Microsoft AI Search Guide That Google Couldn't Write.

In the first two weeks of May 2026, the two largest companies in search published guidance on AI that disagrees at the level of first principles. One post landed on every SEO feed on LinkedIn within hours. It was Google's. The other sat almost unread for two weeks, even though it came out first and came from the company that actually feeds the answers ChatGPT gives your customers.

On May 6, three senior engineers at Microsoft AI published a post on the Bing blog called "Evolving role of the index: From ranking pages to supporting answers." I think it is the most important piece of writing on AI search published this year, and almost nobody in our industry shared it. Nine days later, on May 15, Google followed with its AI Optimization Guide, whose entire thesis fits in one sentence the SEO press repeated everywhere. From Google's perspective, optimizing for generative AI search is optimizing for the search experience, and so it is still SEO. Forget AEO. Forget GEO. Forget chunking, structured data, llms.txt, and inauthentic mentions. Keep doing what you were already doing. We argued a few days ago that this framing leaves out most of the map.

The Microsoft post was written before Google's guide existed, yet it reads like a point-by-point rebuttal of it. That is what happens when one company is describing the technical reality of the system it operates and the other is positioning around a business it needs to protect. Microsoft is not a neutral party here, but it is the party running the retrieval layer behind ChatGPT, the engine sending the overwhelming majority of AI-referred traffic into the clients we work with. Both posts cannot be correct, because they start from incompatible premises. The useful question is which one matches how the machines actually behave, and the field data now answers it cleanly.

Microsoft said the quiet part out loud

The whole argument is in the subtitle of the Microsoft post, "Same Foundations. Different Optimization Problems." Same crawlers, same quality signals, same understanding of the web. What sits on top of that foundation is a different problem depending on whether you are ranking pages for a person to click or grounding facts for an AI to commit to in an answer.

This is the line Google's guide spent four thousand words working around, and Microsoft states it plainly. "A common misconception is that grounding replaces search. It does not. Grounding builds on the same foundational infrastructure but it adds a new optimization layer on top." Google never used the word grounding once, never named the new layer, and built its entire message on collapsing the very distinction Microsoft put in writing. To make the distinction impossible to miss, Microsoft laid the two systems out in a single table, and that table takes apart the position that AI search is just SEO.

Traditional search asks which pages a user should visit, and that question is solved by ranking. Grounding asks what information an AI can responsibly use to build an answer, and that one is solved by a discipline Google has no established practice for, which is exactly why its guide treats the second question as a flavor of the first. The unit of value moves with it. Ranking optimizes around the document, while grounding optimizes around groundable information, the discrete facts that can be cleanly extracted, attributed, and verified. A page can rank beautifully and still be useless for grounding if its facts are tangled up in prose that resists extraction, and a page can rank poorly yet get cited constantly if its facts are dense, sourced, and unambiguous.

The role of the person changes too. In ranking, a human scans the results, skips what does not fit, and corrects course in real time. In grounding, the person reads a single synthesized answer and has to go check the sources to know whether it is true, so the tolerance for error collapses. A slightly wrong ranking is recoverable, while a slightly wrong grounding is a confident false sentence delivered as fact. And the set of valid outcomes is different, because ranking returns options while grounding either answers when the evidence supports it or abstains when it does not. That word, abstain, never appears in Google's guide, yet Microsoft places it in the middle of the table as a first-class outcome, because the correct move for a grounded system with insufficient evidence is to say nothing rather than guess.

The index itself is being rebuilt

Microsoft's second table is the one that should worry the SEO industry most, because it spells out what the index has to measure differently once it is feeding answers instead of ranking pages.

Factual fidelity becomes a hard constraint. In ranking, an indexed representation that distorts a page slightly is a minor nuisance, but in grounding that same distortion becomes a wrong answer, and the culprit is the chunking and transformation step the index runs, the exact process SEOs were told to ignore. Source attribution stops being a nice-to-have and turns into a core signal, because not all indexed content carries equal evidentiary weight, and a grounded system needs to know which sources it can lean on for which kinds of claims. That is provenance, citation lineage, and entity authority operating together as a measurement layer that has no real equivalent on the ranking side.

Freshness picks up a new failure mode. A stale ranking is a degraded experience, while a stale fact inside a grounded answer is an actively misleading one, which is why Perplexity leans so hard on recency and cites recently updated content far more often than older pages, with roughly half of its citations drawn from the past year. That is not a quirk of one engine, it is the correct design for the problem Microsoft is describing. Coverage shifts from pages to facts, so the question is no longer whether a URL sits in the index but whether the specific facts your audience asks AI engines about are retrievable and groundable, and a web that is broadly indexed but stuffed with paraphrased commodity content makes a poor grounding substrate. Conflict detection becomes the index's own job, because when two sources disagree a ranking system can rank one above the other and let the reader arbitrate, while a grounding system has to register the contradiction itself or the model on top of it will pick one silently and assert it with full confidence.

None of these five dimensions appear in Google's guide. The guide is essentially a list of things you do not need to do, and Microsoft's post is a description of what is actually being built underneath you.

