Google named the exact article format not to publish

In May 2026 Google published a guide called Optimising for generative AI features. Buried in it is the single most useful sentence for anyone planning content this year: the worked example of what not to publish is commodity content, for example something like 7 Tips for First-Time Homebuyers.

That is not a paraphrase. Google chose a numbered-tips article on a common topic as its illustration of the problem. It is also, almost exactly, the template sitting in most content calendars right now.

What the guide actually says

Read the whole thing. It quietly kills several claims that are still being sold.

On llms.txt files: Google Search does not use them. They do not help and they do not hurt. The files are fine to publish for other systems, but nobody should be charging for them as a Google tactic.

On structured data: it is not required for generative AI search. Schema still earns rich results in regular search and is worth doing for that reason. It is not the entry ticket to AI answers it has been sold as.

On chunking: there is no requirement to break your content into tiny pieces. The entire cottage industry of restructuring pages into model-friendly fragments has no support from the platform it claims to be optimising for.

On writing style: no need to write differently for AI. And manufacturing brand mentions across the web is not as helpful as claimed.

What the guide does emphasise is unglamorous. The page has to be indexed and eligible to appear with a snippet. It names retrieval-augmented generation and grounding as the mechanism, and query fan-out, described as concurrent related queries, as how the system decides what to retrieve.

So the requirements are: be indexable, be snippet-eligible, and be worth retrieving. Two of those are technical hygiene. The third is the whole job.

The snippet check, first

Before any content work, check this. If you have max-snippet:0 or nosnippet anywhere, you have removed those pages from AI eligibility at the same time. Plenty of sites added those directives during the 2024 and 2025 panic about AI scraping and never reversed them.

Run a crawl and filter for robots meta and X-Robots-Tag headers containing nosnippet or max-snippet. Check your CMS defaults too, because these often get applied at a template level to an entire section without anyone deciding to.

You cannot be cited if you have opted out of being quotable. That is a five-minute check that invalidates six months of content strategy if it fails.

Why the listicle fails specifically

It is worth being precise about the failure, because writing better content is useless advice.

A fan-out query set for a homebuying question generates variations: deposit requirements, stamp duty thresholds, lender criteria, timelines, first home buyer schemes in a specific state. Each of those needs a page that answers that specific thing with a specific number.

A seven-tips article answers none of them. It says save for a deposit where the query needs the minimum deposit to avoid lenders mortgage insurance and what it costs if you go under. The tips article is a summary of things the model already knows. There is nothing in it to retrieve.

That is the actual mechanism. Commodity content fails not because it is short or badly written, but because it contains no fact that the answer needs and could not generate itself.

The replacement test

Before commissioning any piece, ask what a competent model would produce on that title with no sources. If your planned article is within touching distance of that output, do not commission it.

Then ask what you know that the model cannot generate. In practice there are only five kinds of answer, and every one of them is expensive:

• Proprietary data. Your own aggregated numbers, anonymised properly, published with method and sample size.
• First-hand testing. You ran the thing, here is what happened, here is what broke.
• Named specifics. The actual menu path, the actual field name, the actual threshold, the actual date a rule changed.
• Trade-offs stated honestly. Where the recommended approach costs more than it returns, and for whom.
• Practitioner detail. What goes wrong in month three that nobody mentions in month one.

Everything else is commodity content wearing a different headline.

Auditing a calendar against this

Take next quarter's plan and sort every planned piece into three buckets.

Bucket one: pieces whose entire value is the arrangement of widely known facts. Numbered tips, beginner guides, definition explainers on well-covered terms, trend round-ups with no original data. Cut these. Not fewer of them. All of them.

Bucket two: pieces that could carry original substance but currently have none commissioned. A comparison piece with no testing behind it, a guide with no data behind it. Either fund the substance or move it to bucket one.

Bucket three: pieces built on something only you have. Fund these properly and give them more room than the brief suggests.

For most teams the first pass cuts 50 to 70 percent of planned volume. That is the point. The freed budget is what pays for the research in bucket three.

The honest cost

This is slower and more expensive per piece. A guide built on your own client data takes weeks, needs sign-off, and can only be written by someone who understands both the data and the reader. You will publish less.

You will also stop paying for pages that get indexed, receive a trickle of impressions for ten weeks and then disappear, which is what the commodity pipeline produces.

There is one more cost worth naming. Original data invites scrutiny. If you publish a number, someone will check your method. Publish the sample size and the window alongside it, or do not publish the number.

Google told everyone which format is dead and named it. The only remaining question is how long your calendar keeps running on it.