The DNA of a Prompt 1/10: Why Architecture Comes Before Text

You write the text first and organize it afterward. Almost everyone works that way.

The trouble is that the machine reads in the opposite order. It looks at where the passage sits before it decides what the passage is worth.

There is a granted patent that describes this, in exact words.

The machine does not hand you your page. It hands you a slice of it

On October 15, 2020, at the Search On event, Prabhakar Raghavan announced a change in how Google treats long content. The sentence was this: “By understanding passages in addition to the relevancy of the overall page, we can find that needle-in-a-haystack information you’re looking for”. ¹

Understanding passages, on top of the relevance of the whole page. The needle in the haystack.

The feature was introduced first under one name, passage indexing, and Google itself corrected it later to passage ranking. The correction is not a detail of vocabulary. The system does not index passages separately. It ranks passages of pages that are already indexed. ²

The difference changes what you do. There is no separate queue where you enroll your best paragraphs. There is your indexed page, and inside it slices that compete.

The announcement came in October 2020. Production started on February 10, 2021, first for English-language searches in the United States. Three months and 26 days between one and the other. ³

This is called passage ranking.

How the machine knows where that slice came from

If the system cuts a paragraph out of the middle of your page, it needs to know what that paragraph belongs to. Without that, a passage about pricing looks like any other passage about pricing, on any site, about anything.

Patent US 9,959,315 B1, “Context scoring adjustments for answer passages”, from Google, was filed on January 31, 2014 and granted on May 1, 2018. It is active. The claims describe a system that receives a question, identifies candidate passages inside the documents, and determines the heading hierarchy of that document. ⁴

The document names what it does with that hierarchy. It builds a vector defined as “a path in the heading hierarchy from the root heading to the respective heading”: a path through the heading hierarchy, from the root heading down to the heading the passage sits under. ⁵

And here comes the part that decides the subject of this article. That path carries the text of every heading level it passes through. ⁵

Think of an address. “Room 12” says nothing on its own. “Central Building, sixth floor, south wing, room 12” says everything, and each step of the address adds information.

Your headings are that address. The paragraph inherits the meaning of the whole path down to it.

The path is not the only component

It would be convenient to stop here and turn this into a heading recipe. The patent does not allow it.

The same document describes other components of the context adjustment. The depth of the heading inside the hierarchy. The similarity between the question text and the heading text. The coverage ratio of the passage. And the detection of distinctive textual traits, such as a question immediately before the passage, or list format. ⁶

That is four things beyond the path, and they coexist.

Anyone who cites only the heading path impoverishes the mechanism and sells tidy H2s as a complete solution. Architecture matters because it is the layer you control best, and not because it is the only one there is.

Eight years protecting the same idea

That patent family had a continuation. US 11,409,748 B1 carries the same title and was filed on March 16, 2018. Priority is inherited from January 31, 2014. The grant came out on August 9, 2022, and it is active as well. ⁷

From 2014 to 2022, the same engineering lineage stayed under protection.

What the dates prove is the duration of the investment. That the company is protecting ground it considers alive is my reading, and no document says so. ⁸

The prompt

This is the architecture prompt of the series. It builds the heading hierarchy before a single paragraph gets written.

You are a semantic content architect.

I will give you a subject and an audience. Before any text, build the
heading hierarchy of the article, from the H1 down to the H3s.

For each heading, write beside it, in brackets:
1. the question that section answers
2. the full path from the root heading down to it
3. whether that section reads on its own, out of context

Rules:
- each H2 answers a different intent, with no overlap
- an H3 exists only when the H2 has two or more distinct aspects
- no heading starts with "Understanding" or "Getting to know"
- the name of the subject appears in the headings where it is the subject

Do not write the text. Deliver the architecture only.

Subject: [your subject]
Audience: [who reads it]

It forces the structure decision before the writing, and it returns the path of each section explicitly.

One thing it does not do: guarantee position or citation. And it does not replace reporting. It organizes. What goes inside each section stays your work.

Frequently asked questions

Does Google index passages separately?

No. The feature was announced as passage indexing and Google itself corrected it to passage ranking. It ranks passages of pages that are already indexed. ²

How many heading levels should I use?

The patent sets no number. It describes the path carrying the text of every level. And it describes heading depth as one of the components of the adjustment. Too long a path dilutes, too short a path fails to locate. ^{5 6}

Does this work in languages other than English?

Production started in English in the United States, in February 2021, with expansion planned for other countries and languages. The mechanism described in the patent is not specific to a language, and the arrival date for each one is another conversation. ³

Notes and sources

1. Prabhakar Raghavan (Google), “How AI is powering a more helpful Google”, blog.google, October 15, 2020, Search On event. Literal quotation reproduced in the body.
2. Correction of the feature name: Google presented the launch first as “passage indexing” and later corrected it to “passage ranking”, because the original description was not exact. The system ranks passages of pages that are already indexed.
3. Production start on February 10, 2021, first for English-language searches in the United States, with expansion planned for other countries and languages. From the announcement of October 15, 2020 to that date: three months and 26 days.
4. Granted patent US 9,959,315 B1, “Context scoring adjustments for answer passages”, Google LLC. Priority and filing on January 31, 2014, grant on May 1, 2018, status active.
5. Literal definition of the heading vector in US 9,959,315 B1, reproduced in the body. The vector carries the text of every heading level along the path.
6. Components of the context adjustment in the same patent, beyond the heading path: depth of the heading in the hierarchy, similarity between the question text and the heading text, coverage ratio of the passage, and detection of distinctive textual traits.
7. Granted patent US 11,409,748 B1, same title, Google LLC. Filed on March 16, 2018, priority inherited from January 31, 2014, grant on August 9, 2022, status active. Continuation of the application that originated US 9,959,315 B1.
8. Reading of intent from the continuation: the author’s inference. What the dates prove is the duration of the investment in a single engineering lineage.