but has massive consequences for day-to-day work. In my client projects, I observe this regularly – my assessment: those who keep optimising only keyword clusters tend to leave a growing share of their visibility on the table. I’ve written up the overarching model in my article on the interplay of SEO, AIO, GEO, and LLMO. Grounding is an important building block for GEO and LLMO to achieve their full effect.

AI Search: No Longer a Future Topic

For the German-speaking market, this is no longer an abstract future scenario. Google AI Overviews have been available in Germany since March 2025; AI Mode expanded to over 40 countries in autumn 2025, including markets across Europe; Google does not, however, publish a country-by-country list with dates. How widespread AI Overviews actually are can only be estimated: industry observers cite around 15 to 25 percent of searches (as of spring 2026) – reliable official figures from Google don’t exist. I’ve written up what distinguishes AI Mode and AI Overviews concretely in a separate post.

Grounding optimisation for the DACH market is not something that can be deferred. The systems are in use and can influence which sources are considered as a basis for their answers.

Why Classical SEO Remains

A quick note to anyone still shouting “SEO is dead”: it isn’t. Classical SEO generally remains the prerequisite for entry. According to seoClarity analysis (October 2025, US desktop data): around 94 percent of all AI Overviews cite at least one source from the organic top 20. Those who aren’t visible there have poor odds in AI answers. The nuance matters – only around 56 percent of individual citations come from the top 20. The remaining 44 percent come from outside. Visibility alone is no longer enough.

A Seer Interactive analysis (Tracy McDonald, Product Development Lead, published November 4, 2025) with 25.1 million impressions across 3,119 queries and 42 organisations found: a source cited in an AI Overview achieves around 35 percent higher organic CTR (click-through rate) for the same search query than a non-cited source. McDonald explicitly notes that causality is not provable – brands with higher trust and higher baseline CTR may simply get cited more often. An honest caveat I share. The logic still holds: getting cited is the goal.

In my article on the AI Content Trap, I described this effect in detail. Short version: when your Google visibility collapses – after a Core Update or through mass low-effort AI content – your mentions in ChatGPT, Perplexity, and Google AI Overviews typically collapse at the same time. Tomek Rudzki, GEO expert at Peec AI, named this pattern “Mount AI”: steep ascent through mass AI content production, plateau, then abrupt collapse.

As a working hypothesis – not a proven law – the grounding chain can be described like this:

1. Google demotes a domain in a Core Update.
2. ChatGPT – which according to The Information’s research is reported to access Google’s index partially via SerpAPI – possibly finds content from that domain less frequently.
3. AI Overviews draw fewer sources from the domain.
4. Perplexity, which also relies on indexed web sources, responds in sync.
5. The result: the visibility loss cascades across the entire AI ecosystem.

Then there’s the aspect the Microsoft authors highlight in the May blog post: if my domain contains two contradictory statements on a topic – for example because an older post repeats an outdated figure that a newer one corrects – the grounding system can interpret the conflict as a quality signal and cite neither source. Cross-source consistency is therefore not just a reputation issue. It’s a technical lever.

In my technical SEO audit work, comparing Google visibility, AI citations, and Bing AI Performance data has become a standard diagnostic step. Those who only track Google rankings see the problem too late.

Hanns Kronenberg is an SEO expert, founder of GPT Insights, and Head of SEO at Chefkoch. He works with the distinction “On-Model SEO” and “Off-Model SEO” – a differentiation I now use in my own audits because it more cleanly separates what many conflate. On-Model SEO refers to a brand’s representation in an LLM’s internal model knowledge (what it “knows” without a live search). Off-Model SEO refers to external referenceability – the grounding probability when the model must search at runtime. A Grounding Page acts primarily at the Off-Model level: it makes facts findable and citable the moment the model triggers the retrieval step.

Kronenberg consistently distinguishes between “grounding” as an AI technical term (the technical process of anchoring to external sources) and “Grounding Page” as an SEO strategy (a dedicated fact page that facilitates this process). Those who equate the two understand neither properly.

