Google says you don’t need to optimize for AI. We agree — about three of the four things they listed.

Yesterday, Google’s Search Central team quietly updated a document that a lot of people in this corner of the industry have been waiting for. The page is called Guide to Optimizing for Generative AI Features on Google Search, and the punchline is roughly: there is no GEO. There is just SEO, and it still works.

I want to take that seriously, because Google has been wrong about its own systems before (the 2024 API leak is the largest cache of receipts on that), and parts of the practitioner community are already in "vindicated" mode. But the leak should also have taught us the opposite reflex: when Google describes its own pipeline, the description is technically careful, the omissions are often the interesting part, and the gap between "what they said" and "what an operator should do tomorrow" is where the work lives.

So I want to walk through Google’s new doc the way I walk through an audit report. What did they say. Where does it match what I see surfacing brands. Where does it leave a gap that an operator needs to close themselves.

What the doc actually says

The relevant passages are short and worth quoting in full, because every secondhand summary you’re going to read this week will reshape them. From the June 15, 2026 update:

“Our generative AI features on Google Search are rooted in our core Search ranking and quality systems.”

— Google Search Central, June 15, 2026

That’s the architectural claim. The optimization claim is in the “Mythbusting generative AI search: what you don’t need to do” section, and it’s four bullets:

1. No machine-readable files: “You don’t need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search (including its generative AI capabilities), as Google Search itself doesn’t use them.”
2. No content chunking: “There’s no requirement to break your content into tiny pieces for AI to better understand it.”
3. No AI-specific writing: “You don’t need to write in a specific way just for generative AI search.”
4. No special schema: “Structured data isn’t required for generative AI search, and there’s no special schema.org markup you need to add.”

The doc also names two concepts Google has spent the last year avoiding in official text. Query fan-out — “a set of concurrent, related queries generated by the model to request more information and fetch additional relevant search results” — and RAG, which Google defines as “A technique (also known as grounding) used to improve quality, accuracy, and freshness of AI responses by relying on our core Search ranking systems.”

The two technical admissions and the four mythbusters are the whole take. Everything else in the doc is the standard E-E-A-T, helpful-content, technical-SEO checklist that has been the documented Google position for a decade.

Where Google is right

Let’s start with the part of the doc I think operators should accept, because it matches what audits show.

The inputs really are the same

The page that Gemini grounds onto for a buying-intent query in a vertical I audit weekly is, almost always, a page that also ranks well in classical Google Search for the head term. The trust signals are the same. The entity model is the same. If your site isn’t indexed, isn’t fast, doesn’t have a clean Organization schema, isn’t cited by trusted third-party platforms in your category, you don’t get pulled into the AI surface for Google’s engines. There’s no separate index, no separate trust score, no secret AI-only signal. The leak hinted at this and the new doc just makes it explicit.

The cleanest evidence I have for this is something I find embarrassing to admit. When we audited our own site in May, our site-readiness grade came back at B. Schema, semantic HTML, freshness, quotability — all in decent shape. Our actual GEO score came back at 17 out of 100, an F. We were recommended in 8% of buyer-intent queries and ranked 3rd of 103 brands AI engines surface for “AI search visibility audit.” Site quality is necessary. Site quality alone moves nothing. Google’s “same inputs” framing is true at the foundations layer and silent on what happens above it.

Practically: if your audit reveals broken canonicals, a JS-rendered SPA with no SSR, an Organization schema that’s missing the sameAs links to your Wikipedia and LinkedIn entities, or a domain whose entire content portfolio drifts off-topic, you don’t have an “AI problem.” You have a foundations problem, and fixing the foundations is what Google is correctly pointing at.

llms.txt is not the magic file some people sold it as

Google’s position on llms.txt is now on the record: they don’t use it. I’ve been agnostic about this for a year. I now have to be agnostic with more conviction. There are zero verified cases where adding llms.txt to a site changed citation behavior in a way you could attribute to the file rather than to the content the file pointed to.

I still recommend keeping llms.txt on most client sites. Two reasons. First, the cost is approximately zero — it’s a static text file. Second, the convention is novel enough that a major non-Google engine could adopt it tomorrow and the difference between “ready” and “not ready” would matter. But I should be clearer than I have been: llms.txt is a low-probability hedge, not a lever. If you don’t have one, putting one up is a fifteen-minute task. If you do have one, do not credit it for anything.

You don’t need AI-specific schema

Schema.org has not added new types for AI extraction. Google does not have a hidden generativeAIVisible: true flag. The same JSON-LD that gets you a rich result for a recipe, a product, a how-to, a Q&A, or an Organization is the same JSON-LD that helps the AI surface understand your content. The work is to use the existing types correctly — the linked @graph, the entity references, the consistent sameAs array — not to invent new ones.