The argument stops being a reading of two blog posts the moment you look at the citation data. If ranking and grounding were the same problem, the pages that rank would be the pages that get cited. They are not, and the gap is widening fast. An Ahrefs study of 863,000 keywords and four million AI Overview URLs found that only 38% of the pages Google cites in its own AI Overviews still rank in the top ten for that query, down from 76% just seven months earlier. A separate Ahrefs analysis put the overlap even lower, with only 12% of AI-cited URLs ranking in Google's top ten for the original prompt, and independent research from 5W found the overlap between top rankings and AI-cited sources collapsing from around 70% to under 20%. Google's own AI surface is already drawing most of its evidence from outside the pages its ranking system rewards. That is the separation Microsoft described, showing up in production, inside Google's product.

Follow the incentives behind each post

Read the two posts back to back and the motive behind each is hard to miss. Google's incentive is to keep SEO anchored to Google, because its business depends on search optimization staying functionally identical to whatever works on google.com. The moment practitioners decide AI search is a different problem that demands different work, the gravitational pull of Google's documentation, tools, and surfaces weakens, so the AI Optimization Guide functions partly as a defensive document, a reminder that the playbook is still Google's playbook.

Microsoft's incentive runs the other way. Bing's organic search has been a distant second for a decade in a market where second place stopped mattering, but the Bing index powers ChatGPT, which now handles roughly 2.5 billion prompts a day and about 17% of all global digital queries, the first real dent in Google's near-monopoly in twenty years. Microsoft has every commercial reason to convince the industry that grounding is a separate, demanding discipline, because the more grounding is treated as its own field, the more relevant Microsoft becomes inside it. Both companies are talking their book, which is normal. The difference that matters is that Microsoft's commercial interest happens to line up with how the systems actually work, while Google's requires flattening that reality into "still SEO." When you have to choose between a marketing document from the incumbent who needs the old framing to survive and a technical paper from the challenger who profits by explaining how the new system really runs, the technical paper is the one to bring into a client conversation.

The practical work splits into two tracks

If Microsoft is right, and the citation data plus what sits in our clients' GA4 dashboards both say it is, the optimization work for the next two years runs on two tracks instead of one.

The ranking track is everything Google's guide describes, and it still matters. Strong technical SEO, helpful non-commodity content, clean indexability, top-ten organic positions, structured HTML, and real E-E-A-T are the foundation, and Microsoft says as much, because grounding builds on this and skipping it means nothing downstream works. The grounding track is the one Google's guide refuses to name, and it asks for different work on top of that foundation.

Start with information design at the passage level rather than the page level. A claim that an engine can lift more or less verbatim, with a source attached, is groundable, while the same claim buried in a six-hundred-word section that needs summarizing to extract is not, and most enterprise content fails that test today, including pages that rank in the top ten. The fix is concrete. Lead sections with the claim, keep one idea per paragraph, attach the supporting number or source to the sentence that makes the claim, and stop hiding your most citable facts inside narrative build-up.

Treat provenance as a discipline, not a design choice. Microsoft now calls provenance a core grounding signal, so named authors, named sources, dated statistics, and visible citation lineage become retrieval signals rather than aesthetic flourishes. An article with a credible byline, clear expertise markers, and inline citations can out-cite a stronger but unattributed page from a bigger brand, so put real authorship and real sourcing on anything you want an engine to trust.

Make freshness a fact-level commitment instead of a page-level one. The question is no longer when the page was last touched but when each specific claim on it was last verified, so date your facts and not just your articles, and revisit the high-value numbers your audience actually asks about on a cadence that matches how fast they change. The engines that ground aggressively, Perplexity chief among them, reward exactly this behavior.

Build contradiction discipline across everything you own and earn. If your site, your profiles, and your earned coverage assert different versions of the same fact, you now have a problem you did not have on the old SERP, because the next generation of grounding indexes registers contradictions explicitly rather than just ranking one version above another. Audit your owned and earned surfaces for conflicting claims about pricing, specifications, dates, and positioning, and reconcile them before an engine resolves the conflict against you.

Finally, take Bing seriously, because it is the gate to ChatGPT. Bing Webmaster Tools is not a backup tool, it is a primary discovery surface for the largest AI search engine, and a brand that treats Bing as a second-class deployment target is invisible to ChatGPT in a way Google's guide will never warn it about. None of these five moves contradicts good SEO, all of them go beyond it, and Microsoft has now published the technical justification for why that beyond layer is the work.

What we do with this internally

Search Agency has run the layered model for two years. We work the SEO foundation, we work the grounding layer on top of it, and we measure and optimize the two separately because the engines themselves keep them separate. A few days ago we said Google's guide was incomplete. Microsoft's post, published before Google's and ignored by most of the field, fills in the part Google left out, and together the two form the real map, with Google drawing the territory closest to itself and Microsoft drawing the rest, including the ground the two share, in technical language Google's guide could not have used without conceding the problem was bigger than its own document.

If you read both as a practitioner, the takeaway is not to pick a side. It is to notice that the side telling you nothing has changed is the side that needs you to believe it, while the side explaining the new mechanics is the one being rewarded for getting them right. Build for the world Microsoft described, because that is the world ChatGPT, Perplexity, Claude, and increasingly Gemini already operate inside. Anyone still telling you it is just SEO is reading one of the two documents, or selling you the framing that happens to benefit the company that wrote it.

See where your brand stands in AI answers today, benchmarked against your competitors, no pitch required. Or talk to us about building the grounding layer into your search program.