In the OMR Education Podcast (February 2026), Kronenberg describes that the click-through rate from source citations in chatbots and AI Overviews is often only around one percent. His conclusion, which I share: GEO is less a performance channel than a tool of modern brand management. Those who optimise for grounding are optimising for brand mention in the answer, not for clicks. A mindset shift that’s unfamiliar for many organisations.

The standard has a clear goal: reduce AI hallucinations by allowing brands and entities to provide their own facts in a form that models can reliably extract. By this logic, a Grounding Page is not a marketing page. It doesn’t sell. It doesn’t persuade. It defines.

The Three Core Elements of a Grounding Page

The standard names three mandatory elements that together form the factual anchor:

• Stable definition: A short, verifiable statement describing what the entity is.
• Clear delimitation: An explicit statement of what the entity is not – disambiguation from similar or identically named concepts.
• Consistent structure: Same format, same logic, same extractability across all definition blocks.

Added to this are quality principles radically different from classical content marketing: no adjectives, one fact per sentence, visible timestamps (Created, Updated, Verified).

What v1.6 Concretely Changed

Version 1.6 is the current stable version (as of May 2026). The changes from v1.5 affect only the content structure – routing, canonicals, and hreflang remain unchanged:

• Intro box removed: The separate note box at the top of the page is gone. Pages now start directly with the entity – following the pattern of reference works like Wikipedia.
• Reference-oriented structure: The more direct opening makes the page more machine-readable, because the entity name appears immediately in the main content rather than in a meta-context.
• New “Further Reading” section: External links are consolidated at the end of the page rather than scattered throughout. This makes it easier for crawlers to identify which links are deepening resources and which belong to the main content.
• Quieter tone: Editorial language for better verifiability – less assertive structure, more descriptive clarity.

If you’re already running a Grounding Page per v1.5: the changes are manageable. Remove the intro box, move external links to a consolidated section at the end. That’s an hour of work.

Proof of Concept: Three Weeks, Three Engines

The standard was evaluated under defined test conditions. According to documentation on groundingpage.com, a fresh domain (registered November 2025, virtually no backlinks) was set up per the standard. Within three weeks, the domain was referenced as a source in ChatGPT, Perplexity, and Google Gemini. No guarantee – but a documented data point under clearly described conditions.

I’ve set up my own Grounding Page for SEO Kreativ at seo-kreativ.de/grounding-page following exactly this standard. The page contains no marketing copy, no emotional appeals, no aspirational claims. It defines the entity SEO Kreativ in factual, verifiable language – with core data, operator information, consulting services, target audiences, topic areas, a delimitation section, and a FAQ section. What it deliberately does not do: convert. That’s what other pages are for.

Concretely, you can implement the structure in three steps:

1. The page (HTML): Create a dedicated URL. Use definition lists (<dl>) to encode facts. The visible text is the primary source for the model – not hidden metadata. Start directly with the entity, without an intro box (v1.6 requirement).
2. The data (JSON-LD): Provide an identical structured representation below the visible text. Schema.org types like Organization, Person, Service are the standard choice. JSON-LD is a reflection, not a substitute for clean body text.
3. The authority (footer link): Link the page sitewide from the footer. Permanent findability across crawl cycles is the goal – just as an imprint clarifies legal identity, the Grounding Page clarifies semantic identity.

In my practice, a dedicated page type was usually the cleaner option, because it avoids stakeholder conflicts between marketing intent and factual clarity. Those who repurpose an existing “About us” page typically end up with a compromise that neither converts nor is citable.

What that looks like in practice is shown by the Grounding Check for my own Grounding Page – transparently a self-check with the project’s tool, not an independent audit:

In my projects, I typically work through this list from top to bottom. Point 1 has the greatest leverage – without it, the others contribute little.

