How to structure a how-to for AI answers

Guide · SEO / AEO / GEO · 5 min read · last verified 2026-07-25

When a buyer asks an assistant "how do I migrate off tool X" or "how do I set up SSO," the model rarely returns a paragraph verbatim. It reassembles a sequence — do this, then that, watch for this — and it draws those steps from pages that already read as ordered procedures. Prose forces the model to guess where one step ends and the next begins; discrete steps hand it the boundaries. This piece covers how to structure step content so answer engines can lift it cleanly, where HowTo schema still earns its keep and where it no longer does, and how to confirm the rewrite actually changed your position.

What "structured for a how-to" actually means to a model

A how-to is structured for AI when every step is a self-contained instruction — an action, the inputs it needs, and the result it produces — that reads correctly with nothing around it. Answer engines extract at the passage level, not the article level, so the unit that has to survive being pulled out of context is the individual step, not the page as a whole.

That single reframe changes how you write. You stop composing a narrative that flows from paragraph to paragraph and start composing a list of blocks, each of which could be quoted alone and still be correct. A reader skimming your page and a model parsing it want the same thing: to find the one step that answers the question without reading the other nine.

Open with the finished outcome, then number the steps

State what the reader will have accomplished in one sentence, then give the ordered steps. A model matching a "how do I..." query looks for a page that names the completed outcome up front and follows it with a clean sequence — that shape is easy to recognise and easy to extract.

Compare two openings. "There are many considerations when configuring single sign-on..." tells a model nothing about what the page delivers. "This guide sets up SAML single sign-on for your workspace in about fifteen minutes; you will need admin access and your identity provider's metadata URL" states the outcome, the time, and the prerequisites in one breath. The second is the one that gets lifted.

Write each step as a block that stands alone

Lead every step with the action verb, keep it to one action, and include both the input it needs and the result it produces. "Step 3: In Settings, open Security, paste your provider's metadata URL into the SSO field, and click Verify — a green check confirms the connection" is complete. "Then set it up in the security area" is not.

The table below shows the difference between a step a model can extract and one it has to reconstruct.

Does HowTo schema still help?

Honest answer: HowTo structured data no longer produces a rich result in Google Search — Google deprecated the HowTo rich result in 2023, according to Google's Search Central documentation. So do not add it expecting a visual feature. It still encodes step order and step text in a machine-readable form any parser can read, which makes it a cheap clarity signal rather than a guaranteed placement.

If you use it, keep the markup faithful to the visible steps — mismatched schema is worse than none. A minimal example, indented rather than fenced:

{

"@context": "https://schema.org",

"@type": "HowTo",

"name": "Set up SAML single sign-on",

"step": [

{"@type": "HowToStep", "position": 1, "text": "Gather your IdP metadata URL and confirm workspace-admin access."},

{"@type": "HowToStep", "position": 2, "text": "In Settings, open Security, paste the metadata URL, and click Verify."},

{"@type": "HowToStep", "position": 3, "text": "Sign out, then sign in via your IdP to confirm the redirect."}

]

}

The clarity of the visible steps does the heavy lifting; the schema just removes ambiguity about their order.

What the Princeton study says actually lifts citations

According to the Princeton GEO study (2024), the changes that most improved a page's visibility in generative answers were citing sources (+40%), adding statistics (+37%), and including quotations (+30%), while keyword stuffing reduced visibility by roughly 10%. Those are optimization-method effects, not guarantees of placement.

For a how-to, the practical read is: give each step a concrete number where one exists — a time estimate, a version, a setting value — and cite a source whenever a step rests on a non-obvious claim. A step that says "this takes about fifteen minutes" and links to the vendor doc it relies on is more extractable and more trustworthy than a vague instruction, and it costs nothing to write that way.

Structural mistakes that bury otherwise good steps

The most common failures are structural, not editorial. A buried lead hides the outcome three paragraphs down. Combined steps pack three actions into one sentence so a model cannot isolate the relevant one. Prerequisites scattered through the body mean the extracted step is missing an input. Screenshot-only steps with no alt text are invisible to text parsers. And keyword stuffing — repeating "how to set up SSO" in every heading — actively works against you, per the Princeton finding above.

Fix these before you touch schema. Structure a human can skim cleanly is structure a model can parse cleanly; the two goals almost never conflict.

How third-party how-tos shape the same answer

Your own page is rarely the only source a model consults. Community walkthroughs on Reddit, answers on Stack Overflow, and step-by-step videos with transcripts frequently appear in the same answer, and pages you do not control are cited disproportionately often. According to published analyses of AI citations, Wikipedia accounts for roughly 7.8% and Reddit roughly 1.8% of ChatGPT's cited sources — small shares individually, but evidence that pages you do not own carry real weight.

The implication is not to spam those platforms. It is to make sure the community version of your how-to is accurate, because a model reconciling your page against a contradictory Reddit thread may end up trusting neither.

Confirm the rewrite actually moved your position

Restructuring a how-to is a hypothesis, not a result. The way to know it worked is to fix a set of "how do I..." questions your buyers actually ask, record whether and how you appear across the assistants they use before you publish, ship the change, and re-scan the same questions against that locked benchmark. If your steps start showing up where a competitor's used to, the structure earned it; if nothing moves, you learn that too. This measure-act-re-measure loop — the same one Magrios runs continuously, with every claimed appearance linked back to the answer it came from — is what turns "we rewrote the docs" into evidence rather than a hunch.