If you have been sold a $5,000 engagement to add “AI-readiness schema,” you have been sold the same schema you should have already had.

Where the framing leaves a gap

So far so good for Google. Now the part that matters for anyone making decisions about where to spend the next two quarters of content budget.

1. The unit of selection has changed, even if the inputs haven’t

Google’s claim is that you don’t need to chunk your content. Their claim is technically true: the chunking happens in the retrieval system, not in your CMS. The implication people are drawing — that the structure of your prose doesn’t affect what gets cited — is not.

Here’s what an audit actually measures. We capture the AI response. We capture the source URL. Then we read the source URL and find the passage that was lifted. In something like 70–80% of citations across the four engines we audit, the lifted passage is one of the first three paragraphs on the page, or a self-contained chunk inside a list, a definition, or a Q&A block. It is almost never the conclusion. It is rarely the long expository middle.

The window cleaning vertical gave me the clearest single example of this. We ran four audits against one regional market between April and May. AI surfaced 101 distinct window-cleaning brands across 103 captures. When we read the lifted passages, the pattern was uniform: the operator’s meta description was quoted, or the H1, or the opening sentence on their services page. One operator with a strong reputation locally got cited less than a smaller competitor because their homepage opened with a brand story (“Family-owned since 1998…”) instead of a service description. The competitor opened with “Residential and commercial window cleaning in [City] — insured, bonded, same-week scheduling.” That sentence got lifted into AI responses verbatim, in three of the four engines. The brand-story site got skipped.

2. “Don’t write for AI” is true. “Don’t write differently” isn’t

Google’s mythbuster on writing style is aimed, I think, at the strain of GEO advice that tells you to write in robotic question-and-answer format with no voice. They’re right to push back. Hollow Q&A pages don’t get cited.

What does get cited, repeatedly, is content with first-person experience markers. “In our audits we measured.” “We tested.” “Our research found.” The Princeton GEO paper isolated this as the single largest signal separating cited content from filtered AI slop, and our own site-readiness checks weight it accordingly. It maps almost one-to-one to Google’s own “Experience” in E-E-A-T — which Google does still emphasize. The mythbuster says don’t write “just for” AI. It doesn’t say don’t write with your hand on the work. The second reading is the one that pays.

The example that turned me around on this came, again, from the audit of our own site. We configured BeCited with six differentiators we’d been pitching in sales calls. The most distinctive was “Founder-led: every audit is done by Owen Kurth personally, not a junior analyst or an LLM.” A real claim. A defensible claim. A claim I had spent weeks polishing. The rubric measured whether that claim surfaced anywhere in AI responses to buyer-intent queries. It surfaced in zero of fifteen captures where it should have. We went back to the site to see why. The claim was on the homepage, but it lived in the third section, inside a longer paragraph about methodology, qualified by another sentence about why that matters. There was no first-paragraph version of it. There was no quotable, self-contained sentence anywhere on the site that said “Owen runs every audit personally.” The retrieval models couldn’t extract it because we had never written it in extractable form. Six of six differentiators came back the same way.

Google’s mythbuster says no special writing. That’s right. What we needed wasn’t special writing — it was the same writing, structured so a passage retriever could find it. Sentence one, on the right page, says the thing. That fix was on the priority list our own tool generated.

3. Google describes Google. The AI surface is bigger than Google

This is the omission that bothers me most about the doc, because it’s the one that will most distort operator behavior if read uncritically.

Our standard audit captures answers across four engines: ChatGPT, Gemini, Perplexity, and Claude. Of those four, Gemini is the one that uses Google’s ranking pipeline directly. Google AI Overviews and AI Mode also draw from that pipeline. ChatGPT uses Bing as one of several inputs and increasingly its own crawl. Perplexity uses a hybrid of its own index plus selected providers. Claude uses a search tool with its own source-selection criteria. None of them read Google’s docs. None of them treat Google’s mythbusters as guidance.

The window cleaning audit showed me what this looks like in practice. The three platforms most cited across all four engines in that vertical were Angi (48% of captures), Yelp (32%), and Thumbtack (23%). None of those are part of “optimizing for Google.” A window cleaning operator who follows Google’s new doc to the letter — clean schema, fast site, helpful content, no fancy AI tricks — can still be invisible in AI answers because every engine is asking Angi who the good operators are. The fix is “claim and complete your Angi listing,” which isn’t in any Google guide. 21 of the 31 gaps in that audit root-caused to platform absence, not site quality.

The self-audit had the same shape on the SaaS side. Perplexity, for example, leans heavily on therankmasters.com, tryprofound.com, amplitude.com, and nightwatch.io when answering “best GEO audit tools” type prompts. Those are the platforms Perplexity has chosen to weight. None of them are blessed by Google’s ranking pipeline. None of them show up in Google’s doc. Getting BeCited reviewed in one of them moves Perplexity citations in a way no amount of on-site optimization will.