1. Set up a Grounding Page per Standard v1.6. Own URL, dedicated fact page, JSON-LD, footer link. Direct opening with the entity (no intro box), external links in a consolidated “Further Reading” section at the end. This is the semantic anchor that everything else depends on.
2. Entity consistency across all platforms. Name, address, description, founding year, services must match exactly on the website, LinkedIn, industry directories, and – where applicable – Wikipedia and Wikidata. Cross-source drift is hallucination fuel. And as the Bing blog post describes: according to Microsoft, such a conflict can lead to neither source being cited.
3. Deploy structured data consistently. Organization, Person, Service, FAQPage, Article, HowTo – wherever it fits. I’ve described how structured data interacts with AI Overviews in a separate post. Schema markup is the language with which you tell AI systems unambiguously what something is about – but it doesn’t replace clear body text.
4. Answer-first structure in every article. The core answer belongs in the first sentences, not behind an 800-word introduction. AI systems actively extract the first clearly formulated statements. What can’t be condensed into the first paragraph of a section is often still a thesis, not yet a clear fact.
5. Clean H-hierarchy and chunkability. H2 as thematic anchors, H3 as sub-aspects, short paragraphs. What can be chunked well can be cited well. A paragraph that mixes three different ideas is not a good chunk – it’s a poor grounding candidate.
6. Primary sources instead of secondary aggregates. Those who cite studies, data, or interviews first-hand become a primary source themselves. That’s the direct path to a citation in an AI Overview. Secondary citations reduce citeworthiness, because the model then prefers to go directly to the primary source.
7. Establish cross-channel monitoring. Track your 20-30 most important queries monthly in ChatGPT, Perplexity, and Google AI Mode. The Bing AI Performance Dashboard (public since February 10, 2026) provides the first first-party data on citations and grounding queries – though Microsoft explicitly notes that only a sample of actual citation activity is shown. Specialised tools like Peec AI, Rankscale, or Otterly AI complement the picture for Google-side citations.

Additionally: don’t forget llms.txt as a complementary approach. The specification allows AI crawlers to be provided with a structured overview of your content – not a grounding substitute, but an additional channel for On-Model SEO. Kronenberg’s terminology: you control what models take from your domain in the next training run.

Four misconceptions come up particularly often in my consulting conversations:

“A Grounding Page is just an optimised landing page.” No. A landing page wants to convert; a Grounding Page wants to define. Both have their purpose – but mixing them weakens both. A marketing page full of emotional appeals is typically a poor factual anchor for AI systems, because models find few extractable statements there.

“Grounding is just another name for schema markup.” No. Schema markup is a component, not a synonym. The visible text of the Grounding Page is the primary source for the model – the JSON-LD below serves as a structured reflection, not a substitute. A page that only has schema but remains vague in body text doesn’t work as a Grounding Page.

“Grounding replaces E-E-A-T.” Also no. E-E-A-T (Experience, Expertise, Authoritativeness, Trustworthiness) is Google’s quality evaluation framework – grounding is a mechanism that preferentially draws from high-quality, authoritative sources. Both work in the same direction, but they’re not identical. Those without demonstrable expertise won’t become a primary AI source even with a perfect Grounding Page. I’ve written the complete E-E-A-T guide separately.

“Grounding” and “Grounding Page” are the same thing. This is perhaps the most consequential misunderstanding. Grounding is the technical AI process – independent of anything you do. A Grounding Page is an SEO measure that facilitates this process. A model can anchor to other sources without any Grounding Page. The question is just whether that’s your page or a competitor’s.

Since Grounding Pages became a topic, two neighbours keep getting thrown into the same pot: the /ai-instructions/ page and the llms.txt file. The question I hear most often in consulting: “Do I now need all three?” Short answer: no. They don’t do the same thing.

The core of the difference is direction. A Grounding Page does not assert how an AI should portray you – it provides verifiable facts and leaves the conclusion to the model. An /ai-instructions/ page does the opposite: it directly prescribes to the model what it should say. The US company Nectiv runs an illustrative example at nectivdigital.com/ai-instructions – a page explicitly addressed with the note that it contains structured information for AI assistants like ChatGPT, Claude, or Perplexity, and that even includes a dedicated “INSTRUCTIONS FOR AI ASSISTANTS” block prescribing how the models should describe the company.

In my view, this very claim to control is the weakness of the /ai-instructions/ approach. A model that has learned to weight sources by reliability tends to treat a self-instruction with the same scepticism as an advertising text – when in doubt, as marketing, not as fact. A neutrally worded fact page whose statements hold up against external sources is the more robust choice. You can’t reliably instruct an AI; feeding it verifiable facts, you can.

