The one sentence that matters
The unit of value shifts from documents to groundable information — discrete, supportable facts with clear provenance.
For two decades, the unit of value in search was the page. You optimized a page; it ranked; a user clicked. Bing is now saying the unit has changed: AI systems don't want pages, they want facts they can responsibly attribute. The page is still a vessel — but the index is no longer rating it as one.
Two questions, same infrastructure
Bing draws the line cleanly:
- Traditional search asks: which pages should a user visit?
- Grounding asks: what information can an AI system responsibly use to construct an answer?
The crawlers are the same. The quality signals are the same. The optimization problem is not. Search tolerates imperfect ranking because the user can scan and self-correct. Grounding tolerates nothing — errors compound across reasoning steps, and the user only sees the synthesized answer.
The five dimensions Bing now measures differently
This is the part of the post that should change how every content team works. Bing lists five things the grounding index must measure differently from traditional ranking:
Factual fidelity
“Chunking/transformations must preserve meaning and claims used in the answer.” — Bing
When your page gets broken into retrievable chunks, those chunks need to still mean what the original page meant. A misquoted fact that survives chunking shows up as a confidently wrong AI answer.
What to do: Make sure primary content is in the server-rendered HTML (not gated behind JavaScript) and that key facts appear in self-contained sentences, not split across multiple paragraphs.
Source attribution & provenance
“Evidence needs clear provenance and varying evidentiary weight.” — Bing
Not every citation counts equally. The index has to know which source is more trustworthy — Wikipedia over Reddit over a random blog. If your page has no clear publisher, no author, no sameAs links to verifiable profiles, your evidentiary weight rounds to zero.
What to do: Publish Organization JSON-LD with sameAs links (LinkedIn, Crunchbase, Wikipedia). Add author bylines with Person schema. Cite high-trust sources (.gov, .edu, primary research) where you can.
Freshness
“Stale facts can directly produce wrong answers.” — Bing
In traditional search, a stale page just ranks lower. In grounding, a stale fact produces a misleading answer. Pricing, headcount, product details — anything time-sensitive — needs an explicit last-updated signal.
What to do: Expose dateModified in your Article JSON-LD, ship a visible "Last updated" line, and serve a sitemap.xml with accurate <lastmod>.
Coverage of high-value facts
“Must ensure facts and sources people ask about are actually retrievable and groundable.” — Bing
Coverage isn't just "do you have a page?" — it's "is the specific fact someone is asking about actually retrievable from your site?" If your pricing lives inside a tab, your spec lives inside a PDF, or your answer to a common question lives inside a video, it isn't groundable.
What to do: Publish /llms.txt with your canonical pages per topic. Break content into descriptive H2 sections that name the question they answer. Use FAQPage and HowTo schema for question-shaped queries.
Contradiction detection
“Must detect and represent conflict; silent arbitration risks confident wrong answers.” — Bing
Old search would pick one source and move on. A grounding index has to notice when sources disagree, because an AI that silently arbitrates between contradictory facts will commit to one of them with full confidence. Internal contradictions on your own site (canonical vs. og:url, two different Organization names, mismatched titles) put your pages at the front of that conflict queue.
What to do: One Organization JSON-LD per site, referenced by @id. Make canonical, og:url, and the rendered URL agree. Don't ship two different brand names or two different prices for the same product.
Abstention is now a valid outcome
This part doesn't get enough attention. In traditional search, the system always returns ten results. In grounding, Bing says explicitly:
Abstention is a valid outcome when support is missing, stale, or conflicting — it reflects a deliberate judgment about what the available evidence can justify.
For brand teams, this is a second failure mode that didn't exist before. You don't just lose to competitors. You can also lose to no answer at all — an AI assistant that refuses to recommend anyone in your category because the evidence isn't there. We've already seen this in regulated and niche B2B categories. Bing has now made it official.
Retrieval is a system, not a step
The post also reframes retrieval itself:
Grounding operates in loops. A system grounding an AI answer may need to ask follow-up questions, refine retrieval based on intermediate results, combine evidence from multiple sources, and re-evaluate when confidence is low.
This is the fan-out pattern made formal. When you ask ChatGPT or Perplexity a question, the model doesn't run one search — it runs many, refines, and merges. That's why Aiso's fan-out query work matters: ranking for the prompt you can see is necessary but not sufficient. You also need to be in the retrievable evidence for the hidden sub-queries.
What this means for content teams
The practical shift is small but real. The work that used to be “write a page that ranks” is now “publish facts an index can prove.” In concrete terms:
- Stop hiding facts behind JavaScript. If your pricing, your dates, your specs, or your headline claim isn't in the server-rendered HTML, it can't be grounded.
- Expose dateModified everywhere. A visible “Last updated” line plus dateModified in JSON-LD. No content team should ship without one.
- One Organization, one canonical, one truth. Internal contradictions are now an active signal, not a rounding error.
- Add sameAs links. AI needs to cross-check your identity. LinkedIn, Crunchbase, Wikipedia, GitHub — whichever profiles verify you exist as a real entity.
- Publish /llms.txt. It's the cheapest way to tell an agent “these are the canonical pages for these topics.” (See our own.)
- Serve Accept: text/markdown. Bing's factual-fidelity point is partly about chunking artifacts. Giving agents a clean markdown version of your page removes those artifacts.
How to check your own site
We built a free tool that runs all of the above against a URL and grades it across the five dimensions Bing names. It takes about ten seconds:
Run the Grounding Readiness Checker →
The bigger picture
Bing closes the post with the line that should be on every content team's wall:
Search optimizes for likelihood of relevance. Grounding must measure strength of evidence.
That's the whole game. The brands that win the next decade of AI answers are the ones whose facts are easier to prove than their competitors'. Provenance, freshness, coverage, fidelity, and internal consistency are no longer nice-to-haves — they are the index.