So when the headlines this week say “Google says GEO is dead,” what they actually mean is “Google says GEO is unnecessary for Google.” That’s a defensible claim. The leap from there to “therefore optimizing for AI is unnecessary” requires you to believe Google’s share of the AI answer surface will remain dominant, that the other engines will converge on the same source-selection criteria, and that the differences we measure today between engines will narrow rather than widen. None of those are obvious to me from where I’m sitting.

4. Query fan-out makes “rank #1” insufficient

The doc names query fan-out and moves on. Mike King at iPullRank has spent a year measuring what it actually does to citation patterns, and his finding is worth sitting with: a page that ranks #1 for the head term appears in roughly a quarter of the AI Mode responses generated from variants of that head term. The user asks one question. The model asks ten. Ranking for the one no longer guarantees that you participate in the ten.

That changes the surface area of content strategy. If you have one cornerstone article ranking #1 for a high-intent query, you also need supporting content that lives at the sub-query level — the comparisons, the “how does it work” explainers, the “what are the alternatives” pages, the named-entity references that the fan-out is going to ask about. Google’s doc says nothing about this because it isn’t Google’s problem. It’s yours.

What the audits actually show moves the needle

Here is the more useful framing, because it’s grounded in what I see in deliverables rather than in what either Google or its critics say at the policy level. When we compare brands that climbed two or more scoring tiers between audits against brands that stayed flat, the moves correlate with these four things. Each one has a specific audit behind it.

1. Entity consistency across the open web. Wikipedia, Wikidata, Crunchbase, LinkedIn, Google Business Profile, schema.org Organization — same name, same description, same founders, same dates, all linked through sameAs. We saw this concretely in the window cleaning vertical: one operator had a slight name variation between their Yelp listing and their site (“Crystal Clear Window Cleaning” vs. “Crystal Clear Windows LLC”). Three of four engines treated those as different entities and split the citations. Reconciling the name to a single canonical form across all platforms is the kind of thing Google’s doc never names, and the kind of thing that moves recommendation rate within an audit cycle.
2. Presence on the platforms the engines specifically lean on for that vertical. Reddit for many SaaS comparisons. G2 and Capterra for software. Yelp, Angi, and Thumbtack for local services. Industry trade publications for B2B. Source analysis in our own self-audit showed G2 cited 23 times and Capterra cited 7 times across 132 captures, and BeCited was listed on neither at the time. That gap was the single highest-leverage move on the priority list our tool generated, ahead of any on-site change.
3. Answer-first paragraph structure with first-person experience markers. The lifted passage is almost always one that says the thing plainly, early, with the person who knows attached to it. “In our research.” “We measured.” “Our clients see.” This is not robotic Q&A formatting. It’s good editorial habit applied to web pages that historically didn’t need it. The version of this article you’re reading is itself a deliberate test — the opening paragraphs name the doc, the date, and the conclusion, and the “we saw” passages above are written to be extractable on their own.
4. Freshness on the pages that matter, not the whole site. Buyer-intent pages, comparison pages, “best of” pages need a real dateModified that’s within the engine’s freshness window. Evergreen explainers don’t. Updating everything at once dilutes the signal. In one local-services audit between April and May, a single dated case-study update on the operator’s “recent projects” page corresponded with a measurable uptick in Perplexity and ChatGPT citations within three weeks. We’re still measuring whether the correlation holds at n > 5, but the pattern showed up the same way in two other audits since.

None of these contradict what Google’s new doc says. All of them are things the doc doesn’t explicitly name. That gap is the discipline.

The honest operator’s playbook

If you read Google’s new doc this morning and you’re trying to decide what to do about it, here’s the way I’d hold it.

Believe Google about the inputs. The same content, schema, entity, and authority signals that drive classical ranking drive AI visibility on Google’s surface. There is no separate AI optimization layer to bolt on. If your foundations are broken, fix them first.

Don’t over-interpret Google about the surface. Google’s doc governs Google. The AI surface includes engines that don’t use Google’s ranking pipeline and don’t follow Google’s guidance. Audit across the engines your buyers actually use, not just the one Google operates.

Treat passage-level structure as editorial craft, not GEO. Write answer-first. Make claims that stand alone. Put your specific number in the first paragraph. Attach the person who knows it to the claim. None of this is “optimizing for AI” in the way Google’s mythbuster pushes back against. It’s how good writing has always read.

Stop crediting low-leverage tactics. llms.txt does not move citations. Adding novel schema types nobody parses does not move citations. Renaming your blog “AI Search Insights” does not move citations. Pruning these from your hypothesis list saves the budget for the things that do.

Google is mostly right about what doesn’t work. They’re partially right about what does. The gap between “mostly” and “completely” is where the operators who outperform their peers spend their time.

Frequently asked questions