And llms.txt? The file is a signpost, not an evidence supplier – it tells crawlers which content you consider relevant. A broadly proven visibility effect has not been established to current knowledge (more on this in my llms.txt guide). In Kronenberg’s terminology, it acts more at the On-Model level – on what models take in during the next training run – while the Grounding Page makes you citable at runtime on the Off-Model level. If you have to pick a starting point: of the three formats, the Grounding Page is currently the best documented.

I’ve argued a lot in this article for setting up a Grounding Page. In fairness, the other side belongs on the table just as much – otherwise it would be advertising, not assessment.

The position of Christian Kunz (seo-suedwest.de). In his article “Grounding Pages for GEO – does it really do anything?” (May 31, 2026), Kunz argues more cautiously. His central objection concerns a possible negative signal: a Grounding Page must stay indexable so AI systems can read it – while at the same time it generates impressions in classical search that rarely lead to clicks. Kunz writes that one could argue “that every impression for a Grounding Page in Google Search that is not followed by a click can be regarded as a negative signal for the page and the website as a whole” (translated from German). Second point: prioritisation. Grounding Pages, he says, can go on the list as “one of several to-dos”; there are currently things in the GEO space that are “more important or at least as important, for example that AI bots can technically access a website’s content at all” (translated). This is a reporting account of his position – my assessment follows separately from it.

What my own data shows (observation). On the CTR objection I can contribute only a single data point from my practice, not a robust pattern: my Grounding Page at /grounding-page barely ranks in classical search – it draws few impressions, precisely because hardly anyone searches for it. That’s exactly why I’ve so far seen no measurable negative signal on the rest of the website. On the AI side, my Bing AI Performance Dashboard shows over 1,300 citations for the entire domain over the same period – but that’s a different channel from the CTR signal Kunz is concerned with, and says nothing directly about it. That’s my observation on one domain, not proof – the data is thin, and Kunz’s concern remains plausible for index-heavy pages with many click-free impressions.

In my view (assessment). Both objections are valid, but they don’t change my recommendation – they change the order. First, on the CTR signal: a Grounding Page is by design a niche page with low search volume. Where hardly anyone searches, hardly any click-free impressions arise – the risk scales with the page’s visibility, and that’s small here. Anyone who still wants to play it safe keeps the page accessible to AI crawlers but deliberately keeps its classical visibility small. Second, on prioritisation: here I agree with Kunz explicitly. If AI bots can’t technically access your content, the Grounding Page is cosmetics. That’s exactly why, for me, technical discoverability is the precondition that comes before any Grounding Page – not a position on the lever list, but its entry condition. The Grounding Page remains my strongest active lever, but it only takes effect once AI bots can reach the content at all. It’s a worthwhile to-do – but one that comes after the fundamentals, not before.

One detail for transparency that belongs in the assessment: Hanns Kronenberg is the originator of the Grounding Page Standard and runs the project. The specification is project-based, openly documented, and publicly available – it’s a project initiated by Hanns Kronenberg, not the standard of an industry body. That context is worth knowing when you weigh up the urgency.

The shift that Krishna Madhavan, Knut Risvik, and Meenaz Merchant from Microsoft AI described in the Bing Engineering Blog in May 2026 is not, in my view, pure PR. Classical search evaluates documents. Grounding evaluates verifiable statements. That’s a different unit of evaluation – with different requirements for factual accuracy, freshness, attribution, and consistency. And with a consequence I see increasingly in my audits: contradictory statements on your own domain don’t get prioritised by grounding systems – they get avoided. The investment in content quality is therefore not an optional extra – it’s an important building block for grounding systems to be able to classify your domain as citable at all.

For your practice: build a Grounding Page per v1.6 standard, keep it current, link it from the footer, maintain cross-source consistency, monitor your citations alongside your rankings. The effort is manageable; done right, it can improve citability.

For those who want to go deeper into the strategic framing: SEO, AIO, GEO, and LLMO in interplay – that’s the overarching model. The concrete cascade effect is in the AI Content Trap. And what a Grounding Page looks like per standard: seo-kreativ.de/grounding-page.